Zelenskyy adviser Serhii Beskrestnov said SpaceX had capped Starshield, its military-grade version of Starlink, with no prior warning, and that the cap disrupted missions. The drops at lower speeds stopped the next day. The cap stayed. On October 21, the Wall Street Journal reported that Trump called Musk and asked him to lift it. Service improved the next day, but Ukrainian officials said it was still not fully restored on October 23.

Then came the blame. Musk posted that Starshield systems are “owned and controlled by the US government” and that the U.S. military and ultimately the President set the limits. Reuters cited officials saying the order came from the White House. The Journal said Musk made the call. Trump said he was not involved. Zelenskyy said both Starlink and Starshield were hit. Secretary of State Marco Rubio said the speeds are “governed by a private commercial contract,” and Pentagon chief Pete Hegseth pointed to Starlink’s terms of service.

Nobody outside the room can yet say who gave the order. The reporting conflicts. The timing is clear: the cap landed in the middle of a fight between Washington and Kyiv over Ukraine’s drone strikes on Russian oil refineries.

Three questions matter more: who holds the switch, what the network is worth, and which lane, commercial Starlink or government Starshield, should carry offense operations.

Rubio and Hegseth have already given one answer, and it runs against the law: Arms and military services for a foreign army normally move only with Washington’s approval, through Foreign Military Sales or a State Department export license. Both men described the network a foreign army fights on as a matter for a company’s contract terms. That puts a commercial contract where U.S. export law usually puts a government decision. In practice, a U.S. company cannot sell an offensive military service to a foreign government without a State Department license.

As of October 13, astrophysicist Jonathan McDowell counted 13,015 Starlink satellites launched since 2019, 11,156 in orbit, 11,142 working. CelesTrak, a public satellite-tracking catalog, puts Starlink at about 66 percent of all active satellites around Earth. Two of every three working satellites in orbit belong to one network.

SpaceX’s third-quarter earnings release reported 13.6 million Starlink consumer subscribers as of September 30, up from 7.2 million a year earlier, plus hundreds of thousands of enterprise subscribers. Connectivity revenue was $5.2 billion in the quarter. Service now covers 174 markets.

The launch side is the moat. In 2025 SpaceX flew 165 of 315 successful orbital launches worldwide, about 52 percent. Starlink rode on 122 of them.

Drones need a backbone

Ukraine and Russia both bet on cheap, plentiful drones and built drone armies. The result is a war of endurance, a 10 to 20 mile kill zone where neither side can mass for a breakthrough. Russia’s 2026 offensive has stalled, with a net loss of ground since September.

The stalemate is not quiet. It is expensive. CSIS estimates Russia lost 275,000 to 325,000 soldiers killed through 2025, and Ukraine 100,000 to 140,000. That averages roughly 6,000 to 7,000 Russian dead and 2,000 to 3,000 Ukrainian dead every month for nearly four years, about 300 soldiers a day across both sides. The pace has not slowed. CSIS put Russia’s total casualties, killed and wounded, at about 35,000 a month in 2025. Ukraine’s General Staff claims 46,000 Russian casualties in September 2026, its highest month of the year. NATO’s Mark Rutte puts Russian dead alone at 20,000 to 25,000 a month, well above independent counts. Both sides contest these numbers and hide their own losses. The direction is not contested. Cheap drones did not end the killing.

Both armies fly drones, but they do not fly on the same network. Russian units lost the smuggled Starlink terminals they had been flying on in February, a cutoff that blunted Moscow’s advance. Ukraine kept its link. Then one speed setting changed what Ukraine’s fastest drones could do overnight. The advantage sat in the network, not the airframe.

The United States knew this in 1998. Vice Admiral Arthur Cebrowski and John Garstka published “Network-Centric Warfare: Its Origin and Future” in Proceedings. Their argument: shift from platform-centric to network-centric war. Shared awareness, speed of command, self-synchronization. The platform matters less than the grid that connects it.

The doctrine grew out of a judgment about drones. In the 1990s, those of us at defense contractors like McDonnell Douglas, along with DoD strategists and DARPA, looked hard at unmanned aircraft and concluded they would not provide a sustainable asymmetric advantage. Airframes get copied. Any adversary with a machine shop and a supply chain can build one, and the edge disappears. Hence network-centric warfare: put the advantage in the layer that is hard to copy. Ukraine and Russia have proved the point. Each side copies the other’s drones within months. Neither has been able to copy the other’s network. Russia tried to borrow Ukraine’s instead, until the February cutoff.

The U.S. military invented that doctrine and has advanced it for nearly three decades. Starlink is the backbone the doctrine now runs on. Starlink was conceived well before Musk built it.

Drones are munitions. The network is the weapon system. Lose the network and a drone force loses its edge. For drones that need a live link, value moves from the airframe builders to whoever owns the link.

Drones sound new. The logic is old. A drone without a link is a model airplane with a warhead. It cannot see past the horizon, cannot be retargeted, and cannot tell friend from foe. The military name for what makes it lethal is C4ISR: command, control, communications, computers, intelligence, surveillance, and reconnaissance. In plain English, the sensors that find the target, the network that carries the picture, and the people who decide what to hit.

Airframes are commodities. Ukraine and Russia each build millions of drones a year. C4ISR is the force multiplier. Starlink is the backbone of Ukraine’s C4ISR. And it is not just bandwidth. That is why the Pentagon buys its own lane. Take away the network, and a commodity drone fleet is blind. Fiber-optic drones dodge jamming, but they trade range for it, and they still need someone upstream who knows where to send them. Anyone can buy the drone. Few can build the network. Especially one with 11,142 satellites.

The hedge against the assassin’s mace

China’s military has a long-standing concept called shashoujian, the assassin’s mace: asymmetric weapons that let a weaker power strike the critical node of a stronger one. For the United States, the critical node has been space. A few large, expensive satellites carry communications, navigation, and warning. Kill those and the American network goes blind.

Starlink inverts that math. You cannot blind 11,000 satellites with a handful of missiles. Lose a few hundred and the network routes around them. SpaceX can launch replacements faster than an adversary can shoot them down one at a time.

RAND’s study of Chinese military writing on Starlink shows analysts in the People’s Liberation Army treat proliferated low-orbit constellations as a real problem and are hunting for countermeasures. Academic work finds Chinese planners favor soft-kill attacks, jamming and hacking, over missiles, partly because they carry less risk of escalation.

I wrote in September 2025 that Starlink is a hedge against China’s assassin’s mace. Against missiles, it still is. This month showed where the new weak point sits. No weapon touched a satellite, yet one speed setting knocked out Ukraine’s fastest drones. The satellites are redundant. The control plane, the software and contract terms that decide who gets what speed where, is not. The software is where Chinese planners say they would aim. The contract terms are Washington’s to fix. Whether a CEO, a government, or an intruder touched that setting, the weak point is the same.

In January I wrote that the network still wins no matter who sits in the Oval Office. That holds against missiles and against presidents. It does not hold against a setting no one accountable owns. The same week I warned that Musk with satellites and AI but no checks on power would be no improvement, just “a new king with better tools.” The only check that worked this month was a phone call from the President. The check I asked for in January is the one this piece asks for now: a government signature on the switch.

The hedge has a second exposure. Russia has already said that commercial satellites supporting Ukraine’s military could become targets. The United States has its own reasons to keep partners from turning its hedge against China into a Russian target.

Starlink is not Starshield

Rubio and Zelenskyy used the two names differently this month. They are different things.

• Starlink is a commercial service. Anyone can buy a terminal. It runs under commercial terms, and those terms say Starlink is “not designed or intended for offensive or defensive military end uses.” The same terms warn that modifying it for military use can bring it under U.S. export-control rules.

• Starshield is SpaceX’s government business, and it has two parts. One is classified. Reuters reported in 2024 that SpaceX is building a spy satellite network for the National Reconnaissance Office under a $1.8 billion contract signed in 2021. That work runs under government contracts, government security rules, and cleared personnel. The other part is a service. The Pentagon buys Starshield access on the Starlink network under “unique terms and conditions” not found in commercial contracts.

Same technology base, different customer, different control.

When Ukraine says Starshield and Rubio says Starlink, they may be describing different terminals, different contracts, or the same hardware routed through different arrangements. The Pentagon has bought Starlink service for Ukraine since 2023, and Bloomberg reported an August 2024 contract that connected 2,500 more Ukrainian terminals to Starshield. A source told journalist Alex Raufoglu that Germany now pays for the service through April 2027. So a government lane already exists. The public record offers only competing claims about who, inside that lane, sets the speed. That ambiguity is most of the fight.

Three questions, not one man

1. Who holds the switch? Rubio says a commercial contract governs Starlink and that Ukraine is not on Starshield at all. Ukraine says Starshield was capped. Either way, no one in government will sign for the setting. That belongs in a government contract, with written terms for speed limits, geofences that set where the service works, and notice before either changes.

2. What is the network worth? It is the American hedge against the assassin’s mace, and it scaled because millions of commercial subscribers help pay for it.

3. Which lane carries offense? The government lane, under the export and neutrality law built for it.

Commercial volume drives down cost. Military need drives up capability. That combination is rare, and anger at a billionaire is not a defense policy. If Washington answered this by canceling contracts, sidelining the company, or pushing allies toward thinner networks, it would lose the commercial scale that makes this capability cheap. If Washington wrote the switch into its own contract, the decision would get an owner, on the record, without breaking the network.

Below the line for paid subscribers: the law that should decide who holds the switch, my theory on the 100 mph cap and the evidence against it, why Washington may want Ukraine’s refinery strikes slowed, and four readings that need no villain. Then who paid for SpaceX’s foundation, including the Super Hornet toolkit I worked on, and three things to watch.

Read more

Suppose we could end chronic disease. What would an agriculture, food, or healthcare business need to become?

A food company might develop products that earn more by improving health. A farmer might need a different buyer and a different spec. A healthcare business might need a way to earn more by avoiding healthcare.

I spent three days at Real Food, the Crusonia Unconference, in Orlando. Health systems, food purchasing groups, measurement specialists, seed and feed producers, and farmers were in the room. Essentially every part of the value chain.

One conversation moved from whether better food was possible to what the specification should be. Each answer changed what the next person could do.

As an investor with 70+ portfolio companies, we are using the unconferences to explore these retooling points and the value-added opportunities between them. Hayek, the AI tool we built in Crusonia, helps connect the work. AI accelerates understanding. The people in the room figure out how to adapt.

Lessons of the Week

1. Singapore spends under 5% of GDP on health care. The U.S. spends about 18%. Singaporeans live about four years longer.

2. Superintelligence isn’t the threat. Soft humans are. For $200/month we have the dev capability of 20 programmers. That is an opportunity.

3. Six $0.75 Cokes or two $2.25 fiber drinks: $4.50 either way. Do we buy gallons or satiety?

4. An A1c test now costs $49. Food that claims health will be judged by a lab result, not waiting to find a doctor’s available time.

5. Costco shoppers pick the house brand 79% of the time. At Kroger it’s 18%. A retailer that owns the label owns the formula. Food is Health may reach the shelf through the house brand faster than through the big CPG brands.

6. When hospitals pay $3 a meal, the waste bin tells you what patients will not eat. Bad hospital food is an excuse for not understanding how food links to health.

7. San Francisco: a $556M fire-department budget for 2025–26; 5,502 fires in 2025, including 245 coded as building fires. The department answers more than 73,000 medical calls a year. Are we designing around fires or medical calls?

8. Every major company has spies inside. The FBI just arrested an OpenAI engineer accused of hiding a recording device in an executive conference room. Design as if the perimeter is already breached. Stealing secrets weakens the spies' ability to innovate, locking them in as a copycat economy.

9. Politicians were the beta test for self-replicating, attention-seeking code. They proved a message tuned for attention outruns one tuned for truth, and AI tuned for engagement will learn that lesson faster.

10. A cow named Veronika picks which end of a brush to use based on where she itches. It’s the first documented tool use in cattle. If a blue wave caps AI and data centers for humans, the cows will pick up Claude and join Musk in taking Animal Farm to the stars.

Share

Some Thoughts

The knowledge exists. It is in the wrong places

Everything we need to build a food system that prevents chronic disease is already known by someone. A rancher knows what a feed change costs. A hospital buyer knows what patients will actually eat. A lab knows what it can measure and what it cannot. None of them knows enough to act alone.

That problem has a name. In 1945, Friedrich Hayek argued that the knowledge that runs an economy never sits in one place. It lives with the people who know “the particular circumstances of time and place.” Three years later, Claude Shannon defined information as what reduces uncertainty. Put the two together and you get our design of unconferences: find the people who hold the missing knowledge, and put them where they cut each other’s uncertainty fastest.

AI covers half of that brief. A model has read the published record. It has never stood on a feedlot floor, negotiated a hospital food contract, or watched a test fail at a packing plant. That knowledge is tacit. The people who hold it rarely write it down, so no model has trained on it.

So we built the unconference as a plugin to AI. The humans supply the tacit knowledge. Hayek, the AI we named after the economist and his theory of coordination, captures it, connects it, and carries it into the next room. Neither works as well alone. Together they compound.

Orlando showed the mechanism. A health-system food buyer knew the outcome it wanted but not how to measure it. A measurement specialist told the buyer what micronutrients to test. A group purchasing organization turned that test into a specification. A meat producer told the room what the specification would take to produce. Four parties answered, in one conversation, questions none of them could answer alone.

Each answer changed the next person’s problem. The measurement reset the buyer’s requirement. The requirement changed what the producer would attempt. The producer’s constraint sent everyone back to the measurement. That loop is where innovation happens, when every link sits in the same room.

Next we price it. The first production run will cost more than anyone wants to pay. We treat that run as the beta. Then we repeat it and drive the cost down.

We work backward from one outcome: better human health at a price the buyer can afford. We trace that outcome through the food, the animal, the feed, and the field. We push testing toward the field, where the interactions start.

I have written about connecting food, health, and retail. Orlando made those connections concrete. Our target is 10x: an order of magnitude less time from a question in the room to a test in the field than one company working alone.

Four rooms, four broken links

We are building the next unconferences around the dependencies Orlando exposed.

1. Health Capital. Health develops over decades. We will work out who pays to improve it, and how that payer shares in the value.

2. Healthy Harvest. We will design the buyers, contracts, measurement, and financing that make regenerative production pay farmers across 50 million acres, and find the value-added products that fund it.

3. PreChronic PCP. We will define what a primary care physician needs from a network built to prevent chronic disease: the food, the measurement, the care, and the payment.

4. Metabolic food design. We will turn Robert Lustig and colleagues’ 2023 Metabolic Matrix into products we can make, measure, and afford, and then test them on human outcomes.

The rooms feed each other. A financing answer exposes an outcome we must measure. The measurement changes a food specification. The specification reaches back to the farm. Hayek carries that thread from one room to the next so nobody starts over.

PepsiCo is competing with the fuel tank

The IRS’s September 2026 guidance lets qualifying regenerative feedstock practices enter the emissions math for the 45Z clean-fuel production credit. The credit goes to the fuel producer. The feedstock bid and the contract decide how much reaches the farmer.

PepsiCo is paying to expand regenerative practices in its supply chain. When a fuel buyer offers a better return on the same acre, PepsiCo has to pay more, sign longer contracts, or find another supplier.

That squeezes food margins. It also pushes food companies to find value beyond the commodity ingredient. A crop that makes a measurably better food can earn a premium that helps fund the farmer’s transition. Regenerative practice and nutritional value need separate proof; one does not establish the other.

Fuel policy now pressures food companies to find a product with a more valuable human outcome. The company that proves that value, and finds a buyer for it, wins the acre.

A shared bill for proof

Lilly’s SURMOUNT-MAINTAIN trial, published in June 2026, tested three paths after tirzepatide weight loss: stay at 10 or 15 mg, drop to 5 mg, or stop. The reduced dose preserved more weight loss than stopping, though the higher dose worked better. Nobody tested food alongside the drug.

We want to run that test: can better food hold the human outcome with less drug? That question gives agriculture, food, and pharma one commercial problem to solve together.

Run the numbers on a shopper who buys 25% less food but keeps a $120 weekly food budget. Spending per unit rises by a third. Richard Thaler’s 1999 work on mental accounting explains why it might: the category budget adjusts more slowly than the appetite. Less food leaves room for better food.

Cornell’s December 2025 research found household grocery spending fell only 5.3% within six months of GLP-1 adoption. Most of the money is still on the table. Food companies have to earn it.

Proof sits between the opportunity and the product. In July 2026, the Seventh Circuit allowed the Chobani Zero Sugar case to proceed because allulose counted as sugar under the regulation it interpreted. That was not a final liability finding. The FDA recognizes allulose’s negligible glycemic response. A useful physiological difference and a defensible label are separate questions.

PepsiCo needs a premium a food product can support. Lilly needs evidence of a durable treatment outcome. Both gain from funding human studies and proposing labeling standards that survive regulators and courts. Nutrient-content and health claims follow different rules, so we design the evidence for the claim we intend to make.

AI Applications

A group second brain. We created Hayek in Crusonia for everyone working on System C Food is Health innovation. One collective hive. This is our own tool.

In Orlando, Hayek listened across the conference and integrated the discussions into action items, lessons, and a future cast. The AI listened, while applying Chatham House rules: shared learning without identifying who said what. The future cast explored possibilities, not achieved outcomes.

The opportunity now is continuity. Hayek could carry a question, the reasoning behind it, and the unresolved dependencies into the next gathering. People could arrive knowing what was tried and where the argument changed.

This is a practical test of the coordination problem in The End of the Sloan Model. The humans bring purchasing requirements, production constraints, judgment, and commitments. Hayek helps retain and connect the knowledge. That is our AI and human maxing experiment. How much sooner can a room get to a test or a purchase?

Three other AI findings in the news:

1. In simulated teamwork, AI agents did worse in 15 of 16 model/context combinations when asked to predict a partner’s next move. Shared context helped more.

2. Robot pursuit simulations used 77.5% fewer messages by sending updates as uncertainty or collision risk grew. Zero collisions in those tests.

3. Anthropic says 5.75% of sampled US Claude conversations touched politics, policy, or social issues. About 0.01% asked whom or what to vote for.

Some Data and Links

Healthcare Is Eating America. Jeff Giesea’s September 29 essay follows healthcare through the rest of the economy. “We spend universal-healthcare money without getting universal healthcare.” My question: who captures the value if better food improves the health outcome?

The chart below compares public healthcare spending and public coverage using 2022 data.

What kills us? Heart disease and cancer killed more than 1.3 million Americans in 2024. What would change if those outcomes organized more of the food innovation agenda?

What limits GDP? Anthropic’s October 2026 study maps staff survey answers to model-implied annual GDP growth of 17.1% in 2030. That is a scenario, not a company forecast. What would have to change to make even 5% growth sustainable?

Skynet. Open-weight models that are self-replicating, able to run on a microwave?

Subscribe now

On September 2, New York City announced a one-year moratorium on student-facing generative AI for students from the youngest pre-K class through eighth grade, about 600,000 in all. The same day, Los Angeles Unified disclosed that it had already blocked generative AI on every district device, all 378,000 students.

Eight days later, Jennica Pounds, the software engineer who writes as DataRepublican, writes,

Bobby did not spend these six weeks in front of a computer screen. All the work he did, was on pen and paper worksheets generated by AI.

Her son had failed a private school placement test. Two years behind in Saxon math. Could not tell a noun from a verb. Six weeks to the retest.

AI wrote a diagnostic. An adult gave it on paper and scored it. AI read each miss and sorted it: missing concept, shaky procedure, or slip. Then it wrote the next page. Sixty worksheets, 398 pages, one child. He passed.

New York cannot watch 600,000 screens, so it wrote one rule for the worst case: the child who pastes homework into a chatbot and learns nothing. A rule for a population cannot see a person, and every one of the 600,000 is curious about something.

The ban and the Utah kitchen table agree on one thing. Both took the screen away from the child. They differ on where the AI went. New York sent it home. Pounds put it upstream.

The screen was never the variable. A World Bank pilot in Nigeria got the same two years in six weeks with the child on a screen. The loop was the same: diagnose, rewrite, repeat.

Upstream AI

The machine never meets the student. It meets the diagnostic, the scored misses, and the adult who scored them. Downstream AI does the homework. Upstream AI writes it.

A web filter drops downstream AI from the Chromebook. It cannot reach upstream, because upstream happens on a parent’s laptop after the child is asleep. Los Angeles school board member Nick Melvoin raised the problem at the meeting: students with their own devices keep going, students on district devices stop.

The rule applies to everyone, and only some families can get around it. A family with an adult who has the hours, a chatbot subscription, and the nerve to ignore the district runs a private curriculum team around one child. Pounds calls hers a high intensity family intervention. A family without those hours hears “stop” and stops. A rule the district cannot enforce at the kitchen table is a grey market. The ban did not end the arbitrage.

One pace for everyone

School systems teach millions the same thing at the same pace. “Stop” keeps the pace intact. It does not sort AI-to-cheat from AI-to-diagnose, which are the same tool in different hands. Telling them apart takes curiosity.

Evolution did not build one kind of learner. It built a spread: ADHD, dyslexia, minds that run on numbers, minds that run on words. One hypothesis is that small groups needed the mix. A 2008 study of Ariaal pastoralists in Kenya found that the gene variant tied to ADHD-like traits was associated with better nourishment among the nomads and worse among the settled. Restlessness is an asset on the move, a liability at a desk. Weakly held. But if evolution paid for a diversity of minds, who decided that every sixteen-year-old, give or take six months, should learn Algebra 2 in the same thirty-six weeks?

Before Gutenberg, knowledge sat in monasteries and you learned from the monk who held the manuscript. The printing press meant you no longer needed the monk. The school that came after, thirty children, one textbook, one chapter a week, was a production line. For two centuries, it worked. It also fixed the pace to the book, because the book could not change.

Now the book can. A worksheet that notices a child confuses discounts with sale prices, and rewrites itself overnight, is a different production function. The best way to learn is not one way. It may be a tool that learns how each human wants to learn, and tailors to curiosity instead of the calendar.

I learned more in life than in school. Would I have learned more with college done by the end of high school and a master’s by twenty-two? I do not know. I suspect yes, and not for the degrees. For the years.

Who runs the loop

A ban teaches a child one of two things about AI. It is bad, so be afraid of it. Or it is banned, so it must be worth trying. Maybe students cheat with AI for a related reason: they do not see the value in the learning, so they use the machine to jump a gate that means nothing to them. At Brown this spring, a take-home midterm averaged 96 and the proctored final averaged 48.

Pounds’ six weeks is the smallest working draft of the other case. She ran the math as a Terminator game; each concept mastered promoted Bobby to the next model. A child in that game has no reason to cheat. He is John Connor, working to be smarter than the machine. Or maybe a different John Connor who works with AI as a partner. The AI sorted the misses. The adults scored the pages and kept him company.

Pounds’ line is the one I keep: “Humans were still the scarce resource.” Now run her loop for thirty. Last bell. The worksheets hit the scanner. The AI reads every answer, flags every miss, and writes the teacher one line per child. The teacher adds what the scanner cannot see: who rushed, who quit, who never read the instructions. The AI writes thirty different worksheets for tomorrow. Next morning the teacher works the room, one desk at a time. Repeat. The machine keeps the records, a benevolent clerk.

Picture Bobby the day percentages clicked. He did not look at the printer. He looked up at the adult across the table. He wanted to hear affirmation from a person. A child who gets that has no gate to jump, and something better to do with the machine: compete with it.

Neither ban reached the teacher’s laptop. New York still lets teachers plan lessons with AI. Los Angeles trains them to use it to differentiate instruction. Rather than ban AI tools or data centers… Maybe a better press conference would have told the story of how teachers are using AI to prepare students for a future where humans excel. And AI assists. Aspire instead of ban.

Subscribe now

Real institutions, agencies, locations and a small number of public figures appear as settings and reference points. Everything else is invented: the two agents, the technician, the officers on both sides, the travelers, the collection programs, and every deliberation depicted in Beijing and in Washington. No real person is depicted as having done, known, or concealed anything.

Do read Chapter 1 First

Chapter Two: Not Yet

Bluffdale, Utah. 24 November 2019, 0340 MST. Erin Vance, contractor, ticket 19-11-2241.

The building has no windows, which is the least interesting thing about it.

It sits on the south end of the Salt Lake Valley, on the shoulder of Camp Williams, which is a National Guard post, which means the perimeter looks like a National Guard post. A chain fence. A gate with a bored kid in it. Sagebrush and a berm and the Traverse Mountains standing up behind it in the dark. From I-15 it reads as a warehouse complex for a company that distributes something to Nevada.

Its real signature is not visible from the highway. It is on a county utility ledger. The building draws about sixty-five megawatts, which is a small city, and something over a million gallons of water a day, in a desert, to keep the floor from cooking itself. That is the trace. The largest archive of human movement ever assembled announces itself to the public record as a water bill.

Inside is TALLGRASS.

TALLGRASS is not a program in the sense of an operation. It is a corpus. Handset position reports. Every mobile device on earth that has attached to a network, every time it attached, with the tower or the beacon or the satellite fix that saw it, and the timestamp. A little over six billion position reports a day, held on a rolling sixty-month window. It is not targeted. It does not need to be. It is the exhaust of the fact that four and a half billion people carry a radio that is required, by physics and by protocol, to keep telling the network where it is.

Nobody at Bluffdale reads TALLGRASS. You cannot read it. You can only ask it something.

Erin Vance had been in the building for four years and cleared for the tool for three. She was thirty-six, contracted through a company in Annapolis Junction, and she drove in from Herriman on nights because nights paid a differential and her daughter was at her mother’s until Thursday. The floor at 0340 held nine people and the sound of forty thousand fans.

Her queue had a ticket in it that had come over from Meade with a priority of routine, which is the third of four, and a requester field that said only a compartment code. The ticket asked for a co-location reconstruction around a fixed civilian location in Wuhan, Hubei, for August 2019, with a note that the requester believed a household disease event had occurred and could not identify the household.

That note was the whole thing. That note was four months late.

Because TALLGRASS answers questions and TALLGRASS cannot ask them. It has no idea what is in it. To run a query you need a selector: a handset, a polygon, a corridor, something the machine can hold onto and walk outward from. In August there had been no selector. There had been a dual-construct sequence with the job title stripped off it and a rumor about a dead man with no name attached, and you cannot seed a graph with a rumor. The corpus had every fact required. Nobody had a way to point at it.

What Peter Lang had brought back from a walk in Wuhan on the twentieth was not intelligence in the classical sense. It was an auntie’s remark that half a kindergarten class had stayed home in the second week of August.

A kindergarten is a polygon. A kindergarten gate at 0740 and 1630 on weekdays is a signature you could pick out of noise on a bad day with half the corpus missing.

Vance drew the polygon. Forty meters by sixty, the courtyard and the gate. She set the window to 1 July through 31 August and pulled every handset that had dwelled inside it for more than four minutes during gate hours on more than six weekdays. Two hundred and eleven devices. Of those, one hundred and sixty-three also showed an overnight dwell at a residential address within four kilometers, repeating, which is the pattern-of-life signature of a parent rather than a teacher, a delivery driver, or a man waiting for a bus.

Then she inverted it, which is the move, and which is why they pay for people and not only for racks.

She did not ask who had been at the gate. She asked which of those handsets had stopped being at the gate. Whose weekday gate-dwell terminated inside a narrow window and never resumed, and whose residential dwell went to twenty-four hours a day at the same time.

That is what “kept home sick” looks like from four hundred kilometers up, rendered as an absence in a repeating pattern.

Nineteen households, clustered on the sixth through the eleventh of August.

From nineteen households she walked backward. In one of them, a handset that had run an eleven-month commute of monotonous regularity, a residential address in Hongshan and the 413 bus corridor and a fixed destination in Wuchang, five days a week for eleven months, had gone dead on 30 July at 1147 local and had never moved again. Not powered off in a way that resumes. Dead in the way a phone is dead when it is in a bag on a shelf in a room the owner does not come back to.

The fixed destination in Wuchang resolved to the Xiaohongshan campus of the Wuhan Institute of Virology.

Vance did not know what that was. She wrote down the coordinates.

Then she ran a rosewood on the dead handset. That is the co-location analytic, all-pairs, which takes one device and scores every other device on earth by how long and how closely and how often it stood near it. She ran it against the nine days before the handset died, because the ticket had said household disease event and nine days was the number the requester had asked for.

The 413 came out of the arithmetic on its own. Not because anyone told the machine there was a bus, but because a bus is what it looks like when forty handsets hold a constant relative position at eleven meters per second along a fixed line for twenty minutes at the same time every morning. The machine did not know the word bus. It knew the shape.

Eleven thousand four hundred devices with any contact. Two thousand three hundred above the exposure-minute threshold the ticket specified. Ranked.

Then the last filter, which took ninety seconds, because it is the cheapest thing the system does. Of those two thousand three hundred, which ones subsequently attached to a foreign network.

Fourteen.

Of the fourteen, twelve resolved to short business trips within the region, Seoul and Singapore and Osaka and Taipei, with home attach resuming inside a week and no onward dwell of interest.

Two did not.

One attached to Vodafone Italia at Malpensa on 9 August at 1841 local, moved by rail to Bergamo, and then dwelled for nine days in a town called Nembro. A parish, an apartment above a shuttered bakery, a factory in Albino on the fifth day, and a forty-minute Sunday dwell inside a church with two hundred other handsets in it.

The other attached to AT&T at Kennedy on 10 August at 1502, moved through Jamaica and Woodside to a residential address in Flushing, Queens, and spent four hours that evening inside an eleven-meter radius with fourteen other devices.

Vance flagged both, attached the reconstruction, and pushed the ticket back up the queue at 0512.

She never learned what it was for. That is not a complaint; that is the arrangement. She was cleared for the tool, not for the question. The compartment code in the requester field was not one she could resolve, and the ticket closed itself out of her queue when someone at Meade acknowledged it, and by then she was on I-15 northbound behind a plow truck, thinking about Thursday.

Two floors below her, the thing that had actually done the work was making most of the heat.

Graph traversal at that scale is not a database problem. It is linear algebra. Co-location scoring is an all-pairs operation. Every device against every other device, a matrix the size of a civilization, multiplied and multiplied again. The only economical hardware on earth for that arithmetic in November 2019 was a floor of Volta-class accelerators wired into each other, sixteen to a chassis, several thousand on the floor, each one drawing three hundred watts and dumping every one of them into the water loop that showed up on the county ledger.

That floor was the most concentrated graph-processing capability in the world, and it existed because the United States had decided that knowing who stood next to whom, everywhere, forever, was worth sixty-five megawatts.

Nobody in the building thought of it as a computer that could learn anything. It was a machine for answering a question you already had.

Here is the part that is only visible from where we are standing now.

The arithmetic that scored who stood near whom in Wuhan is the same arithmetic, on the same silicon, that two years later would be called attention. Every token against every other token, weighted by relevance, all-pairs, over and over, until the weights mean something. A paper describing it had been published in 2017 and was sitting on arXiv, unclassified, free, while Vance ran her rosewood. In February of that same year a lab in San Francisco had put out a model that could finish a paragraph, and serious people had called it a toy.

The rack in Bluffdale and the rack that was about to write like a person were the same rack. The difference was not the hardware and it was not the math. It was the question.

TALLGRASS could answer anything and could ask nothing. That was not a limitation of the corpus. That was the design, and the design was correct in 2019, and it had roughly thirty-six months left.

Because the other thing coming down the same wire, the thing being built on the same accelerators by people with no clearances, for reasons that had nothing to do with any of this, was a machine that generates its own questions. Point that at a corpus like TALLGRASS and you do not need an auntie in Wuhan to remark that a kindergarten class stayed home. The absence is already in the data. In August it was already in the data. It had been sitting in Utah since the second week of August, six thousand miles from anyone who suspected it existed, waiting on a human being to think of the right thing to type.

And the asymmetry that made all of it possible, that only a state could afford to ask all-pairs questions about the entire world, had about three years to live. By 2023 the compute on that floor would be rentable by the hour, with a credit card, by anybody.

The reconstruction was runnable on 7 August 2019.

The corpus was complete. The polygon existed. The bus existed. The two travelers were in the archive before either of them boarded, and the arithmetic would have taken the same ninety minutes it took on the twenty-fourth of November.

There was no seed, because a seed is a question, and on 7 August the only people who had the makings of that question were a provincial case officer in Hubei who was not permitted to ask it and a deputy director in Beijing who declined to carry it twelve meters down a corridor.

At the hour Liang Cheng said the two words that mean by procedure, it was 0540 in Utah. The floor was running. It was answering seventy-one other tickets, all of them routine, none of them this.

Fort Detrick, Maryland. 26 November 2019, 0510 EST. Rachel Keene, ranking the boxes once, for herself.

The August product was still in the system. The city had since thinned. The Institute’s bat-virus database had gone dark on 12 September. USAMRIID was still a parking lot with a badge reader.

And she now had two names, which in August she had not had.

The names came backward, the way names usually do. Lang’s auntie became a polygon, the polygon became a household, the household became a dead handset with an eleven-month commute, and a floor in Utah turned the nine days before that handset died into two thousand three hundred ranked contacts and then into fourteen and then into two. Ninety minutes of arithmetic, on a ticket marked routine, four months after it would have mattered.

Corroboration arrived from a different direction and gave her a date to hang it on. A Thai hospital had a leftover nasopharyngeal swab in a research freezer from a traveler out of Wuhan on China Southern 3034, banked in a respiratory study and never sequenced. Bangkok ran it at CDC’s request and it matched the November pair. Two independent systems, neither aware of the other, produced the same August.

Lin Yating had slept in Nembro. Zhao Jun had slept in Flushing.

That is a pattern, not a proof. Chinese municipal tracers flag hundreds of people a week and most of them fly somewhere.

She would not close the boxes. She would rank them, once, for herself, and then she would refuse to rank them in a room with a President.

The least bad fit was a breach, and then a decision made after the breach. A technician. A cough that was already 17. A spray that was 19. A city that was not locked. August tickets that might have walked the spreader out. Everything after day one available to a man who could do arithmetic.

The exfiltrated Chinese progress report, the one Galaxy had lifted off an Institute server rather than the NIH paperwork in Bethesda, described a transmission model and a severity model. It contained neither QIU designation and no complete construct sequence. The functions matched the two clusters. That was a rhyme, not a proof. The travelers looked like couriers and like buyers. She had no vial, no order, no confession. She had a pattern that fit a wound you would rather share.

Natural still sat there. 17 looks like a bat virus because it was built from one, or because it was one. A market in a city that also houses the laboratory is not a joke. It is the coincidence that eats the next six years whether the laboratory is guilty or not. She had no affirmative market evidence. She also had no affirmative courier evidence. Absence is not a ranking.

Weapon, meaning they built it to fire, she liked least. The benches had vaccine language. The dead man had a daughter and a bus. That does not clear a ministry. It keeps her from saying launch.

She wrote one more paragraph, because the severity numbers did not agree with each other and someone was going to notice.

The technician died in a week. The ward at Jinyintan was losing something closer to one in ten of the old and almost none of the young. Same agent, two pictures, and the difference is the door it came through. A cabinet spray across broken skin puts a million particles into a man at once. A spoon at a banquet puts in what a spoon puts in. Dose is not a footnote. Dose is why a laboratory accident looks like a bioweapon and an epidemic of the same organism looks like a bad flu that kills grandparents. It is also why the two pictures in this file will be read as two diseases by anyone who wants them to be.

Low confidence on each box. Higher confidence, still not high, that box two had happened and that someone had since done math. Moderate confidence that the first-hour decisions did not move if you changed the box.

She sent it upstairs a second time. This time, someone was in town.

She did not have Ma Jian or Deng Yu. Neither man had yet been given a reason to matter.

Bethesda, Maryland. 27 November 2019, 2140 EST. Alan Breit, NIAID, reading 2R01AI110964.

Breit did not have Keene’s finding. He had a renewal file due Monday and an overseas subaward that had been running for five years.

The money was not a secret. NIAID had funded EcoHealth. EcoHealth had sent part of the award to the Wuhan Institute. The progress reports said bat coronavirus, transmission model, vaccine work. They did not say QIU-17. They did not say QIU-19. They did not say a technician was dead.

Fauci knew the portfolio the way a director knows a portfolio: the program, the purpose, the people responsible for it. He did not know the sequences in Keene’s compartment. Keene knew the sequences. She did not know the grant machinery well enough to say what the American money had bought.

Breit put the file in a folder for Monday. No email.

Nobody had decided to hide two halves from each other. One half was classified. The other was administrative. The wall was already there.

The White House, Situation Room. 28 November 2019, 0800 EST. The President, hearing it for the first time.

Thanksgiving morning. Six people. Keene had eleven minutes.

The West Wing smelled like turkey and Sterno. Pompeo was not in the room. The Vice President was in Indiana. The National Security Advisor had been read in at 0610 and had spent the drive deciding what this was going to cost.

She did the facts. Then she refused to pick a winner for him.

A novel coronavirus in Wuhan, weeks not days. Possibly two agents under one name. Vaccine language in a Chinese progress report. A technician, patient zero, dead in a week in August after a night at home, if the rumor holds. Voss saw the city thin. Radar saw Jinyintan fill. Lang pulled virus out of the sewers and antibodies out of the blood supply. Every sequence the hospitals ran, we already had. The August read sat in a bucket with no job title on the tube. A Thai freezer had a swab off China Southern 3034 that nobody had thought to sequence until last week; it matches. A woman flew to Nembro in August. A man flew to JFK the next morning. The Institute’s database went dark on 12 September. USAMRIID, next door, has been shut since July. An August finding sat in the system. No one walked it here.

In August we had one dual-construct read and a rumor. In November we have the same pair in two hospitals, a city thinning, virus in the sewers, antibodies in the blood supply, and a matching sample in Bangkok.

Current estimate inside Hubei, based on mobility drop, hospital fill rates, and the two clusters: three to six thousand already dead, most of them still carried as seasonal pneumonia or cardiac. Outside China the numbers are still small, but the signals are no longer theoretical. Residual sequence in Bangkok. Passenger manifests and early clinical noise consistent with seeding in northern Italy and in the New York metropolitan area. The August travelers are the most likely vectors. If the second agent is already in those chains, the severe picture will appear in those cities first.

This is a major outbreak the public will not see for two to four months. If the second agent reaches the seeded nodes, Bergamo and New York will light up badly. That is true if someone aimed. It is true if a bench leaked. It is true if a market did this on its own.

She put the three boxes on the table.

• One. Weapon.

• Two. Containment breach, and then a decision made after the breach.

• Three. Natural.

“I will not rank these for you. I can tell you the second one fits what we have with the least strain, and I can tell you I cannot close it. The travelers look like buyers. The research language rhymes. The market has not been searched. The first-hour decisions do not move. Warn or wait. Close travel or hold. Put a doctor at the podium or stand there yourself. Name what you have not ruled out, or spend the trust later.”

The President asked the only question that mattered.

“Is it ours?”

Keene answered carefully.

“American money funded work at the Institute through an NIAID award. We do not know that it funded these agents. We do not know that these agents came from the work we funded. Those are different statements.”

“NIAID?”

“The infectious-disease institute at NIH,” the National Security Advisor said.

“Who there can answer it?”

“Fauci.”

The President looked at him.

“Who is Fauci?”

“Anthony Fauci. He runs NIAID. He has run it since 1984.”

The name produced nothing in the President’s face.

“Should he be here?”

“Not in this room. He knows the grant portfolio. He is not read into this collection compartment. If we bring him in, we tell the civilian health side how we are reading Chinese hospital sequences.”

“Can he tell us whether our money bought these things?”

“He can tell us where the money went,” Keene said. “He cannot see what we can see.”

“And you can see the agents, but you cannot tell me what the grant bought.”

“Correct.”

“So nobody can answer me.”

Keene looked at the three boxes.

“Beijing can.”

No one moved.

The President sat forward.

“Then that’s who I talk to.”

He meant it the way he had meant it for forty years. There is a party on the other side of the table. The other party has the information, and you have something the other party wants, and between those two facts is a room and a number and a Tuesday.

Keene took longer with this answer than with any other.

“Sir, they are not a counterparty. They already hold everything we would offer. They have our collection. Not the fact of it, but the effect of it; they will assume we are reading them. They have the grant paperwork, because it is public. They have the sequences, because they generated them. And they have the only thing we actually want, which is what happened in that building in July.”

“Everybody wants something.”

“They want the severe picture not to have one address. Sir, that is the trade. That is the only trade available. It is already priced, and we are not the ones who set it.”

The National Security Advisor said nothing, because there was nothing to add and because he had watched three principals in two administrations arrive at this same door and try the handle.

Keene broke the silence herself.

“Sir, there is a more immediate point. This is already crossing borders. The estimated dead inside Hubei are not small. The signals in Italy and New York are consistent with early seeding. Waiting for a public-health word does not change the biology.”

“The public record is still Wuhan and one swab,” the National Security Advisor said. “Italy and New York are manifestations and inferences. WHO has not used the word. HHS has not used the word. CDC has a working group and no case definition. You are asking the President to be four weeks ahead of every institution whose job this is, on a holiday, on sources we cannot show, in a season when nothing this President says will be heard as a statement about a virus.”

“The virus will not wait to be heard.”

“No. And the country will not hear it. Both of those are true and neither of them is a plan.”

Keene did not have an answer to that, because there was not one. He was right that it would not be heard. She was right that it would not wait. The room had two correct arguments in it and no way to run both.

The President looked down at the boxes.

He had asked six questions in eleven minutes and every one of them had been a good question. Whose paper is this. What is our exposure. Who can answer it. Who do I talk to. All of them the questions of a man who had spent forty years finding the party on the other side of a bad deal and getting to him before somebody else did.

Not one of them had been how many.

Keene noticed it and understood, standing there, that he was not going to ask it. Not out of coldness, but because nothing in a life of buildings and counterparties had ever required the question. A number of dead is not an exposure. It does not sit on either side of a table. There is nobody to bring it to.

“We do not say pandemic. Not today.”

He pointed at Keene’s folder.

“Give HHS the clinical picture. Novel coronavirus. Unusual severity. Probable early export to at least two countries. Current internal estimate of dead inside Hubei. Nothing about two agents. Nothing about the pipeline or the technician.

“The compartment stays closed. That includes NIH. Not the man, the building. Have NIAID map every dollar that reached the Institute and send the answer back through HHS. If they ask why, call it routine oversight.”

He turned to the National Security Advisor.

“More collection. Lang, the hospitals, the outbound traffic, and any signal that the severe agent is already moving. I want options that survive all three boxes in forty-eight hours.”

“Forty-eight hours is time the virus gets too,” Keene said.

“I know what I’m buying,” the President said. “I’ve bought it before.”

He stood.

“Issue the tasking. Keep the circle small.”

The collection orders went out before noon. HHS received a sanitized warning that said major outbreak, elevated internal fatality estimate, and possible early export. It did not say pandemic. It did not say two agents. It did not say American money.

NIAID received a grant query with no source attached. Fauci already knew that NIAID money had reached the Institute. He did not know what had put the question on the President’s table.

Keene knew what the Chinese hospitals were showing. Fauci knew where the grant money had gone. They were not put in the same room.

Nobody had decided to keep them apart. One half was classified and the other was administrative, and there is no procedure anywhere in the United States government for a person holding one to discover that the other exists.

Beijing had both.

Wuhan, then Malpensa, then Flushing. Late November into December 2019. 19, on purpose, in two places.

Wu did not use Lin. Wu did not use Zhao. Those tickets had been 17, and they had not known, and that was the point of them.

He had watched the American tasking come alive on the twenty-ninth. More traffic against the hospitals, more against the Institute, a case officer working the blood supply. None of it changed the arithmetic. It only shortened it. Whatever Washington was assembling, it was still assembling, and the twelve-meter sentence still stood: the severe cases cannot be only ours.

Wu did the same three questions he had done in August. This time only the third one had an answer he could act on.

What has already left. 17 had. Two chains, both four months old, both warm, both in cities where a household still eats from a common dish. Nembro had a parish and a bakery and a valley of towns that visited each other. Flushing had a table for fourteen and a cousin who would invite anyone.

That was the finding, and it was not virology. Seeding is expensive. Seeded ground is not. Wu did not need to find two cities. He already had two cities, and the file already told him how a virus moved through each of them, because a virus had already moved through each of them.

He picked for the ground, not for the man.

Ma Jian, thirty-eight, machine tools, was going to Bergamo because Lin Yating’s tannery had opened a follow-up in June and someone had to close it. He would be inside the same valley, at the same kind of table.

Deng Yu, forty-four, had a cousin’s banquet in Flushing he had been promising since spring. Wu chose him because the file already had Flushing in it. The address was in the folder before the man was.

Neither was told anything. Neither was a weaponeer. Both had real invoices, real customers, and a reason to travel that predated any of this. The Institute ran a traveler’s clinic that week, the way institutes do before a delegation season. Screening, a shot, a form. Two men went through it. Two men went to the airport.

Ma Jian flew CA 439 on 29 November. Albino, not Nembro. A dinner for six above a workshop in the Serio valley. He coughed into a napkin and left before the second bottle. He was on CA 440 on 2 December. He died in a Wuhan fever ward on 6 December. The certificate said seasonal. Two households from that table seeded chains that reached the valley clinics in January.

That was 19. It does not race. It sits in a lung, and then in a household, and then, in winter, in a hospital.

Deng Yu flew CA 981 on 3 December. He shook fifteen hands, because that is what a banquet is. He flew home on the eighth. He made it, which was luck, and which Wu had not planned for either way. Two of those handshakes seeded household chains that reached Elmhurst by February.

That was 19 also.

Wu wrote nothing that said 19. He wrote nothing that said nudge. He filed two travel notifications and a clinic roster, all of which were true.

Sitting in the office that smelled like cigarettes and wet paper, he understood exactly what he had done and exactly how little of it existed. He had a sentence from a man twelve meters from a study. He had two invoices. He had a clinic that treated everyone. Every element was ordinary and the sum was not, and there was no document anywhere in which the sum appeared.

He also understood that he could have done nothing, and that nothing would have happened to him. No one had ordered this. No one would have asked why it had not occurred. A permission is not a tasking, and the difference between the two is that a tasking has a name at the bottom and a permission has a man who decides for himself what it meant. That is the part Wu would not be able to explain later, if anyone had ever asked him, which no one did.

That is the whole operational history of the killer, from the Chinese side, that anyone will ever find.

17 had been in both rooms since August. 19 arrived in December, into cities that already knew how to pass a plate and a spoon. When Bergamo and Queens lit up they would light up badly, and they would light up as one disease, because Galaxy, and every other pipeline in the world, was very sure.

The White House, Situation Room. 6 January 2020, 0940 EST. The second brief, and the decision that was no longer about the virus.

Eight people this time. Keene had the file again. The circle was still small.

She did not restate the August history. She went fast, because the room already knew the shape.

The mild agent was established in northern Italy and in the New York metropolitan area. The severe agent was now present in both. Bergamo and Queens were beginning to see the drowned picture and were calling it undiagnosed pneumonia. The sequences matched the November pair.

Uncontrolled community transmission in multiple countries inside sixty days. No available intervention stops the geographic spread. The windows that existed in August, and still existed in late November, are closed.

“So it’s going to happen,” the President said.

“Yes, sir.”

“All of it.”

“All of it. What is still open is how many, and how fast, and whether the country believes what it is told while it happens. Nothing that is still open is geographic.”

He took that better than anyone expected. It was, at least, a definite answer, and he had spent a career preferring a bad definite number to a good range.

Then Keene put a single page on the table. The NIAID mapping, back through the sanitized channel.

The money reached the Institute. Some of it supported bat coronavirus work that included transmission and virulence models. Nobody could prove the specific constructs in the file were built under that funding; the documentary trail stopped short of the agents themselves. But the functional overlap was no longer a rhyme.

“If the origin becomes public as a laboratory event, the funding link will be read as American participation. Whether or not the law or the science supports that reading.”

The President looked at the page for a long moment.

“So we may be to blame.”

“We may be made to appear to blame,” Keene said. “The distinction will not survive the first news cycle.”

He turned the page ninety degrees, the way a man turns a term sheet toward himself.

“Who signed it?”

“A program officer, sir. In 2014. Under a competitive renewal reviewed by a study section of outside scientists. Nobody signed a decision to build these agents. Somebody signed a decision to fund a category of work.”

“There’s always a signature.”

“There is. It just isn’t on the thing you want it to be on.”

Nobody spoke. The Vice President’s chair was empty again. Somewhere down the hall a printer was running.

Keene stayed standing, because the next part was the part she had come to say.

“Sir, I want to name something before it happens quietly. As of ten minutes ago, this stopped being a public-health question in this room. Not because anyone decided that. Because the public-health question no longer has a live answer, because the spread is fixed, and this one does. From here forward the only variable anybody in this building can still move is how the United States is seen while it happens.

“That is a trap, and I would rather walk into it with my eyes open. Everyone who assesses this from now on, including me, has an interest in the answer. Not a bias. An interest. The question how bad is this and the question are we on the hook for it are going to be answered by the same people, in the same rooms, using the same file. I do not have a fix for that. I can only tell you that it has started, and that it will not feel like anything while it is happening.”

“What would you do?”

“Say it. This week. All of it. The grant, the subaward, the dates, what we knew and when. It costs a brutal quarter and it will be the worst month of the administration. It buys the only asset that matters in March, which is being believed when we tell two hundred million people to do something difficult. Credibility cannot be manufactured during a crisis, sir. It can only be spent.”

“And if we hold it?”

“It breaks anyway. The documents are unclassified. A grant file, a subpoena, an angry program officer. And then the story is not American money reached a Chinese lab. It is they knew in November and hid the money. The first version costs a quarter. The second one costs the podium for the length of the epidemic.”

The National Security Advisor said what he was in the room to say.

“They will send the articles to the Senate this month. Possibly this week. Anything this President says in January about a Chinese laboratory and American money is a story about impeachment before it is a story about a virus. It will not be heard. It will be used, by both sides, and against us by Beijing within the hour.”

“And in March?” Keene said.

“In March we will need the podium.” He did not soften it. “Which is the argument for holding it and also the argument against, and I do not know which way it cuts. Neither does she.”

That was the most honest sentence anyone said that morning, and it did not help.

The President put his hand flat on the funding page and did the thing he had done ten thousand times.

He did not sign.

“The compartment stays closed. The funding assessment stays inside this circle until I say otherwise.” He looked up. “That is not a no. That is a not-yet.”

He said it the way a man says it when a not-yet has always been the right answer. Nothing he had ever waited on had gotten worse while he waited. Markets moved. Counterparties softened. A deal you did not sign on Tuesday was a better deal on Friday, and forty years of being right about that is not a habit. It is a conviction, and it is the correct conviction, in every arena he had ever been in.

“Build the package. Grant, subaward, dates, what we knew and when. Have it ready to go on twenty-four hours’ notice.”

“Ready when, sir?” Keene said. “Who decides when?”

He was already turning to the domestic side.

“Prepare public health for a novel coronavirus with higher severity than seasonal flu and established export. Do not use the laboratory word. Do not use the funding word. First thirty days of domestic response, on my desk tomorrow.”

He stood.

“And find out how much of this Beijing already knows we know.”

The meeting ended at 1012.

The package was drafted. It was reviewed by four people. It was never issued, and no one ever decided not to issue it, because a not-yet is not a concealment. It is a decision that has not arrived, and a decision that has not arrived has no owner, and a thing with no owner does not appear on anybody’s calendar. Keene asked the question twice more, in February, in writing, through a channel that logged it. Both times the answer came back that the matter remained under consideration at a level above her.

Four months earlier and twelve time zones away, a man in a room people walked through had declined to walk twelve meters, for the same reason and to the same effect, and had said so in two words that mean by procedure.

Neither man lied. Neither gave an order. Neither was in the other’s file. Both systems produced the identical output, which was silence, and both produced it the same way, which was by making certain the decision never formally arrived in front of the person who would have had to own it.

The severe agent was already in Bergamo and in Queens. The mild agent had been there since summer. Nothing decided on the morning of 6 January would put either of them back. And in the strict sense, nothing was decided on the morning of 6 January.

Subscribe now

Real institutions, agencies, locations and a small number of public figures appear as settings and reference points. Everything else is invented: the two agents, the technician, the officers on both sides, the travelers, the collection programs, and every deliberation depicted in Beijing and in Washington. No real person is depicted as having done, known, or concealed anything.

Chapter One: By Procedure

Wuhan Institute of Virology, Xiaohongshan campus. 29 July 2019, 1840 CST. Chen Wei, technician.

The two viruses lived in the same room because that was the design, and the design was a vaccine.

QIU-17 was the spreader. High transmission, low kill. A bat SARS-like spike trained onto human ACE2 so you could watch a virus move through a model the way it would move through a city. If you wanted to know whether a vaccine blocked spread, you needed something that spread. 17 was that something. Chen Wei had been on that bench for eleven months.

QIU-19 was the killer. Lower transmission, high kill. Same backbone family, hotter in the lung. If you wanted to know whether a vaccine still saved you when the bat virus came in angry, you needed something angry. 19 was that something. He had been on that bench for six weeks. Director Hao had said the good hands should see both. Hao said it the way you say promotion.

Chen Wei was twenty-nine. He had a daughter in kindergarten in Hongshan. He knew the freezers, the gloves, the centrifuge, the logbook that wanted a time, a lot number, and a pair of initials. He did not know a mission. There was no mission in his head. There were two tools for testing a vaccine against a bat coronavirus, and a lid that stuck when the room was humid.

17 had already left.

Chen Wei had been coughing for nine days. He had told his wife it was the air conditioning. He had told Hao nothing, because a cough is not a report, and a man who reports a cough comes off the 19 bench. He had ridden Wuhan’s No. 413 bus with it. He had read the rabbit story to his daughter with it. He had stood at the kindergarten gate with it. 17 does not announce itself. That is what it was built to do.

The mistake on July 29th was 19.

A late tube. A centrifuge that had been complaining since June. A lid he seated with the heel of his hand because the lock ring stuck, and July in Wuhan is humid.

And Xiaohongshan was not the BSL-4.

The BSL-4 was thirty kilometers south at Zhengdian, new, badged, half empty. 19 belonged in it. 19 was not in it, because the hands that knew 19 were here, and moving a program to Jiangxia is a memo with a signature on it. So 19 sat on a BSL-3 bench in a downtown building with a centrifuge that had been complaining since June.

The lid walked. A spray, fine, almost pretty under the cabinet light. Across the cuff, across the gap where the glove meets the sleeve, onto skin he would not look at for four seconds because he was watching the rotor stop.

QIU-19. The killer. The one you used to see if the shot still worked when the disease was bad.

He did not follow the protocol.

A man who follows the protocol becomes a line in a book. A line in a book becomes a career that is over, and a daughter who will be told her father was careless, and a wife who will be moved. He had been on the 17 bench for eleven months. If he was lucky, the cuff was nothing. If he was not lucky, reporting it would not save him and would name him. He showered. He logged the run as complete. He put the tube back. He left through the ordinary door at 19:10, badge-out, the way he left every Monday.

Home was a third-floor walk-up in Hongshan, twelve minutes on the 413. His wife had rice on the table. His daughter wanted a story about a rabbit. He washed his hands in the kitchen until the water ran cold, which is not a protocol; it is a superstition. He ate. He held the child. He told his wife he was tired. He was not tired. He was waiting to see if the cuff had been enough.

At 0210 he had a fever that was not the cough. He did not call Hao. He took two pills that were for influenza and sat on the edge of the bed so he would not wake the room.

Wuhan Institute of Virology. 30 July 2019, 0740 CST. Director Hao.

Hao came in early because the 19 bench had a run scheduled, and Chen Wei was the pair of hands he trusted with 19.

The log said the Monday run was complete at 18:52. The shower log said Chen Wei had showered at 18:47, four minutes before the run was complete, which is not how a complete run looks. The camera at the ordinary door showed him badge out at 19:10, with wet hair and a bag he did not usually carry.

Hao sat with that for eleven minutes. He was a scientist who had become a director, meaning he was a man who thought about what a log would look like in a hearing. He called security. He did not call Shi. He did not call Beijing. He said: find him, quietly, and bring him to the rooms behind the animal wing, and do not discuss it with the night staff.

He was already writing the sentence he would need if this were nothing. He was already writing the sentence he would need if it were not.

Hongshan. 30 July 2019, 1140 CST. The gate, then the wife.

They came to the kindergarten gate. Two men from the Institute and one who did not introduce himself. Hao was not among them. Chen Wei saw the car and understood. He handed his daughter to the teacher. He did not run.

The wife was two minutes behind, on the 413, a plastic bag of oranges in her hand. She saw the car. She saw her husband put the child into the teacher’s arms and get into the back seat without looking at her. She ran. One of the men turned and said Institute business, go home, someone will call. The child was crying in a way that was not about oranges. The teacher did not know who to give her to until the wife reached the gate.

No one called that afternoon. No one called that night. On the third day a woman from the personnel office came to the walk-up and said Chen Wei had a workplace fever and was under observation and the family should not visit and should not post. The wife said she had been coughing too, since before they took him. The woman from personnel wrote nothing down.

On the morning of 5 August, the same woman came back with a death certificate that said seasonal pneumonia, and a box of things from a locker. A transit card. A photograph of the daughter. The wife asked for the page of the logbook with his initials on it. The woman from personnel said the logbook was Institute property.

The wife asked if it was the laboratory.

The woman from personnel said it was seasonal. She said it twice.

Wuhan Institute of Virology. 5 August 2019, 2310 CST. Hao, ordering the record he did not want.

By the third day in the small rooms the fever had not been a drill. By the fifth, Chen Wei could not make a sentence that held. One week. That was what 19 did to a man who took it across a cuff, at a dose no one on a bus would ever get, with no treatment and no luck.

Hao signed seasonal for the family, because a family asks the question once. He let Beijing be told cardiac, because Beijing asks a different question forever, and cardiac closes a file that pneumonia keeps open in a season with no flu in it. Shi was in Beijing and was told cardiac. Wu Jun, who ran human sources on the Institute because the Institute had foreigners, was told a technician had an accident on the 19 bench and had gone home first.

Then Hao ordered the stored admission sample sequenced that night.

He should have ordered it on the 30th. He had not, because a sequence is a record, and a record is a story with an address. He ordered it now because the alternative was a director who had not looked.

17 was not a question of whether it had left. 17 had been in a cough for nine days, in a walk-up, on a bus, at a gate. A man who spends eleven months on a transmission bench and then a night at home is not a contained event. He is a first case you will pretend was not.

Wuhan. 6 August 2019. Wu Jun, case officer, doing the arithmetic.

Wu was not a scientist. He was a man who kept files on people who kept files on viruses. He sat in a provincial MSS office that smelled like cigarettes and wet paper and he listened to Hao’s liaison, who wanted this to be a workplace accident and a death certificate and a closed loop.

Wu asked the only three questions a case officer has.

Who else was on the benches.

Who the technician saw.

What has already left.

Hao’s man said the 17 bench and the 19 bench were vaccine work. Bat coronavirus. Transmission model. Severity model. American money in the grant, Chinese hands on the pipettes. No military tasking Hao knew about. Wu believed that, or believed Hao believed it. It did not change the third question.

The stored sample had come back at 0140. Both constructs. 19 loud, the thing that had filled the lung. 17 in the background, older, already there. The wife had been coughing before they took him. The kindergarten had not been called. Tianhe was still launching.

17 was already theirs. 17 was already leaving.

Municipal tracing gave Wu two outbound names, not because either traveler was important, but because both had crossed the same chain and both had a foreign ticket in the next four days.

Lin Yating, a provincial commerce officer, had shared the 413 with Chen’s wife on 2 August and reported a cough at her office clinic on the fifth. She had a ticket to Milan on the ninth, booked in June for a real tannery visit in Nembro.

Zhao Jun, a Shanghai plastics buyer, had spent three days with a supplier in Hongshan and eaten at a noodle shop beside the kindergarten on 31 July. He had flown out of Tianhe on the third and told an airline desk he was coughing, to explain why he had missed a connection. That is the only reason the cough exists on paper. He had a Beijing–New York ticket on the tenth.

Neither link proved infection. They were simply the two flagged contacts whose tickets left the country before anyone could close the circle.

19 was different. 19 had killed one man in a week. 19 was still, as far as Hao could swear, in a dead technician and in the rooms behind the animal wing and in whoever had stood too close. 19 was still mostly theirs.

If this stays in Wuhan, it is our death, our growth number, and our standing committee. The world sends condolences and writes a story with one address. That is a wound.

If 17 leaves and 19 stays, we have the mild picture everywhere and the drowned picture only here. That is a worse wound. The severe cases become a map with one city on it.

Wu did not pull the tickets. He also did not let 19 fly on his initials.

He wrote the note he was allowed to write. Accident confirmed. Technician deceased. Two constructs. Vaccine rationale. 17 already in household and municipal circulation. Outbound commercial traffic normal. 19 not confirmed outside the Institute loop. Recommend no disruption that creates a record.

He sent it to Shen Weimin, who ran the Hubei file at Ministry headquarters and could not sign this either, but who sat close enough to a man who could. Then Wu waited.

A provincial case officer does not ground Tianhe. He does not lock a city. He does not put the killer on a plane. He can see the arithmetic. He cannot sign it.

Zhongnanhai. 7 August 2019, 1940 CST. Liang Cheng, Deputy Director of the General Office, the man who decides what the president hears.

Shen brought it in person. That is how you can tell it was already too large for a cable.

Liang’s office is a room people walk through to get to a study. He has no public title anyone abroad can spell. He reads what the president will read, which means he also decides what the president will not.

Shen laid it down without adjectives. A technician. A night at home. Two vaccine constructs, American money, Chinese benches. 17 already moving. 19 still mostly here. A city still open. Two outbound tickets a provincial officer had flagged and had not touched. If we lock Wuhan, we own the story. If we do nothing, 17 acquires other addresses. 19 does not. Not yet.

Liang asked if it was a weapon.

Shen said he did not think so. He said the Institute’s people did not think so. He said that thinking so, out loud, in this compound, was its own decision.

Then Liang asked a question that had nothing to do with any of it.

“How much room does the American central bank have left?”

Shen had to think about it. “Two points. Two and a quarter. They spent most of the last decade at zero and they have not rebuilt.”

“And their election.”

“November of next year.”

Liang nodded once and did not follow up, and Shen filed the exchange the way you file a thing you do not understand yet. It would be four months before he understood that the meeting had turned in that moment, and that everything after it had been a formality.

Because Liang had read what everyone in the building had read. The doctrine is not secret. It has been in open Chinese military literature for twenty years and you can buy it in a bookstore: that a weaker power does not defeat a stronger one by matching it, but by finding the single place where the stronger one cannot defend itself, and that in the modern era that place is almost never a battlefield. The old word for it is the mace in the sleeve. The assassin’s mace. You carry it hidden and you produce it once.

Liang had always found the metaphor slightly childish, because it assumes a swing.

Liang asked if the president needed to see it tonight.

That was the only question in the room that mattered, and both men knew it.

If Liang walked it twelve meters down the corridor, the president would have to act, or be seen later as a man who was told and did not act. Acting means a city closed, flights down, a record. Not acting, after being told, is also a record. Just a quieter one.

Liang looked at the note. He did not pick up the phone. He did not stand. He said it in two words, the way it is always said: by procedure.

Procedure, for a dead technician and a transmission strain already on a bus, is a municipal health matter. It is not a standing-committee matter. It is not a grounding of CA 439. It is not a conversation in the study. It is not, in August, a decision about 19.

Shen waited to see if anything else was coming. Nothing else was coming.

Liang did not pick up a pen. He left the provincial officer’s recommendation where it was. No disruption that creates a record. The tickets stayed live. 19 stayed here. The president went to sleep in a compound that had not been asked a question.

Wu received the instruction the next morning. By procedure.

He did not put 19 on anyone. August was not that. He did nothing at all, which was the instruction.

Doing nothing, on paper, is procedure. Procedure, with 17 already on two tickets you could have killed, is a nudge. The nudge did not belong to a case officer. It belonged to a man twelve meters from a study who chose not to walk the twelve meters.

Fort Meade, Maryland. 2017. how the pipe got built.

Two years earlier, and this is the part that has no cable.

NCMI at Fort Detrick asked for something narrow: every sequence Jinyintan and Wuhan Union run, in a form that does not look like we asked. Not a target list. A subscription.

The problem was that Chinese tertiary hospitals were about to modernize. They were going to buy a clinical sequencing pipeline. That is the software that takes raw reads off a machine, assembles them, calls a consensus, and tells a tired physician at 0300 what is in the tube. There were four credible vendors. Three were American. The best of them was a Maryland company forty minutes from Meade that had never knowingly worked a day for the government.

The Chinese procurement preference, in 2017, was not to license the best one. It was to take it.

So NSA let them take it.

That is the whole operation, and it took eighteen months. A subsidiary’s build server was left one patch behind. A contractor’s credentials aged out of rotation on a schedule someone chose. The exfiltration was made to cost effort, because a thing that is easy to steal is a thing people examine. It was made to look like a win, because a win gets briefed upward, and a tool that has been briefed upward gets deployed.

Beijing’s own people took the credit. They renamed it Galaxy. They stood it up on the benches in Wuhan and Guangzhou and on the wards at Jinyintan and Union and, because it was good, on Xiaohongshan.

A tool you have stolen is a tool you do not audit. Auditing it means someone in the room admits how you got it.

What the implant returned was not dramatic. Galaxy phoned home for reference-database updates, the way every clinical pipeline does, because a pathogen caller that never updates its references is worthless inside a year. The update handshake carried a payload out: every consensus sequence the instance had called since the last check-in, plus a hash of the underlying reads so Meade could ask for the raw file if the consensus looked interesting.

There is no path anyone will find. That is the only sentence in this operation that is fully true. Not the cover, not the legend, not the vendor’s belief about its own breach. The take is real and the trail is not there.

There was one design flaw, and it was not in the implant. It was in the ethics.

Galaxy stripped occupational metadata at the hospital end. It had to. A clinical pipeline is not supposed to know which samples came from a laboratory worker, because a hospital that sorts its patients by employer is a hospital with a lawsuit. So the tube arrived at Fort Meade with an accession number, a city, a date, and a sequence, and nothing that said technician, Institute, 19 bench.

And the second flaw. The functional descriptions of what the Wuhan Institute had been funded to build, which is to say transmission model, severity model, humanized receptor, lived in NIH grant files in Bethesda. Unclassified. Searchable. Nobody had ever ingested them into the compartment, because the compartment was built to read Chinese hospitals, not American paperwork, and because putting NIH documents inside a SIGINT compartment is a memo, and a memo has a signature on it.

A perfect pipeline that does not know what it is looking at is a filing cabinet.

Fort Meade, Maryland. 5 August 2019, 1352 EDT. Michael Voss, and the take that arrived before the printout.

Wuhan runs twelve hours ahead of Maryland in August, which is the whole reason this file is shaped the way it is. Everything Beijing decides at night, Washington gets at lunch, a day earlier by the calendar and a day later by the argument.

Chen Wei’s stored admission sample, the one Hao had ordered at 2310, cleared Galaxy’s update handshake at 0140 on the sixth, Wuhan time. It was early afternoon of the fifth at Fort Meade. Voss had the read four minutes before Director Hao had the printout in his hand, and neither man would ever know that.

Voss pulled it because the consensus caller had thrown a flag it almost never threw: two distinct clusters resolving under one accession. Two conversations in the same room. 19 loud. 17 older, quieter, already there.

He read the metadata. Wuhan. 5 August. Adult male. Nothing else.

He marked it unusual.

Unusual, in August, with no job title on the tube and no city behind it, is not a principals brief. It is a file that waits for a second event.

Fort Detrick, Maryland. 8 August 2019. Rachel Keene, NCMI, writing a finding that will not travel.

USAMRIID next door had been dark since July. CDC cease-and-desist, containment failures, a badge reader on an empty building.

She had a rumored dead technician, a Galaxy take with two constructs and no job title on the tube, vaccine language in an exfiltrated Chinese progress report, and a city that was not yet thinning.

She wrote it.

Accident, low confidence. Dual constructs, moderate confidence. Outbound commercial traffic not restricted. First-hour decisions if this is a leak are the same first-hour decisions if this is a market. She asked for a principals brief.

She did not get one.

August. Recess. A dual-construct sequence with no occupational metadata was a DIA cable, not a principals brief. The deputy who would have walked it upstairs was at the beach. The cable was legal, complete, and unread by anyone who could put it in front of a President.

She marked it and left it in the system. That is how a perfect take becomes a late take. Not because the data failed. Because a country cannot hear a maybe in August.

Air China CA 439, Beijing Capital to Milano Malpensa. 9 August 2019. Lin Yating, in seat 32A, carrying 17, on a trip she booked in June.

She was thirty-four, provincial commerce, Milan twice before, furniture and machine tools. She spoke the Italian of a person who has learned it for work. She had a mother in Wuchang and a ticket that had been sitting in a drawer since the tannery invited her in the spring. She had a cough she had decided was the cabin.

She flew Wuhan to Beijing on the morning 4301. Tianhe in haze. In Terminal 3 she bought water and did not drink it, because the cabin would be dry, which is the kind of thing she thought about. She was not thinking about a laboratory. She was thinking about Luca Bruni’s finishing line and whether the invoice would clear before National Day.

CA 439 pushed at 14:10. The man in 32B wanted to talk about leather, then about his son at UCLA. She said she was tired. She watched the map crawl across Kazakhstan and the Alps. Purpose of visit: business. Address: Hotel San Lorenzo, Bergamo Alta, then the apartment the tannery kept in Nembro.

Malpensa at 18:25. The officer looked at the visa, at the invitation from Conceria Bruni, and stamped her in. No swab. August at Malpensa is families and buyers and no one looking for a coronavirus that does not yet have a name.

Malpensa Express to Cadorna, Trenord to Bergamo, brake dust and cold cuts. A taxi past the hospital, into Nembro after dark. Second floor above a closed bakery. A parish church. Scooters. A town that would later be described as an inexplicable mortality event, as if towns become inexplicable by themselves.

Sunday the family took her to Mass because that is what the family did. She passed the plate. She had a low fever on the third night. She took the factory tour on the fifth day. Luca wanted her to understand the finishing process and did not understand why she kept a meter of air, except at lunch, when she did not.

She left on the eighteenth, CA 440. She posted nothing. The cough was gone by Beijing.

She never heard the name Zhao Jun. She never heard the name Wu Jun. She had a tannery to see, and then she went home. What she left in Nembro was 17. It would look like influenza until something else arrived.

JFK, Terminal 1. 10 August 2019. Zhao Jun, carrying 17, a name Wu did not pull.

Zhao Jun, forty-one, Shanghai passport, CA 981 out of Beijing, seat 47L. Plastics conference at the Javits that existed. Hotel on West 35th that had the reservation. Sent to secondary, and out of it in eleven minutes.

AirTrain to Jamaica, LIRR to Woodside, the 7 to Flushing. Dinner by eight.

His cousin’s table had fourteen people. He served himself with the common spoon because that is how the table works. He flew home on the sixteenth. He was never very sick.

His name had crossed Wu’s desk because of the Hongshan exposure, and Wu had left the ticket live. That is the whole operational history of Zhao Jun, from the Chinese side, that anyone will ever find.

Two tickets. Two cities that had 17 in August, quietly. Not 19. 19 was still in Wuhan.

Or two buyers leaving a city that was already sick, which is what buyers do.

Jinyintan Hospital, Wuhan. 16 November 2019, 2240 CST. Dr. Gao Min, attending, looking at two pictures.

She had eleven patients on the floor who were not the same disease, except the tests said they were.

Four of them looked like a bad flu that had learned to stay. Fever, cough, they sat up, they asked for their phones. That was 17, though she did not have that name.

Seven of them looked like they had been drowned from the inside. That was 19, which had not stayed in the rooms behind the animal wing. Chen Wei had spent five days in those rooms. Rooms have staff. Staff have buses.

Same admission screen. Same Galaxy pathogen label. The consensus sequences were close enough that the software called them one virus. The raw reads were not close enough, and the patients were not close enough, to be one story you could tell a family.

Galaxy sat on the terminal at the nurses’ station the way it sat on every terminal in the building. A prize. Stolen, people said, which is how you knew it was good.

She ran the sequences because that is what you do when the picture does not match the name. She signed the reports. She sent them to the municipal desk. She did not send them anywhere else. She had no idea the reads had already left the building. She had no idea they had left the country before her shift started.

She wrote in the chart: clinical discordance, query two agents, query lab confirmation. A senior physician crossed out the second agent in the morning and wrote influenza-like, pending. He was not a fool. He was a man who had learned which sentences come back.

Gao Min went back to the seven. She did not know Meade had the file. She did not know Chen Wei’s name. She knew only that the software was very sure, and the lungs were not.

Fort Meade, Maryland. 18 November 2019, 0314 EST. Michael Voss, watching a city go quiet.

Wuhan thinned. Jianghan, Wuchang, Qiaokou. Eleven days. Call volume off six percent, then eight, then nine. Mobility worse.

Changsha flat. Zhengzhou flat. Those are the controls. Same region, same season, same holiday calendar. If the drop were a festival or a weather event or an artifact of his own collection, Changsha and Zhengzhou would drop too. They did not. The signal was one city wide, and it was holding through the morning commute, which is the hour a city cannot fake.

He already had the sequences. He had had them since August. They had sat in the bucket because a dual construct with no job title is a curiosity until a city starts to empty.

Jinyintan had run sixty-one samples in eleven days. Union forty. Xiaohongshan a silence after Chen Wei. Two clusters in both hospitals. The hospitals thought they were talking to themselves.

He pulled the August accession and put it next to the November ward. Same pair. 17 common now. 19 no longer a single dead man.

Lacrosse tasked at 0322. Radar imaging, which is the point, because Wuhan in November is under haze and cloud and an optical bird sees nothing at all. The radar saw Jinyintan full. Overflow structures in a lot that had been dirt in June. Hankou station still.

He wrote two paragraphs to Keene. He did not say I told you in August. He did not have to. The file had been true in August. It had not been hearable.

Wuhan. 20 November 2019. Peter Lang, CIA case officer.

Wastewater off the Shahu trunk line, four pulls across nine days. A coronavirus adjacent to SARS in all four, rising.

The Wuhan Blood Center, forty-one archived units, antibodies in eleven of them against something nobody had named. Donors who had been infected and recovered without anyone recording that they had been sick at all.

He asked about Chen Wei. The family had been told seasonal pneumonia. Beijing had been told cardiac. Two different lies to two different audiences is not sloppiness; it is a director allocating. A neighbor said the Institute took the body. The wife had been coughing in July. Half the kindergarten class had stayed home the second week of August, which is a thing you can learn if you ask the right auntie and do not write it down in front of her.

He sat with it on the walk back. 17 had been in the city since July. 19 was in the hospitals now. A virus that stayed in Hubei was Beijing’s problem. A virus that left was everyone’s. 17 had already left, if the August tickets were what they looked like. 19, so far, had not, not in any number you could prove.

Someone on their side had done that math. You could see it in what had not happened. No city lock in August. No board down at Tianhe. No disruption that created a record.

He wrote the incentive into the take. Then he wrote that he could not prove it. An incentive is not a cable that closes a box.

Zhongnanhai. 21 November 2019, 2015 CST. Liang Cheng, again. The twelve meters still unwalked.

Shen came back because the hospitals had made the August note true, and because 19 was no longer a single death certificate.

He did not bring adjectives this time either. 17 is in the wards and has been in two foreign cities since August if the tickets were dirty. 19 is in Jinyintan. 19 is still mostly ours. If the drowned picture stays in one province, the map has one city on it. If it acquires other addresses, the map is a pandemic.

Liang did not ask about the hospitals.

“Sit down,” he said, which he had never said before.

He had spent four months on a question that was not a medical question, and he had spent it alone, because the useful thing about a question you have not asked anyone is that no one can later say you asked it.

“We have been thinking about this as a wound,” he said. “I want you to think about it as an inventory.

“Suppose the severe picture reaches Europe and North America in winter. Not a weapon. Nobody fires anything. A respiratory disease arrives in countries that cannot close themselves.

“Ours closes. We close Hubei on a Tuesday, we absorb it, and in eleven weeks we are manufacturing again. It costs us a quarter and a number we control. Nobody votes on it.

“Theirs cannot. A democracy cannot impose a cost on itself quickly. It has to argue first, in public, with cameras, for six or eight weeks, and then it does the same thing we did on the Tuesday, except later and worse and after the argument has already taught half the country that the other half is lying to them. The disease does not do the damage. The argument does the damage.”

Shen said nothing. This was not a briefing anymore.

“Now the arithmetic. Their central bank has two points of room. They will spend it in a week and then they will print, because there is nothing else to do, and the printing will not feel like a decision to anyone. It will feel like weather. But it moves ten trillion dollars of wealth from people who work to people who hold assets, and that transfer will not be reversed in our lifetime, and every one of those people will vote afterward.

“Their supply chains. They will discover in about ninety days what they do not make. Masks. Ingredients. Simple things. That discovery is worth more to us than a fleet, and it costs us nothing, because we already own the answer and they will have to come and ask us for it on camera.

“And an election in eleven months, run by a country that has just spent a year being told by its own institutions that the institutions were wrong. Whichever side loses will not accept it. That is not a prediction about their politics. That is a property of what happens to a society that has been frightened and then lied to by accident.”

He stopped.

“None of that is something I can do to them. All of it is something they will do to themselves, at a speed we cannot achieve by any other means, if a respiratory disease reaches them in winter and nobody warns them in time to be calm about it.

“I do not want a war. A war we would lose, and a war would unify them, which is the one outcome that undoes all of the above. This is the opposite of a war. This is the largest transfer of relative position available to this country in two hundred years, and its entire cost is that we say nothing for a few weeks.”

Shen finally spoke, because someone had to. “You are describing next winter as though you had seen it.”

“I am describing a distribution. Most of it will be wrong in the details. I have been wrong about the Americans before and so has everyone in this compound.” Liang was quiet for a moment. “And I will tell you the part I am least sure of, because you should hear me say it once.

“I said we close Hubei and absorb it. Eleven weeks. I believe that. But I have never seen this country close and then reopen, and I do not actually know whether we can stop. A system that discovers it can close a city on a Tuesday may find it cannot explain why it should ever open one again. If that happens, the cost is not a quarter. It is years, and it is ours, and it will be the most expensive thing in this room.”

He let that sit.

“I still take the trade.”

Shen said: “And if it is traced to Wuhan.”

“Then it is a laboratory accident in a poor province, which is a humiliation, and humiliations pass. Ask the Americans. They have had several.” Liang turned his hand over. “The trade is a decade of embarrassment for a generation of position. I would sign that.”

He paused on the verb, because he had used it by accident.

“I am not going to sign it.”

Shen understood then what the four months had been.

“The mace,” Liang said, “is a bad metaphor. It assumes you swing. This one does not require a swing. It requires that a man in this building does not walk down a hallway, and that a doctor in Wuhan writes influenza-like, pending, and that nothing anywhere is ever written down. It is the only weapon in history whose entire operation is an omission. There is no order to give. There is nothing to intercept. In thirty years, when someone opens every file we have, the files will be true.”

He looked at the note and did not pick it up.

“The severe cases,” he said, “cannot be only ours.”

That is not an order to a case officer. That is not a flight number. That is a sentence a man can say in a room people walk through, and then go home.

Shen carried the sentence downstairs without adding a verb. Wu understood what it permitted. Nothing was written. Nothing needed to be.

The president, again, was not asked. That was no longer an oversight. It had become the instrument.

Subscribe now

Higher healthcare consumption in the US underwrites the creative destruction that improves per-incidence outcomes globally. Healthcare costs are untenable in total, because a failure to prevent chronic disease is a free rider on Medicare and Medicaid.

When national health expenditure is measured against Actual Individual Consumption (AIC) rather than GDP per capita, the United States aligns much more closely with the expected trend. High spending largely reflects higher real quantities of care consumed in a high-income, high-consumption society. Source: https://randomcriticalanalysis.com/2014/11/24/national-healthcare-expenditure-united-states-versus-other-countries-the-us-is-not-really-an-outlier/

This AIC principle has direct policy implications.

Most of the volume that drives expenditure is concentrated in chronic disease. It is substantially cheaper to prevent or delay the incidence of obesity, diabetes, cardiovascular disease, and related conditions than to treat their complications for decades. Financing reforms that merely compress payment rates leave the underlying incidence untouched. Prevention does not.

High-AIC demand plays a complementary role. The greater willingness-to-pay that accompanies elevated individual consumption underwrites the fixed costs of innovation, particularly in acute and complex care. This private demand signal finances discovery and early diffusion that would otherwise be harder to sustain under uniform administered prices.

The resulting innovations generate positive spillovers. Once refined in the high-AIC segment, novel acute treatments can be extended to lower-AIC patients through Medicare and Medicaid at administered rates. Public programs thereby convert privately financed innovation into broader access without having to bear the full cost of development.

Policy sequence that follows from the AIC principle:

1. Reduce chronic disease incidence at the population level.

2. Preserve the high-AIC demand signal that finances acute-care innovation.

3. Focus ARPA-H and NIH facilitating technology spillover.

4. Use Medicare and Medicaid as the channel that delivers the resulting spillovers to lower-AIC patients.

Expanding administered pricing across the entire population without first addressing disease burden mutes the innovation signal while leaving the volume problem intact. Prevention plus targeted public access to high-AIC-financed breakthroughs is the more coherent path.

Subscribe now

In January 2025, Washington University followed 215,970 patients on these drugs, matched against controls, across 175 conditions. Risk fell on 42 of them. It rose on 19.

Sixty-one conditions moved. One molecule.

They were never sixty-one conditions.

A hormone is a chemical messenger that travels in the blood, which means it reaches every tissue the blood reaches. Nothing about the molecule changes between tissues. Only the cell receiving it. GLP-1 receptors sit on the pancreas, in the hypothalamus, on the stomach lining, and in the brain’s reward circuitry. The drug does not report to a department. It goes everywhere at once.

Medicine is not built that way. Medicine is built by organ. Endocrinology, cardiology, nephrology, pulmonology, hepatology, addiction. Separate boards, separate journals, separate clinics, separate billing codes. That structure came from how you train a doctor and how you pay one. It did not come from the body.

The indication cascade reads like discovery. Diabetes, then obesity in 2021, then a 20% cut in cardiovascular events, then 24% on kidney outcomes, then sleep apnea, then liver disease. Every trial surfaces an effect nobody designed for, and the next molecule is already on the gray market for $25 a dose.

The day Novo said the kidney trial had been stopped early for benefit, DaVita and Fresenius each fell about 17%. Dialysis runs about 1% of the entire federal budget, two companies hold most of the market, and three-quarters of kidney failure cases trace back to diabetes and high blood pressure. A drug that slows kidney decline shrinks the population those companies were built to serve. Nobody had to win that argument. The stock settled it in an afternoon.

Walmart’s U.S. CEO can see GLP-1 users buying smaller baskets. Nestlé sells a frozen meal line designed for them.

Addiction is on the list of 42, and it is the only one there with no incumbent to reprice. Thirty-six million Americans, no dialysis chain, no device franchise, nothing to short. Nobody’s stock moved on that one.

So the thing coming down is not obesity. It is a specialty as a unit of medicine. An org chart drawn around organs and paid by code, meeting a molecule that does not recognize any of it.

We already knew the body was one system. Food said it first. Evolution made is so, without filing any patents.

Subscribe now

A quick reminder. Capitalism is often defined as a capital-controlling means of production. And socialism is some version of State control of the means of production. Or the state persuading the workers that the workers control the means of production - see Animal Farm.

When debating profit and fair wage, they always seem to gravitate to the business of making pencils.

In either case of socialism or capitalism, we discuss the “Control of means of production”. My take is that the real means of production is knowledge. This is the argument in Defoe’s novel Robinson Crusoe. Up until Defoe’s book, people thought of Adam Smith and factories controlling slaves or labor. Defoe’s little island example demonstrated that a single individual could improve their productivity through innovation and build their own productive society. Moving from starvation to leisure.

I think knowledge is really the means of production. If knowledge is the means of production, then how can a socialist control it? Thomas Jefferson realized this and changed “Life, liberty and property” to “Life, Liberty and the Pursuit of Happiness”

Since we produce fewer pencils and more knowledge today, as much as socialists wish to control the means of production, they really can not without inducing those with knowledge via price, which leads to profit. Profit means some people being paid more than others. If you pay a few more than others, those few reveal their knowledge, and then everyone benefits more.

A Tweet that went Viral. A socialist student tells Ben Shapiro that the profits of a pencil company belong to the workers. Give me the machinery and the raw materials, he says, and without me you have a pile of wood, paint, graphite, rubber, and aluminum. Capital needs labor infinitely more than labor needs capital. Shapiro answers that he has millions of people willing to do the work voluntarily, and that money is just stored labor anyway. One of us is completely wrong, he says, and I am pretty sure it is not me.

A socialist student explains to Ben Shapiro why the profits of his company should belong to the workers...

He was left in shock by the response he got.

"Imagine you have a pencil factory and I am a worker. You can have all the machinery and all the raw materials, but without me, you have nothing but a pile of wood, paint, graphite, rubber, and aluminum."

"If you're a single person trying to assemble pencils, you're not going to get very far. You need workers. Capital needs labor infinitely more than labor needs capital."

Shapiro:

"But the thing is, I do have them. I have millions of people willing to do it voluntarily for me."

"I fundamentally disagree with that distinction between capital and labor. Capital is just a word for money. And money doesn't grow out of the ground."

"Money only has value because at some point it was exchanged for labor. If I buy machinery, I'm investing my labor, because the money didn't just appear out of nowhere."

"One of us is completely wrong. And I'm pretty sure it's not me."

They are both arguing over the pencil. Neither of them can make one.

That was Leonard Read’s point in I, Pencil, written in 1958. Not one person alive knows how to make a pencil. The logger does not know the chemistry of the lacquer. The graphite miner has never seen a ferrule press. The owner cannot fell a tree, mill a slat, or run a freighter. The knowledge sits scattered across thousands of people who will never meet, who do not share a plan, and who mostly do not care about pencils. The pencil shows up anyway. Price is what does the coordinating.

Indispensable is the normal condition of an input. Workers do not get profits because their labor cannot be removed. Nothing in the process can be removed. Take out the graphite, and you have a painted stick. Cut the power and the line stops. Being necessary is not a claim on the residual, or every supplier and every utility would have one too. Workers receive what they can contract for in a competitive market. Owners receive the residual because they assembled the arrangement, advanced money before a single sale, and accepted the loss if nobody bought. That is not a ranking of moral desert. Nobody has to argue the owner worked harder.

The measurement is the part that cannot be built. Assign profits politically, by contribution, and someone has to measure each person’s true contribution. Name that someone. No committee, no agency, no ministry holds the knowledge to do it, because the knowledge is not in one place to begin with. The market’s answer is imperfect, and everyone knows it. It also emerges from millions of separate choices instead of one imposed theory of who deserves what. Markets rarely fail for lack of demand. They fail when the signals do not integrate.

And the door is open. Any socialist can assemble a competing firm, buy the machines, and bid for labor in the same market. That is what voluntary exchange means. It does not check your politics at the entrance.

Why are we still talking about pencils? Which brings me back to a more important point. The scarce input now is knowledge. One person knows something, sells it at a price, and everyone with less knowledge lives better for it. But you cannot conscript what is in a person’s head. (Maybe Anthropic can). You can command hours. You can command attendance. You cannot command insight, and you cannot audit whether the best version of it came out. Paying more is the one method that reliably persuades people to release what they know. A state can nationalize a factory. It has no mechanism for nationalizing a mind.

That is the argument worth having, and the clip never gets near it. When the cost of expert-level work collapses, the gap between who knows and who does not becomes the market itself. The measurement problem gets harder from here, not easier. Try assigning a contribution share in a firm where the output is judgment, the tools are rented by the token, and the productive asset walks out at six.

The pencil factory is a comfortable place to argue. Two men, one building, a visible pile of parts, and a number to split. Almost nothing valuable is made that way now. The wood and the graphite were always the easy part.

If AI flattens knowledge, the socialists might just get enough of an edge. I wonder if Bernie Sanders realizes AI is labor’s friend, not its enemy?

Subscribe now

Creative destruction is capitalism’s superpower. The Dow Jones is the perfect example: 30 companies in 1976. Only 1 remains today. Free markets do not preserve the past. They create the future!

He’s right. One name survived fifty years. Procter & Gamble. The other twenty-nine got torn down and replaced.

No one chose the replacements. No committee looked at Anaconda Copper in 1976 and slotted Nvidia into the seat. There was no plan. There was capital moving toward return, consumers picking what they wanted, a government open enough to let the swap happen, and innovation adapting to the next thing. The 2026 list is the residue of billions of small decisions. Not one large one.

The United States, for all its faults, does this better than anyone.

So the question for our decade. Does creative destruction speed up with AI, or slow down?

Hayek framed the stakes in 1945:

The peculiar character of the problem of a rational economic order is determined precisely by the fact that the knowledge of the circumstances of which we must make use never exists in concentrated or integrated form but solely as the dispersed bits of incomplete and frequently contradictory knowledge which all the separate individuals possess.

That was the case against central planning. The planner can’t gather what he’d need to know. The knowledge lives out at the edges, with the people on the spot.

It was also a case for AI, and the mastery of knowledge. Claude Shannon, Hayek’s contemporary, saw that. Three years later, in 1948:

These semantic aspects of communication are irrelevant to the engineering problem. The significant aspect is that the actual message is one selected from a set of possible messages.

Shannon threw out meaning and kept structure. Knowledge became a quantity. Something you could measure, move, and compute. Hayek said it was too scattered to centralize. Shannon said any piece of it reduces to bits and moves anywhere. Put the two together and the dispersed knowledge is workable by machine. The only question left is whose machine.

AI cuts both ways here.

One reading: AI finally builds the planner Hayek said couldn’t exist. Gather everything. Model everything. Plan from the center with a machine that never sleeps. The old dream, now with compute.

The other reading: AI arms the edges instead. It hands each of eight and a half billion people a tool that reads the fine print, prices the option, drafts the contract, checks the seller before the money moves. The person on the spot already had the local knowledge. Now they understand the deal without a specialist, and negotiate from a stronger position. Friction was the tax on being small. AI cuts the tax.

I think the mass of humanity probably has SkyNet beat. Not on principle. On arithmetic.

One planner, however good, is still one search running against the world. Eight and a half billion, each with an agent and their own local knowledge, is a different kind of engine. The center has to be right about everything. The edge only has to be right about one thing. Its own.

The wildcard is rent-seeking. A lot of the incumbents on the 1976 list didn’t die of old age. They died slow, propped up by regulation and captured allocation long after the market wanted them gone.

Musk put the line on it:

Physics is the law, everything else is a recommendation. Anyone can break laws created by people, but I have yet to see anyone break the laws of physics.

Regulation is a recommendation. It buys the incumbent time, not permanence. A better product arriving cheaper is closer to physics, and physics wins eventually. AI in the hands of the people writing the rules stretches the delay. AI in the hands of the people the rules are spent on ends it. Same tool. Whether the next fifty-year list turns over this fast depends on which hands hold it and on capital finding the future rather than guarding the incumbent.

Creative Destruction is the natural order.

Subscribe now

Serrano’s class of 86 had no such trouble this spring. They agreed, nearly all of them, on one thing. Let ChatGPT do the work. But that was not agreement. No one deliberated. Each reached for the same frictionless tool at once. Less a social choice than Huxley’s soma, a whole room dosing in unison.

The take-home midterm averaged 96. Forty perfect scores. A take-home, unproctored, in 2026. Serrano seems to have missed the memo. Only after the numbers came in did he move the final into a room with no phones. The average fell to 48, a record low.
A professor testing a skill the world has retired, then failing to prepare his students for the future he will toast at their graduation.

Paul Graham posted the chart….

The exam graded the school

The gap between 96 and 48 is not a fact about students. When most of a room hands the work to a machine, they are showing what they came for. The degree, and what it gives them. Not the knowledge. That was never the point. The degree is a signal. A false one now, long out of date. Peter Thiel saw it more than a decade ago, and started paying kids to skip college. Serrano meant to test his class. He audited his own institution instead. The number was not flattering.

Some knowledge has lives downstream of it

Engineers, doctors, lawyers. They have to be right. They do not get to hallucinate. In early 2024 a door plug blew off a Boeing 737 over Oregon. The jet turned back and landed with everyone alive. It held because engineers who drew that airframe decades earlier built it to survive exactly that. Get it wrong there, and the failure is not a bad grade. It is a body. Serrano’s course is different. It has near zero relevance to anything but socialism. A welfare proof can be, in his words “kind of correct, but very off.” No one dies. None of these students will hold lives in their hands the way those Boeing engineers did. The course simply does not matter. Claude could teach the highlights and test you the course Socratically in a week.

The tool they cheated with is the tool the world runs on

In 1983, my algebra II teacher at Hotchkiss, Mr. Balmer, made us interpolate logarithms and sines from a printed table instead of the HP-41 in my bag. His reason: someday we might be stranded in a desert with no calculator. I never asked the obvious question. Why, in that desert, would we have the log tables?

These Brown students will never see Balmer’s desert. The calculator is always in reach. The one place it is contraband is the room testing whether they can work without it. Serrano caught this class because the answers shared the same sort of responses. That tell works today only because the model is clumsy. It gets less clumsy every release. Soon the flavor is gone. Then the exam catches no one. Then it measures nothing.

Creative destruction does not mask a dishonest act

Each of these students still chose, and chose badly. Serrano is right: a society cannot afford its best young minds deciding cheating is fine. If the course is a hoop, integrity leaves two honest moves. Do the work, or give up the seat. Cheating is the third door. A bad actor walks through it.

I doubted Balmer’s exercise and used the tables anyway, at home, the HP-41 sitting idle on the desk, no one watching. I thought integrity was part of the lesson. I would have been one of the flat lines in Serrano’s chart. Same score supervised or not, because there is nothing to hide. Those are the students Serrano admired and knew well. They did not need the exam to prove anything. The ones who cheated proved it could not.

An institution near the top of its curve

If the credential is all the school still sells, and the knowledge is free from the same machine that graded the cheating, its core good is now a commodity. What is left is a positional token and a portfolio. The unkind version, getting louder, is that an elite university is a hedge fund with a tax loophole and a lecture hall out front. This is the back half of an S-curve, seen from inside. The function matured. The price of what it sells slides toward the cost of a subscription. The endowment becomes the point.

My uncle Terry joined a university’s board in the early 1990s. He mentioned to me…. “We admit A students and hand them a degree. They were nearly perfect walking in the door. What did we add that would not have done more for a C student?” No one dared ask that in 1990. Is the answer obvious now?

Close Brown tomorrow. Pair a student watching YouTube lessons with a Socratic AI. The power to explore anything. Do we then realize higher education is done? Does the sum of human understanding fall? For the surgeon and the engineer, yes, and it matters enormously. For a hall of students who would rather the machine sit the test, the answer is less flattering. The chart never measured the students. It measured the school. The grades are worth less than the endowment.

Call it what it is. Creative destruction. The machine already took the knowledge monopoly. The credential is next. Brown, and the whole tier of higher education built to sell it. All past their sell-by date. Maybe. The old order never posts its own expiry. It just stops being worth the price.

But the school is the smaller loss. The chart could not measure what the students did to themselves. They optimized for the credential and skipped the one thing the machine cannot rent them. The judgment you earn only by doing it alone first. Always armed. Never capable.

Mr. Balmer was right, in perhaps the way he really meant, absurd desert and all. Serrano ran the experiment without meaning to. Armed, the class averaged 96. Disarmed, 48. In an age of infinite tools, the only edge left is what you can do disarmed. The one who can cross the desert on foot gets the most from the machine. The machine gives little to the one who cheats.

I still know how to interpolate log tables. But I never forgot his story about the desert.

Subscribe now

The race is over, and nobody won. A race has a finish line and a leader who pulls away. What we have is a pack that keeps converging, each new front-runner beaten on price within a fortnight by whoever was just behind. Three frontier models dropped in a single day, which one observer called “not a race heating up, that’s the sound of models becoming a commodity.”

That is not a lead changing hands. That is a commodity forming.

Intelligence saturates, the way it always has. Here is the part the benchmark charts miss, and the reason none of this should surprise us. We ran this experiment already, with people. Average general intelligence is set at 100 by definition, and for the overwhelming majority of real work, somewhere past 120 or 140 the extra points stop paying. A brilliant lawyer and a merely good one file the same motion. Above the ceiling of the task, more IQ is a rounding error. Models are now crossing that same ceiling for most of what we ask them, and the people paying the bills have noticed. As one engineer wrote, “once the weekly model soap opera calms down, a lot of people will discover they never needed the frontier model anyway”

Opus versus Fable is a live question at the frontier and an immaterial one for the ninety percent of work that never goes there. AI, like intelligence itself, was always going to become a commodity. I have argued before that the model is necessary but not sufficient. This is what that looks like when it lands.

The gap is real, it is closing, and the best of it was never for sale. None of this means the frontier labs are in trouble at doing frontier things. The closed models still lead, by something like three to five hundred Elo points on the hardest work, and that lead is worth paying for when the work is hard enough.

But the open models have caught up to where the closed frontier sat a year ago, at roughly one one-hundredth of the price, and a year is not much of a moat. If a lab ever built something genuinely world-changing, the last thing it would do is rent it to everyone by the token. It would keep it and use it. What reaches the API, by definition, is the tier a lab is willing to sell to all comers, which is another way of saying the commodity tier. The breakthroughs will keep coming, and they will still matter enormously. They just will not be the thing on the price list. Meanwhile, the price list gets longer: once a buyer has five vendors who all clear the bar, price falls, and so, in time, do the trillion-dollar valuations built on the premise that one of them owns something the others cannot copy. If the model is the commodity, it is worth asking out loud what exactly the trillion-dollar valuation is paying for. I flagged the price collapse coming when DeepSeek did it cheap early last year; the $1.25 trillion being staked on the idea that raw intelligence itself is the prize is the wager on the other side of this one.

The winner is not the one with the best model, or even the most intelligence. It is the human who fuses a cheap brilliant mind with a lifetime of tacit skill and a better memory of the problem, so that one plus one comes out to a good deal more than three.

I will admit I am living this, not just describing it. My own work now runs across a rotating mix of Claude, OpenAI, Grok, and GLM-5.2, in their cloud environments, driven from Cursor or the Claude command-line harness I have come to prefer. I am token-maxing: sending each task to whichever deal is cheapest for the quality it needs, and switching the moment the math changes. The models have become sufficiently interchangeable that the choice is now mostly economic. My productivity keeps climbing, the cost keeps falling, and my mind can barely keep pace with the tools. It is coping. Not one genius to rule them all, then, but a cheap, swappable bench of them, and a human in the chair learning, a little breathlessly, to conduct.

The old order priced the frontier model as the crown jewel and built the valuations to match. The new one treats it as current, cheap, and everywhere, and quietly moves the margin, and the meaning, to the layer above the pipe. The frontier is no longer a spot on a benchmark. It is the narrow, and very human, space between a rented mind and someone who knows what to ask it, and what to do once it answers. Those few are the same people ahead without the AI.

Subscribe now

Does Roundup cause cancer? The Court didn’t say. I won’t either. I don’t have the data, and neither, honestly, does most of the internet weighing in. Who’s really behind the wave of lawsuits? I’ll get there, but only in the cheap seats, and only once.

Monsanto v. Durnell came down 7–2. Kavanaugh wrote it. The holding is narrow on its face: FIFRA, the federal pesticide statute, expressly preempts state failure-to-warn labeling claims. No ruling on cancer. Design-defect, manufacturing, negligent-testing claims all survive. So the headline “Monsanto wins” is doing a lot of work it didn’t earn.

What interests me is everything the fight is standing in front of.

The warning they could have written

Here is the detail almost nobody is discussing. According to the dissent, Monsanto could have added a cancer warning to the label *itself*, years ago, through a routine “minor modification” notification, no prior EPA sign-off required. A Bayer subsidiary has actually done this. So the whole failure-to-warn theory, the one that generated the litigation billions, could have been strangled at step one by the company simply choosing to warn.

Would it have sold less Roundup? My read: barely. People who buy glyphosate to kill weeds are not, on the whole, deterred by a paragraph of small type. They already assume the stuff is not a smoothie ingredient.

So run the counterfactual. Warn early. Kill the theory. Take the money that instead went to a decade of lawyers and pour it into the next herbicide. The herbicide Bayer needs regardless, because glyphosate resistance is coming whether or not a jury ever rules. The lesson there isn’t legal. It’s strategic. Warn early, and fund your own replacement, because the molecule has an expiration date the courtroom has nothing to do with.

Courts are a slow, blunt tool for driving innovation

A jury cannot invent chemistry. It can only move money after the fact, years late, with enormous variance. If the real complaint is that we don’t trust the EPA’s approval, the clean move is to change the law, not to route product-safety policy through tort one verdict at a time.

And there’s a quieter effect. Weakening agency approval as a shield doesn’t punish bigness. It entrenches it. Only the largest players can afford to test everything against open-ended, unbounded liability. The startup with a better, safer molecule can’t buy that insurance. So the instinct to make Bayer bleed mostly guarantees that the only companies left standing are the Bayers. This is the same regulatory machinery that keeps producing outcomes nobody voted for: the rule written to restrain the incumbent becomes the moat that protects it.

Everyone says cancer. I keep thinking about the microbiome

Glyphosate targets a metabolic pathway that human cells don’t have, but our gut bacteria do. So there’s a plausible story where the effect that touches the most people isn’t a tumor in a single plaintiff but a diffuse, low-grade disruption across millions of guts, the kind no tort could ever measure because there’s no discrete injury to point at.

I don’t know the magnitude. Honestly, nobody does. But it fits a pattern I keep noticing in food and health: the variable that ends up mattering is rarely the one on trial. We litigate the scary word and ignore the boring, distributed one. At an absolute minimum, people ought to be able to know whether it’s in their food, and anyone spraying it should take precautions no label is going to require.

The lineup didn’t match the politics, and that’s the fun part

7–2. The two dissenters were Gorsuch and Jackson. Gorsuch silently joined Jackson’s dissent and wrote nothing separate. Meanwhile Thomas concurred with the result, he voted for Monsanto, but wrote separately to argue that FIFRA itself is unconstitutional.

Sit with that. There are two completely opposite “anti-administrative” instincts on this Court, and this case pried them apart.

Thomas’s is constitutional demolition: the whole delegation offends the structure, the EPA shouldn’t be doing this at all. And yet he shields the company the agency he distrusts had approved. Gorsuch’s is textualist anti-deference: read the words. FIFRA says registration is only “prima facie evidence” of compliance and is “in no event” a defense to misbranding. Evidence, not a command. If the approval is merely rebuttable evidence, it can’t also be the automatic, permanent immunity the majority turned it into.

Two skeptics of the administrative state, landing on opposite sides. One shields the company from a regulator he thinks shouldn’t exist. The other refuses to let a regulator’s sign-off harden into a shield the statute never granted. The tidy political model, count the appointing presidents, tells you nothing here. The real fault line is about who wields power and where the text draws the line, and those two questions don’t sort by party at all. I kind of wish Gorsuch had written it up himself; the silent join left the best argument in the case underlined by someone else’s pen.

One detour into the cheap seats

Against my own rule, and purely in the spirit of full speculative abandon: if you’re the sort of person already certain that the aliens landed and nobody told us, then we may as well throw a little random fuel on the Roundup fire.

The rumors, stated neutrally: that litigation-finance investors bankrolled the suits for a cut of the verdicts; that foreign interests are happy to see American agriculture hobbled; that a faction mostly wants the administrative state taken down a peg and found a convenient vehicle. I have no idea if any of it is true, and I’d bet most of it isn’t. But notice the tell, none of these rumors are about whether Roundup is safe. Which is the point. The case became a screen. Everyone projected the fight they already wanted to have onto it.

And it isn’t really about Roundup

The majority itself analogized to statutes covering food, cosmetics, over-the-counter drugs, meat, poultry, eggs. The principle, that an agency’s individual product approval counts as a federal requirement that blocks state failure-to-warn claims, is portable. If you care what’s allowed to be hidden from a food or supplement label, this pesticide case is your dry run.

Here’s the catch under the catch

Lawsuits did a second job nobody gives them credit for. Paying injured people was the visible one. The real one was information: the threat of discovery and a verdict is what forced companies to put data on the table they’d otherwise have sat on. Remove the lawsuit, and you remove the thing that pulled the data out.

So what’s left holding a company accountable? Really just one thing, the EPA’s approval. And that approval is a single snapshot, built from the data the company chose to submit. It’s the thinnest safeguard available, and Durnell just made it the only one.

The obvious response is to attack it: strip the shield, force Congress to write something better. I’d be careful there. When you blow up a law, you don’t get to write the replacement, the company with the biggest lobbying budget does. So you don’t get a better system. You get a better-funded version of the same shield.

The move I’d make instead is to build the missing option before that fight arrives. Not immunity, not endless lawsuits. A third thing: a way to measure what a product actually does once it’s out in the world, where honestly disclosing the risks (and how sure you are of them) is what earns a company legal protection. Build that, and the next time Congress is cornered, there’s a real alternative already on the table instead of a forced choice between shielding the company and suing it.

Because accountability doesn’t vanish when the lawsuit goes away. It just moves to wherever someone is still measuring. Right now, once the EPA signs off, no one is. That empty space is the whole opportunity.

None of which is a doctrine. It’s just the oldest human courtesy: tell people the risk, and let them decide. It would have saved everyone a decade. A measurement layer would only make it normal.

Subscribe now

Nine months later, on April 15, 2026, the agency rejected it again. The stock fell another 75 percent. The company announced layoffs, wrote down its Phase 3 confirmatory trial as likely unviable, and acknowledged it would need to raise $250 million dilutively on a balance sheet now worth roughly $150 million.

The data on RP1 wasn’t bad. In the IGNYTE trial, 33.6 percent of patients who had previously failed anti-PD-1 immunotherapy responded to RP1 plus nivolumab. Fifteen percent achieved complete responses. Median duration of response exceeded 33 months. Grade 3 or higher treatment-related adverse events ran at 12.8 percent. Zero RP1-related deaths.

Yet the FDA rejected it. Twice. And they weren’t wrong, exactly. They were operating the way a 63-year-old law told them to.

The Honest Challenge in the RP1 Data

Before blaming the regulator, look at what the company actually did.

To enroll in IGNYTE, a patient had to have been on anti-PD-1 therapy for at least 8 weeks and have confirmed disease progression. Eight weeks is about the time it takes for anti-PD-1 to start working for the patients it’s going to help. The trial then put those patients back on nivolumab, the same class of drug they had just been declared refractory to, alongside RP1. Hmm……

Published rechallenge data shows anti-PD-1 alone can produce response rates in the 15 to 21 percent range in similar populations. Not 33 percent. But not zero either.

This creates a legitimate problem: in a single-arm trial, you cannot statistically separate RP1’s contribution from nivolumab working on patients who were still inside the window where nivolumab might work. FDA regulation 21 CFR 300.50 requires that each component of a combination make an identifiable contribution. Replimune didn’t design the trial to answer that question. So how could FDA discern it?

This was a choice. The company made it. The board made it. The CEO made it. They chose to assume the 8-week cutoff was enough to establish refractoriness, file on single-arm data, and move fast. The FDA led them along (or rather didn’t say no), granting Breakthrough Designation, Priority Review, and a written pre-BLA agreement.

And then the agency changed its mind.

A randomized Phase 2b with even 150 patients, one arm getting RP1 plus nivolumab and one getting nivolumab alone, would have cost roughly $20 to $40 million more. It would have taken another 12 to 18 months. It would have made the contribution-of-components question answerable. It would have protected $3 billion of market cap from evaporating.

The board optimized for the nearest catalyst. The catalyst blew up. Now the company has to raise $250 million at a market cap 25 times lower than the day before rejection to run the Phase 3 it should have started with. The capital math doesn’t work. A treatment that might have worked will likely never get tested at the scale to prove it.

This is not just an FDA failure. It’s a capital allocation failure trained into the industry by the regulatory system that created the incentive.

The Ambiguity Machine

In 1962, Congress passed the Kefauver-Harris Amendments. The thalidomide disaster had just unfolded. The response was to require drug manufacturers to prove not just safety, but “substantial evidence of effectiveness” before approval.

It sounded reasonable. It wasn’t.

Safety has a bright line. You can count bodies. You can measure toxicity. You can test in animals, then healthy humans, then sick ones. The answer is binary enough to hold.

Effectiveness has no such line. It’s a moving evidentiary standard that depends on trial design, comparator choice, endpoint selection, surrogate validation, statistical power, and which review division you walked into that Tuesday. Every one of those choices is contested. Each creates a political surface.

That’s the design flaw. Kefauver-Harris handed regulators a discretionary lever that gets pulled harder when political winds change. The law creates ambiguity. The ambiguity creates the rent. The rent creates the lobby. The lobby pressures politicians. The politicians pressure regulators. A great way to raise money if you are in politics.

It’s not a conspiracy. It’s a predictable equilibrium.

Run the Numbers

The foundational paper here is Gyawali, Hey, and Kesselheim, JAMA Internal Medicine, 2019. They looked at 93 cancer drugs approved under the accelerated pathway between 1992 and 2017. Only 19 showed an overall survival benefit in confirmatory trials. That is 20 percent.

The Liu, Kesselheim, and Cliff update in JAMA 2024 found that among accelerated approvals with more than five years of follow-up, only 43 percent demonstrated any clinical benefit in confirmatory trials. The other 57 percent either failed or never converted.

Translate that: most cancer drugs that clear the FDA effectiveness bar don’t actually help patients live longer. One in five under the stricter definition. Fewer than half under the looser one. We spend years of trials and expense to get what new information?

Why do we still demand effectiveness before approval?

The honest answer is we don’t, really. We demand a bar that can be argued over. Aduhelm was approved after a 10-to-0 advisory committee vote against, then withdrawn three years later. Exondys 51 was approved over the objections of FDA staff scientists. Relyvrio was approved, then pulled when Phase 3 failed. Makena held the market for 12 years before its confirmatory trial came back negative.

The effectiveness standard doesn’t screen out drugs that don’t work. It creates a queue in which the drugs that clear it are the ones whose sponsors fought the hardest, not the ones whose biology is strongest.

This is rent-seeking. The law says don’t approve low-effectiveness drugs. Politicians show up and yell at regulators to approve anyway. Patient advocacy groups funded by the sponsors testify. Investors call their senators. Review teams get overridden. Directors resign.

Then the next RP1 comes along with clean data and gets rejected, because the pendulum swung the other way that quarter. Same law. Different outcome. Different rent.

A Better Bar

Strip the law back to 1938. Approve for safety. Require public reporting of efficacy.

Don’t force withdrawal when a confirmatory trial fails. Require disclosure of that failure, in plain language, on the label and in every ad. Let citizen scientists call the bullshit. They will.

Two forces exist now that didn’t exist in 1962. Large structured datasets, and independent analysts who can read them. This is where it gets interesting.

Three Things to Watch

1. The citizen scientist with AI and data access

In March 2025, Kevin Bass used an AI system to process 1.4 million pages of the Jeffrey Epstein DOJ document release. He surfaced 261 documented meetings between Woody Allen and Epstein across nine years, 4,095 supporting documents, flight manifests, and dinner logs. Every quote verbatim. Every date from DOJ files. No anonymous sources. No inferences.

By March 2026, Bass had downloaded 600 gigabytes of federal health data. CMS Part D prescriptions. Open Payments. Medicare, Medicaid, Healthcare.gov. The entire publicly available federal healthcare data ecosystem, sitting on his disk, being processed by AI pipelines he built.

Nick Shirley is a different flavor of the same thing. Twenty-four years old, no journalism background, a camera, and a willingness to show up. His viral video on Minnesota daycare fraud froze $185 million in federal funding and triggered an FBI investigation. He testified before House Judiciary in January 2026. He has since moved on to California voter rolls and hospice fraud. He raises money from viewers for private security. He answers to no editor, no network, no party.

This is citizen science plus distribution, running on AI infrastructure that didn’t exist three years ago. The Gyawali paper that took a PhD oncologist six months can now be produced in a day by someone with Claude, a Python environment, and enough stubbornness to keep digging. The Bass Epstein analysis that would have required a team of investigative reporters and a year of work was built by one guy.

A rent-seeking cabal can capture one FDA review panel. It cannot capture ten thousand independent analysts pulling apart the data in public. The cost of calling bullshit has collapsed.

This is the Napster analogy. When MP3s arrived, the major labels were destroyed. But music wasn’t destroyed. It exploded. More artists got paid, more songs got made, more listeners got reached. The rent extracted by the five-label oligopoly collapsed, and the economic surplus was redistributed to the edges. Drug approval can work the same way. Blockbuster pharma has built its business on a regulatory moat that requires $2 billion and a decade per approval. Force public efficacy data, and the moat collapses. It was collapsing anyway…..

2. Data-driven millennials and Gen-Z pressuring reform

This generation is running its own clinical trials on itself. It buys continuous glucose monitors the way earlier generations bought watches. It orders peptides from compounding pharmacies for $25 a week. It pays Function Health for 100-biomarker panels and throws the data into Claude. It doesn’t trust the FDA, doesn’t trust pharma, doesn’t trust the AMA, and doesn’t need to. It has built a parallel system.

The tension this generation creates is productive. They tolerate over-promising on the front end because they distrust the intermediaries who used to filter for them. But they hold companies accountable on the back end through data they can measure themselves. That’s the opposite of the current system, which under-promises, under-delivers, and hides the results.

When this cohort reaches scale as voters and consumers, the political pressure for reform becomes irresistible. It’s already visible. The gray-market GLP-1 economy is what Kefauver-Harris reform looks like in practice. Millions of people are doing their own risk-benefit analysis, buying on the open market, sharing results on Reddit, and proving efficacy through crowdsourced outcomes data that moves faster than any FDA review cycle.

3. Prediction markets on drug efficacy

Polymarket, Kalshi, and the next generation of regulated venues can host binary markets on every confirmatory trial. “Will RP1 show OS benefit in IGNYTE-3?” becomes a standing, liquid market. The price is informative. It aggregates more information than any advisory committee.

The real benefit is limiting the boom-bust cycle in biotech. Right now a single binary FDA decision can move $3 billion in a day. That volatility is not a signal about the drug. It’s a signal about the regulatory capture that precedes the decision. A prediction market gives investors a continuous signal, trading against the expected FDA outcome. When the market price and the FDA decision diverge, you have learned something.

Drug sponsors hate this because it limits the hype cycle. Patients benefit because the market price tells them what the experts actually believe, not what the press release says.

Back to RP1

Replimune could have run a randomized Phase 2b for $20 to $40 million. Instead, it filed on single-arm data, destroyed roughly $3 billion in market cap across two rejections, and now needs $250 million to finish the Phase 3 that its board should have started with.

The board optimized for the nearest catalyst. The system trained them to play that game. The 43 percent confirmatory benefit rate tells you the game is mostly cheating. Polymarket would have forced them to rethink their strategy.

RP1 might actually work. We will probably never know. The system is optimized to reward bad trials that argue well, and to punish good trials that argue poorly. And provide politicians a toll booth to extract rent. The ambiguity at the center of Kefauver-Harris makes the argument louder than the biology.

The Political Opening

RFK Jr. got grilled this week at a congressional hearing over the RP1 rejection. The WSJ editorial board has been treating him as personally responsible for every patient death traceable to a drug the FDA blocks. Every RP1-style case that follows is another chance to run that headline. WSJ did not bother to suggest a better way. 30 years ago, it would have.

He has two options. Keep defending individual FDA decisions case by case, a losing position because the cases are technical, the data is contested, and the optics are always patients dying. Or flip the frame.

Reform Kefauver-Harris.

The coalition is already assembled. MAHA is anti-pharma rent-seeking from the start. The peptide underground is running its own clinical trials and doesn’t want FDA permission. Data-driven millennials don’t trust any institutional filter and never will. Free-market conservatives want the moat torn down. The only defenders of the status quo are the blockbuster pharma companies that built the moat, the advocacy groups they fund, and the review divisions that got captured.

Push safety-plus-public-efficacy-reporting, and you capture the MAHA base, the gray-market peptide users who want legal access, the Gen Z self-quantifiers, the biotech industry that would back any reform reducing regulatory risk, and the libertarian right in one move. Then let Bass pull the CMS data. Let Shirley knock on doors. Let the market price the drugs. Pharma can’t lobby its way past ten thousand independent analysts.

This is the play for the midterms. Not a defensive crouch over RP1. Or blasting more gerrymandering. An offensive push for 1962-style reform with 2026-style tools.

It also solves the real political problem, which is the “patients dying while you block drugs” headline. You cannot defend Type II errors one at a time. You can defend a reform that acknowledges them and builds a different system.

Subscribe now

Its Time

You may not like RFK, anti-vax, drug companies, billionaires, or whatever. That is sort of irrelevant.

The facts on the ground are that the current system is not working. Its the end of the line. We have the tools to make the change. It will be messy. The forces of change are here.

Someone will harness the forces of Creative Destruction.

Schumpeter said Creative Destruction is what keeps capitalism honest. The entrenched incumbent gets replaced by the scrappy upstart, not because the regulator chose it, but because the market did.

We haven’t allowed that in pharma for 63 years.

It’s time.

Get more from Carter Williams in the Substack app

Available for iOS and Android

Get the app

You can buy Eli Lilly’s next obesity drug online today. Retatrutide is not approved. Lilly is suing FDA over how to classify it. A compounding pharmacy in Florida or a research chemical site in Shenzhen will ship you the molecule this week for around $25 a weekly dose. The demand is verified. The synthesis is trivial. The gray market is already running.

This is what creative destruction looks like when the business model breaks before the science does. Speed in healthcare has never been a scientific question. It has always been a business model question. The science is usually ready years before the capital stack and the regulatory architecture catch up. Consumers do not wait. They route around the obstacle.

A Breakthrough in Pancreatic Cancer

Memorial Sloan Kettering and BioNTech just reported six-year follow-up data on a personalized mRNA vaccine for pancreatic cancer. Sixteen patients. Eight produced durable T-cell responses. Nearly all the responders are still alive. In a disease where 87% of patients die within five years, that is a breakthrough.

Now look at how narrow the path to that breakthrough actually is.

Pancreatic ductal adenocarcinoma kills because it hides. There is no colonoscopy equivalent, no mammogram, no PSA test. Symptoms show up when the tumor has already grown around blood vessels or seeded the liver. By the time most patients walk into an oncologist’s office, the disease is Stage 3 or Stage 4. For those patients, this vaccine does nothing. They are not eligible.

Only about 20% of pancreatic cancer patients have tumors early enough and anatomically clean enough to be surgically removed. The operation depends on where the tumor sits. Roughly two-thirds arise in the head of the pancreas, which requires a Whipple procedure. That is a six to eight hour reconstruction where the surgeon removes the head of the pancreas, the duodenum, the gallbladder, part of the bile duct, and sometimes part of the stomach, then rebuilds the plumbing of the upper GI tract. Tumors in the body or tail get a distal pancreatectomy, often with the spleen removed. Either way it is major abdominal surgery. Mortality in experienced hands is under 3%. Serious complications run 30% to 40%. Recovery takes months. You need to be relatively healthy going in, or the surgery itself will kill you before the cancer does.

So the funnel starts at 100% of pancreatic cancer patients. It drops to the 20% who are operable. It drops again to the subset healthy enough to tolerate the resection. It drops again to those who survive the operation intact enough to start an experimental protocol. That is the population eligible for the trial.

Now the manufacturing begins. Surgeons harvest tumor tissue during the operation and ship it to BioNTech. A sequencing team reads the tumor DNA and compares it to the patient’s healthy DNA. Algorithms scan the differences looking for neoantigens, mutated proteins unique to that tumor. Out of thousands of candidates, the system picks up to 20 neoantigens most likely to provoke a T-cell response. BioNTech then synthesizes a custom mRNA strand encoding those specific neoantigens, packages it in a lipoplex nanoparticle, and quality-controls the batch. The full process takes roughly nine weeks per patient.

During those nine weeks, the patient gets a checkpoint inhibitor to prime the immune system. Then the vaccine arrives, delivered in eight priming doses. Then standard chemotherapy. Then a ninth booster dose of the vaccine. The total treatment cycle runs most of a year.

Even after all that, only half the patients who received the vaccine mounted a T-cell response. The other half did not. No one fully understands why. Their tumor biology, their immune systems, something in the manufacturing, or some combination produced a dose that did not activate. Of the eight responders, six remain cancer-free years later. Two recurred. So the headline “nearly all responders still alive” sits inside a funnel that runs roughly like this: 100 patients diagnosed, 20 operable, 16 in the trial, 8 responders, 6 durable responses.

Every step of that funnel is an engineering problem disguised as a biology problem. Better early detection expands the operable population. Better surgical outcomes expand the tolerable population. Faster sequencing and better neoantigen prediction expand the response population. Better manufacturing and cold chain shorten the window between surgery and first dose. Better biomarkers tell us which patients will respond before we spend nine weeks building their vaccine.

None of these are funded the way a blockbuster molecule is funded. They are picks and shovels. They generalize across every personalized therapeutic intervention in the pipeline. And almost none of them fit inside a traditional pharma P&L.

This is why the timeline feels glacial. Every bottleneck in that funnel is a different company, a different regulatory pathway, a different reimbursement conversation. Big Pharma can solve the molecule. It cannot solve the system. The system requires a different kind of operator, a different kind of capital, and a different business model.

The Model Is Wrong, Not the Science

Personalized medicine is an n-of-one business. Every dose is custom. Every patient is a batch. There is no single molecule to patent in the traditional sense. The value does not sit in the compound. It sits in the system that produces the compound.

Big Pharma does not know how to price a manufacturing-as-service oncology product. Medicare has no reimbursement code for it. Private payers refuse to cover “experimental” approaches without ten years of outcomes data. VCs pattern-match to the blockbuster model and walk away from anything that looks like services revenue.

How many breakthrough therapies never get funded because no one can figure out how to capture rent from them? The answer is most of them. The industry invests in what its business model can absorb. Everything else dies in a dean’s drawer or a founder’s pitch deck.

The Food Parallel

Food is the same story. You cannot patent broccoli. You cannot patent the Mediterranean diet. You cannot patent the well-known fact that most chronic disease traces back to diet, stress, and movement. So Big Pharma builds an $80 billion market around GLP-1s and almost nobody builds a serious business around the upstream cause.

But think about what you can patent. You can patent the sensor that reads metabolism in real time. You can patent the algorithm that translates continuous glucose data, microbiome reads, and genetic variants into a meal plan that works for a specific person. You can patent the logistics layer that matches food to biology for the right person at the right moment. The container of value is the coordination system, not the ingredient.

Poor nutrition drives roughly $1.9 trillion in annual US healthcare spending when you count chronic disease comorbidities. That number is larger than the entire pharmaceutical industry. If you build the operating system that personalizes nutrition and moves those outcomes, you are not competing for a share of a drug category. You are capturing value from the cost disease itself.

A blockbuster drug is worth tens of billions. A platform that rebuilds how diagnostics, food, and care coordinate for every person is worth trillions. The container of value has changed. Most investors have not updated the map.

Why Lilly Is Fighting Over One Amino Acid

Which brings us back to retatrutide and the 40 amino acid line. FDA defines a biologic as a protein of at least 40 amino acids in a linear chain. Retatrutide has 39. Lilly is in court arguing that the linker chemistry and side chain should push it over the line.

Why does one amino acid matter? Money. Under the Inflation Reduction Act, biologics get 13 years before Medicare negotiates their price. Small molecules and peptides get 9 years. Four years of monopoly pricing on a blockbuster obesity drug is worth tens of billions.

The deeper reason is manufacturing. A 39 amino acid peptide can be made by any competent synthesis lab on the planet. A biologic with complex folding and glycosylation is Fort Knox. Lilly is not really arguing about chemistry. It is arguing about who gets to manufacture the molecule and at what price.

The gray market already answered the question. Semaglutide and tirzepatide were compounded into billions of dollars of sales during the shortage. Retatrutide will not wait for FDA approval to show up in injection vials at wellness clinics. The cost to manufacture is a fraction of the list price. The demand is a tidal wave. The only thing separating the two is a classification argument.

Three Things to Watch

1. New business models for personalization. Watch for the first platform that sequences a tumor or a metabolism, designs a therapy or a protocol, manufactures the dose or the food, handles reimbursement, and tracks outcomes as a single vertically integrated stack. The winner is not a pharma company. It is a logistics and data company that happens to ship biology. The same pattern plays out in nutrition. The first firm that connects wearable data, lab testing, food production, and behavior change at scale captures more value than any food brand in history.

2. The Napster of personalized medicine. Gray market peptides, compounded obesity drugs, and decentralized biohacking communities are already running ahead of the regulatory system. They are crude, sometimes dangerous, and undeniably popular. Someone will build the Napster of personalized medicine, a platform that coordinates biology, manufacturing, and demand outside the traditional channels. It will be called illegal until it gets too big to stop. Maybe a bit like Uber and the Taxi Commission. Watch who builds the trust and safety layer first, because that is the company that survives the transition from outlaw to standard.

3. Incumbents that pivot versus incumbents that fight. Lilly is fighting. So are the big food companies staring down demand destruction in ultra-processed snacks. Some pharma CEOs will see the shift and rebuild their companies as platforms. Most will defend the old model until their balance sheets break. The history of creative destruction is clear on this point. In the end, they all give in. The only question is whether they give in from a position of relevance or from bankruptcy.

The speed of change in healthcare has never been about the science. The science is ready. The business model is the bottleneck. The gray market is simply showing us what happens when the business model fails to adapt fast enough.

Subscribe now

The popular reading is that AI replaces middle managers. That is the least interesting interpretation of what happened. Dorsey just made the first large-scale public bet on an organizational architecture that Friedrich Hayek described in 1945 and Claude Shannon formalized in 1948. He does not cite either of them. But the logic is identical.

And the GDP data shows why it matters.

The Chart

Pull up the Maddison Project data on long-term GDP per capita by country since 1810. Plot it on a log scale. Every major economy shows the same pattern.

Rapid growth as industrialization takes hold. Then a gradual convergence toward a steady-state growth rate that sits roughly 1 to 2 percentage points below that of the United States.

Japan surged from the 1950s through the 1980s. South Korea from the 1960s through the 2000s. China from the 1980s through the 2010s. Each hit a growth ceiling and leveled off. The U.S. just keeps compounding at roughly 1.95% annually, decade after decade, for over two centuries. A straight line on a log chart for 200 years.

The standard explanation is that these are copycat economies. They take product innovations developed in the U.S. and optimize production. Japan perfected lean manufacturing. Korea mastered semiconductor fabrication. China scaled everything. Process innovation is valuable. It raises living standards enormously. But it has a ceiling.

Product innovation has no ceiling because it creates entirely new categories of value. The iPhone was not a better Nokia. Amazon was not a better Sears. Tesla was not a better GM. These were new architectures that redefined their markets. System C.

The U.S. advantage in product innovation is not cultural mysticism. It is structural. Weak central authority. Strong property rights for ideas. Deep capital markets. And 250 years of institutional practice at realigning the information layer to let entrepreneurs detect and act on market signals. No king. No industrial ministry. Just a system that lets the person closest to the problem attempt a solution.

The question is not why the other countries plateau. That part is straightforward. The question is what organizational architecture sustains product innovation at scale? What happens when that architecture reaches the end of its useful life? And what replaces it?

System A: One Mind

Before Alfred Sloan, American corporations ran on genius and instinct.

William Durant assembled General Motors by acquiring everything he could reach. He kept the critical information in his head. When the stock market broke in 1920, he was leveraged to the point of ruin. What the du Pont family inherited was a corporation with $84 million in unsold inventory and a cash position nobody could accurately state because the divisions kept separate books.

Henry Ford had the opposite problem. One car, one color, one price point. The knowledge required to run Ford Motor Company existed in one mind. When that mind could not process a changing market, the company stalled.

This was System A. One brain trying to hold all the information. It worked when businesses were simple enough for a single mind to coordinate. The Roman Army ran on the same logic for 2,000 years: nested hierarchies, but still dependent on individual commanders at each node. The whole structure existed to route information through human bandwidth.

System A hit its limit when the information required to coordinate a continental enterprise exceeded the capacity of any single human brain.

System B: The Firm

Alfred Sloan fixed General Motors with two principles. Each division operates semi-autonomously with its own P&L. A central office coordinates what the divisions cannot coordinate themselves. The genius was in the interface: operational decisions went to the divisions, policy decisions went to the center, and the boundary was defined by who possessed the relevant knowledge.

He then built the information system to make it work. Standardized accounting. Uniform financial metrics. Return on investment as the universal measure. Non-overlapping market segments.

In 1921, GM held 12 percent of the market. Ford held 56 percent. By 1927, GM passed Ford. By the mid-1930s, General Motors was the largest corporation on Earth. Every major company copied the model. By the 1960s, it was the dominant organizational form worldwide.

This was System B. The firm as coordination technology. It solved what Hayek called the coordination problem: the relevant knowledge is dispersed among millions of individuals, no single mind can hold it all, and the question is how you coordinate action when the information required to act is scattered across people who cannot easily communicate.

Sloan’s divisional model was a knowledge-routing architecture dressed up as an org chart. Enterprise software is its digital encoding. SAP, Salesforce, Oracle, Workday, Epic. These are not merely tools. They are the coordination layer of American business. When a hospital installs Epic, it is not buying records software. It is committing to System B.

System B powered U.S. GDP growth of roughly 4 percent from the 1940s through the 1960s. It was the organizational engine of the American century.

But System B has a design flaw.

The Filter Problem

Shannon published “A Mathematical Theory of Communication” in 1948. The core insight: every channel has a capacity, and every relay introduces noise. The more intermediaries between signal and receiver, the more information degrades.

Shannon was writing about telephone circuits. The math applies to organizations.

In a System B corporation, a customer complaint reaches a sales rep, gets logged, aggregated by a regional manager, rolled into a divisional report, and lands on the executive committee’s desk. At each stage, the signal loses context. The specific customer, the specific failure, the urgency. All stripped away, averaged, delayed, formatted to fit the reporting template.

By the time the signal reaches the person with authority to change the architecture, it is no longer a signal. It is a statistic.

This is not a bug. It is the architecture. Sloan designed the system to filter. Filtering was the mechanism that made coordination at scale possible. But filtering is lossy. Shannon’s math guarantees it. And the information that gets filtered out is systematically biased toward novelty: the signal that does not fit existing categories, the customer need that does not map to an existing product line, the competitive threat from outside the recognized industry.

Rebecca Henderson proved this in 1990. Established firms fail at architectural innovation not because they lack talent but because their organizational structures are tuned to the old architecture. The channels process old signals efficiently and new signals as noise. The organization does not reject the new signal consciously. It cannot see it. The structure makes it invisible.

Sloan built the filter. Shannon explained its cost. Henderson proved the cost is structural. And the cost shows up in the GDP data.

As System B ossified, U.S. growth dropped from roughly 4 percent to 2.5 percent. The difference, compounded over decades, is the difference between an economy that doubles every 18 years and one that doubles every 28 years. Ten years of compounding, lost. System B is not broken. It is at the end of its S-curve. It has reached its natural limit and is protecting its position with rent-seeking rather than innovation.

Getting back to 4 percent requires a new architecture.

What Dorsey Actually Did

On March 31, 2026, Dorsey and Sequoia’s Roelof Botha published “From Hierarchy to Intelligence.” Dorsey’s argument starts where this essay starts. The hierarchy exists to route information. It was the best tool available for 2,000 years. Each layer adds latency and noise. Shannon’s math at work.

Dorsey’s move is to replace the layers with what Block calls a “world model”: a continuously updated picture of everything happening across the company. Every decision, every customer, every transaction, every bottleneck. In real time. No status update. No weekly sync. No manager to translate what is happening on the ground into language the executive can understand. (BTW I have been building a world model for Food is Health over the last few weeks).

When the world model carries the information, you do not need the layers.

Block now runs on three roles. Individual contributors who build. Directly Responsible Individuals who own specific outcomes for a fixed period. Player-coaches who develop people while still doing the work. No permanent middle management layer. The system handles coordination. The humans handle the work.

In his Sequoia interview, Dorsey said he wants all 6,000 remaining employees reporting directly to him. That sounds absurd in a System B frame. In a System C frame, it is logical. If the world model carries the context that managers used to carry, the span of control constraint dissolves. The bottleneck was never the number of people. It was the information bandwidth between them.

What Dorsey gets right, and most commentary misses: he is not automating what managers do. He is removing the relay from the communication chain. The signal from the customer reaches the person who can act on it without passing through four layers of lossy transmission.

His essay traces the history of organizational hierarchy from the Roman contubernium through the Prussian general staff to the American railroad org chart. Two thousand years of the same architecture, optimized but never replaced. Block is attempting the replacement.

Whether Block succeeds is an open question. Current and former employees say 95 percent of AI-generated code still requires human modification. Regulated financial services limit how much decision-making you can delegate to AI. The experiment may break at scale.

Don’t let it escape you that Dorsey is the rumored inventor of Bitcoin and BlockChain. And the DAO is the logical architecture for System C Managment.

But the direction is right. And the direction is what matters for GDP.

System C: Hayek Meets Shannon

• System A was one mind trying to hold all the information. It hit a coordination ceiling.

• System B was Sloan’s hierarchy. A structured filter that routed information through layers. It solved the coordination problem for a century. Enterprise software encoded it. Management science optimized it. Business schools trained people to operate within it.

• System C is what happens when AI collapses the coordination cost that justified the hierarchy.

Hayek argued that the entrepreneur is the quirky resource who detects a pattern others miss. Not through superior processing power, but through a different way of seeing. There is a reason dyslexia over-indexes among entrepreneurs. Dyslexic thinkers process sentences before words. They see the shape of the problem before the details. They solve by synthesis, not analysis.

Shannon argued that the solution to signal degradation is not a faster relay. It is fewer relays. Compress the chain between signal and action, and you preserve the information that matters.

System C compresses the chain. The customer signal does not need to travel through four management layers to reach the engineer. The patient’s biomarker data does not need to travel through three clinical departments to reach the person who can change the diet. The farmer’s soil data does not need to travel through two supply chain intermediaries to reach the food scientist who can reformulate the product.

AI does not speed up System B. It replaces the architecture. Instead of routing information through hierarchical filters, AI agents connect signal to action directly. The coordination layer that once required massive capital investment and years of implementation can now be assembled dynamically. The novel signal, the cross-domain connection, the insight that does not fit the reporting template, all of it reaches the person who can act on it with context intact.

This is the Sloan model going atomic. And it maps directly to the three derivatives of innovation.

• The first derivative, the rate of experimentation, accelerates because System C removes the organizational friction that slows iteration. When an individual contributor can access the full context of the company through a world model, the cycle time between insight and action compresses.

• The second derivative, the depth of architectural change, increases because the filters are gone. Henderson showed that System B organizations cannot perceive architectural innovation because the filters destroy the signals. System C removes the filters. The organization can finally see what it has been filtering out for decades.

• The third derivative, the creation of entirely new coordination capacity, is where System C diverges from every previous model. When agents coordinate with other agents across organizational boundaries, the firm is no longer the unit of coordination. The network is. An engineering agent at one company negotiates specifications directly with a supply chain agent at another. The organizational middleware between them becomes overhead, not infrastructure.

Shannon’s semantic structure meets Hayek’s distributed knowledge. The network finds the signal. The entrepreneur discerns its meaning. The organization executes without waiting for permission from seven layers of middleware.

The Copycat Question

If AI removes the information friction that gave the U.S. its coordination advantage, do the copycat economies close the gap?

Not quite. AI is very good at processing bulk intelligence. Pattern recognition, optimization, translation. These are signal-processing tasks. An AI model trained on enough data can match or exceed human performance on structured problems.

But product innovation is not a structured problem. It requires what Hayek identified and System C preserves: the ability to detect signals that do not fit existing categories. The network of trust, reputation, shared context, and informal knowledge that lets ad hoc teams form around emerging opportunities. These networks operate on merit rather than authority. They are built over decades of interaction. And they are nearly impossible to replicate by fiat.

China can build faster models. But a model needs a network to operate in. The U.S. economy has been building innovation networks for 250 years. The muscle memory of decentralized coordination. The instinct to form teams around problems without waiting for permission. The cultural tolerance for failure and iteration. These are features of the operating environment, not the AI system.

The countries that move to System C fastest will be the ones that pull away on the GDP chart. The U.S. has the structural advantage because it already operates closest to Hayek’s model. AI does not neutralize that advantage. It amplifies it.

India, Southeast Asia, and Africa will line up over the next century to improve process, just as Japan, Korea, and China did before them. The great copycats will do well against the world. But the gap between product innovation and process optimization persists because the gap was never about processing power. It was about the ability to coordinate innovation in a system with limited central authority.

System C makes that system more powerful, not less. The path to abundance runs through the countries that adopt it first.

Three Things to Watch

1. Watch for the emergence of agent-to-agent protocols. Right now, AI tools are mostly human-to-agent interactions. The real shift happens when agents coordinate with other agents across organizational boundaries without human intermediation. When an engineering agent at one company can negotiate specifications directly with a supply chain agent at another, the organizational middleware between them becomes overhead. The protocols and trust frameworks that enable this will be the railroads of the next economy.

2. Watch the enterprise software vendors. The ones who survive will accept their role as standards infrastructure rather than toll collectors. There is a difference between a standard and a monopolist who controls access to the standard. AI creates the ability to rapidly adapt around standards, building dynamic interfaces on the fly. The vendors who embrace this will thrive. The ones who fight it will join Lotus 123.

3. Watch the total factor productivity data by sector. If this thesis is right, TFP should begin accelerating in sectors beyond tech. The chart of TFP by industry has shown technology and software at the top for decades, with agriculture, healthcare, education, and construction barely moving. System C applied through network-centric coordination should start bending those other lines upward. Not because the technology is new, but because the coordination architecture finally lets the technology reach the edges of the economy where the biggest productivity gains remain.

The disruption I am tracking is not in replacing Salesforce, but in connecting a farmer’s soil data to a food scientist’s recipe to a patient’s biomarker to an insurer’s risk model. All without anyone picking up the phone.

That is the path to abundance. That is how you break 4 percent.

System C is coming.

Presidents are ephemeral. Food, energy, and finance are not

Every few years, Americans get trapped in the wrong frame. They watch a president, a strike, a headline, a tweet, and assume history has suddenly lurched off course. It has not. The actors change. The control systems do not.

The deepest logic of national security is older than any administration. People need calories, fuel, and a medium of exchange. If an adversary can control those three things, it controls growth, politics, and survival. That was true in the ancient world. It was true in the Cold War. It is true now.

The most important weapon in geopolitics is not the one that explodes fastest. It is the one that can quietly determine whether populations eat, heat homes, finance trade, and maintain order. A country can survive a bad quarter. It can survive one reckless politician. It struggles to survive if rivals can control food, energy, and money.

I think most commentary on Iran misses the point. People ask whether Trump is insane, whether this week’s move was tactically smart, whether fertilizer prices will rise, and whether we should have spent the money on healthcare instead. Those are real questions. They are not the first question.

The first question is this: who controls the essentials over the long term?

It’s not the next frontier. It has always been the frontier.

The Oldest Frontier

Postwar American strategy understood this. When the United States emerged from World War II with unmatched productive capacity, it did not just build carriers, bombers, and institutions. It also translated agricultural abundance into strategic architecture. Eisenhower signed the Agricultural Trade Development and Assistance Act in 1954, which laid the foundation for Food for Peace. By 1960, he described food for peace as “our chief weapon in the worldwide war against want and hunger.” Later presidents kept the same logic. Johnson argued that hunger and want, wherever they existed, were enemies of all mankind. Reagan described Food for Peace as a program that expanded agricultural markets and helped turn potential allies into stronger allies for freedom. (World Food Program USA)

You miss the plot when you reduce strategy to one president’s personality. Presidents come and go. They change very little that was not already changing. The strategic DNA remains. The muscle memory of national security is food, energy, and finance. A president can move five degrees left or right. He does not rewrite the underlying physics of state power. Trump matters tactically. He does not repeal the fact that open access to food and energy sits near the center of American power.

The American System

The American advantage was never just force. It was architecture. The U.S. used agricultural abundance, open sea lanes, dollar settlement, technology, and decentralized innovation to make famine less common and markets more open. The point was not charity alone. The point was to create a world in which coercive control over food became harder to sustain. A world with fewer famines is not just more moral. It is more compatible with open markets, private investment, and the kind of decentralized innovation at which the United States is best.

This is the real distinction between America and many of its rivals. The American model, at its best, assumes that growth emerges from distributed intelligence. Farmers, shippers, scientists, engineers, insurers, traders, and entrepreneurs respond faster than a ministry can. Open markets transmit information. Freedom creates variation. Variation drives innovation. That is the superpower. Life, Liberty, and the Pursuit of Happiness.

Other systems often default toward centralized leverage. They think in terms of chokepoints, state tolerated gray markets, command over payment systems, and political control over supply. They do not need to out innovate the United States in every category. They only need to gain enough leverage over essentials to make openness more expensive.

What Iran Actually Means

The Iran question cannot be reduced to whether this week’s move raises oil prices or whether next month’s fertilizer bill goes up. Those matter. But the larger question is whether the United States allows a rival network to deepen control over energy chokepoints, sanctions evasion routes, payment workarounds, and the political leverage that comes with them. The strategic risk of doing nothing is not zero. It is cumulative.

Action is dangerous. Inaction is also dangerous.

The Network Under the Network

BRICS is not shy about talking in the language of food security and agricultural coordination. The bloc’s own 2025 declarations put food security at the center of cooperation, and BRICS officials say the member states account for 42 percent of global food production, 33 percent of agricultural land, and 39 percent of water resources. At the same time, the U.S. China Economic and Security Review Commission warns that China is facilitating Russia’s and Iran’s ability to evade sanctions and withstand economic pressure. Reuters reported this week that Russia’s deputy energy minister said Moscow would keep supplying energy without regard to Western sanctions and offer “mutually acceptable payment practices” with partners.

None of that means every BRICS food initiative is sinister. It means serious states understand that food, energy, and finance are strategic terrain. They understand that if they can build enough parallel capacity outside U.S. visibility and outside U.S. pressure, they gain leverage even if they never fully replace the existing system.

The Strait of Hormuz is the reminder in plain sight. On a typical day it carries about one fifth of the world’s crude oil. In the current Iran war, AP reports that traffic has been sharply constrained while Iran has still managed to export well over 16 million barrels of oil since early March, largely to China, even as broad shipping through the chokepoint has come under pressure. That is what leverage looks like. Not total closure. Selective access. Enough control to raise prices, reward friends, punish enemies, and remind everyone that dependence has a cost.

Trump matters tactically. Strategically, he is operating inside a much older system. This is why I am not especially interested in the daily emotional theater around Trump. I am interested in whether the United States preserves a world in which no bloc can weaponize scarcity over the next twenty years. If you think that sounds abstract, imagine the opposite. Imagine a world where an Iran, Russia, China aligned system can selectively tighten energy flows, tolerate food coercion, and weaken the institutions that keep trade open. “We should have spent this money on healthcare” stops sounding like a complete argument when the alternative is a structurally less free and less stable system for the next half century.

In that sense, food power is healthcare, just on a different time scale. Healthcare treats people within a society. Food security keeps whole societies from breaking down. One prevents disease. The other helps prevent famine, inflation shocks, forced migration, political extremism, and war cascades. Serious states have to price both.

Europe often struggles here, not because Europeans are stupid, but because consensus systems tend to prioritize short-term internal stability over external power projection. That leaves them tempted to free ride on American action while criticizing American tactics. The pattern is old. Someone else bears the cost of maintaining the perimeter. Europe debates process. America gets blamed either way. Fine. That does not change the strategic arithmetic.

The better question is not whether one likes the current president. The better question is whether the long-run structure remains intact: open food markets, open energy flows, resilient finance, and relentless innovation. If those hold, the American system retains its edge. If they erode, daily politics will not save us.

So how should people think about this without getting lost in the noise?

Three things to watch

1. First, watch chokepoints, not headlines: Pay attention to Hormuz, Black Sea grain routes, fertilizer feedstocks, maritime insurance, and shipping access. Ask whether these flows remain broadly open or increasingly selective. The moment essentials start moving through narrow political gates, leverage shifts.

2. Second, watch the network under the network: Pay attention to sanctions evasion routes, shadow fleets, payment workarounds, bilateral commodity deals, and agricultural coordination outside Western visibility. Power increasingly travels through tolerated gray channels, not just official treaties. If a rival bloc can move food, fuel, and money while bypassing pressure, it is building real strategic depth.

3. Third, watch where innovation compounds: The real American superpower is not a speech. It is the ability of open markets to produce better seeds, better logistics, better energy systems, better pricing signals, and eventually healthier food at lower cost. The country that can make abundance cheaper than coercion wins the century. It is no surprise that we chose now to deal with Iran. AI for abundance, protect the flanks of energy, food and finance.

That is the frame.

Not Trump. Not one strike. Not one tweet. Not next week’s fertilizer cost forecast.

The durable contest is over whether the world remains open enough that innovation, not coercion, allocates essentials.

Presidents are ephemeral. Food power is not.

If you want to judge what is happening in Iran, or Russia, or China, or Europe, start with food, energy, and finance.

Subscribe now

Over Christmas break I finally dug in and started using AI coding tools. Since then, the tools have dominated my thinking of what is possible. Good, bad, and complex. Tools to remove friction. Rethinking markets.

Claude is a nootropic of sorts. I vary between thinking this is a Red pill or NZT-48. I can’t quite explain it yet.

If you have used ChatGPT or Claude chat, the coding tools are > 10x that. It creates a cognitive load I have not felt in a very long time. I can’t decided if that is good. It is breakthrough. I think humans and these tools working together will deliver some real breakthroughs. Breakthroughs that remove friction, and free human capacity to be more human. These tools are taking new ideas I have thought about for years, and creating the pathways to make them happen. That is not a new app, but an orchestration of technology, marketing, strategy, etc. around an opportunity that previously to hard to coordinate. The human and artificial superintelligence together.

It will be messy for a while.

What I learned this week….

1. The result of the super-rich wealth tax will be 50 super-rich people the voters still hate.

2. The people with the best models are keeping them to themselves. (See lesson #1)

3. Early adopters of AI are compounding advantage. (See lesson #2)

4. There is really no way for Anthropic, OpenAI or any of the model providers to defend their moats. Far cheaper, open-source models are near.

5. While we approach 140 IQ in AI, for the vast majority of use cases, we don’t actually need higher intelligence. Models that are “good enough” will accelerate and commoditize.

6. The US needs to win the AI race.

7. The cost of WMD and weapons is going to zero. We have known this for many years. Its is now obvious to more people.

8. Demoralized IRGC and Basij squads, who no longer get a paycheck from oilfields, are being killed by special ops and microdrones, with no one to send help.

9. DoD using Anthropic for mass surveillance and autonomous weapons is probably less risky for humanity than the rapidly expanding use of AI for open source intelligence.

10. 40M Americans turn to AI daily for healthcare (Link) Will it will only get better?

11. Statins were the first drug class prescribed based on blood tests instead of symptoms, marking one of the most profitable ideas in pharmaceutical history. Is prevention a similar, but bigger opportunity?

12. GLP-1 will disrupts any industry that involves mass. Transportation, buildings, agriculture…..

13. Chinamaxxing is a social media trend in which young Western users aestheticize everyday Chinese practices such as drinking hot water, wearing slippers, eating traditional foods, and dressing in specific fashions. Driven by TikTok. Hollywood did the same for American culture for the last years.

Share

Some Thoughts

Drug Gray Market - We are active investors in alphfold, multiomics, microbiome, proteomics, etc., because they will reshape the nature of care. To give you a sense of the technology acceleration, a tech guy in Australia successfully sequenced the DNA of his cancer-stricken dog and developed an mRNA vaccine targeting the cancer. Dog cancer is a lot differnt than human cancers, but the point is more about empowering patients to control their longevity.

Lilly’s phase 3 drug, Retatrutide, is currently available on the gray market for $24/week. A peptide produced “Off the books”. AI will further accelerate the ability to explore new targets. More testing by consumers will expand data sets to evaluate target solutions. Better multiomics will expand testing, precision, and understanding. Combined the impact is non-linear.

While drug companies may try to protect their IP, they will also need to rethink their business model to address the gray market. In a sense, this mimics the Napster model, stealing MP3 music in the late 90s. Ultimately, the theft of music revealed new musicians who previously had a hard time penetrating the record labels. It also mimics SaaS in a world of AI. In less than 6 months from now, Antropic and OpenAI will be good enough to replicate Salesforce nearly for free. Unraveling the SaaS market. Tech Guy in Australia has signaled the same unlock for drugs. While people will want the reliability of proven approved drugs, the ability to charge $50k for a drug vs. buying on the gray market for $24, changes the risk-reward calculus.

Is Hormuz about destroying the IRGC? Iran doesn’t run on ideology. It runs on a payroll. Fifty-one percent of oil export revenues go directly to the IRGC and military. Miss payroll and the regime fractures.

During the Iran-Iraq War (1980-1988), the US Navy escorted Kuwaiti tankers through the strait under fire. Iranian mines crippled US warships. By 2000, America imported more than half its oil for the first time in history. Hormuz was a choke point on our own supply chain. War game options in that era were constrained by one hard fact: we needed that oil. Every strategic move had to account for the cost of losing it.

Then fracking changed the physics.

Mitchell Energy cracked the Barnett Shale in the early 2000s. The regulatory unlocks that enabled it were controversial at the time (i.e. Dick Cheney) and read by most observers as industry favoritism. Maybe. But the strategic effect was unambiguous. By 2019 the United States was the world’s largest oil producer. The dependency flipped. Hormuz went from a US vulnerability to an Iranian revenue dependency. The off switch changed hands.

That is not an accident of markets. That is what multi-domain network centric strategy looks like from the outside, long after the moves were made.

I spent time at Boeing Phantom Works in the late 1990s running wargames for DoD. We gamed out 2025 threats, built imaginary weapons systems, and put National Command Authority in the chair. This is not 4D chess. It is simply disciplined long-term thinking about hard problems, the kind that is completely routine inside the defense community and nearly invisible to everyone outside it.

Our strategic goal over Hormuz today is less about oil shipments. It is likely more about destroying the IRGC’s balance sheet. Shut down Iranian exports and you do not just pressure a government. You defund Hezbollah, the Houthis, and Hamas simultaneously. The proxy network runs on the same cash flow as the regime. Cut the revenue and you do not need to defeat these forces in the field. They defeat themselves.

Military, financial, and food power are not separate domains. They are levers on the same system of power. The press and chattering class are stuck in the 80s. Agriculture needs to think about their role in the power dynamic.

AI Apps

→ Idea Browser: Converts Reddit threads and search data into startup ideas with verified demand signals, pairing each concept with execution plans, market sizing, and timing rationale instead of leaving you with a vague prompt and a blank screen. Co-founder Greg Isenberg built it to productize the sourcing workflow he uses to build and invest in companies, which makes it worth a look if you want to see where early behavioral signals are clustering before a market gets crowded. https://www.ideabrowser.com/

Some Charts

It was Inevitable: Iran was rapidly building its ballistic missile stockpile, making it impossible for the U.S. to produce enough interceptors to stop them. If successful, Iran could pursue its nuclear weapons program because the West wouldn't be able to attack Iran without suffering serious damage.

“That is what they were trying to do, is put themselves in a place of immunity where the damage they can inflict on the region would be so high that no one can do anything about their nuclear program or their nuclear ambitions.” - Mark Rubio

Every administration worried this. The tipping point revealed itself. It does make you wish we figured another way long ago.

Hardware vs. Software: Average intelligence is IQ of 100. The best AI models are approaching IQ of 130-140. While smarter feels better, we may be moving to a point of commoditization for most AI applications. Once the software models are easy, the model means less. It becomes a commodity business of selling hardware. Cheap AI models sell Hardware. China is good at profits on hardware.

Peptides: Short chains of amino acids, the same building blocks that make proteins. Your body already produces them as signaling molecules, tiny instruction sets that tell cells to repair tissue, regulate hormones, reduce inflammation, or ramp up growth.

The pharmaceutical model works by patenting synthetic molecules your body does not recognize and charging accordingly. Peptides are different. Many are bioidentical or near-bioidentical, which means they work with existing biological machinery rather than overriding it. That is precisely why pharma suppressed compounding access. You cannot patent what the body already makes.

Reclassifying these 14 peptides does not just expand a product category. It reopens a low-cost, physician-directed pathway to tissue repair, immune modulation, and metabolic optimization that bypasses the $50,000-a-year drug model entirely. If compounding pharmacies can legally produce BPC-157 for tissue healing or Thymosin Alpha-1 for immune function at a fraction of branded drug costs, the longevity and performance market stops looking like a pharma revenue stream and starts looking like a deflationary consumer category. That is the disruption. Not a new drug. A price collapse on biological tools that already work.

Deinstitutionalization: At peak institutionalization, roughly 560 per 100,000 adults lived in state psychiatric hospitals. What replaced it was not treatment. It was the street, and eventually, the prison.

1. In 1941, Joseph Kennedy arranged a lobotomy for his daughter Rosemary, then 23, because her mood swings and mild cognitive disability embarrassed the family. She spent the rest of her life institutionalized.

2. In 1961, Thomas Szasz published “The Myth of Mental Illness,” arguing that psychiatry was not medicine but social control dressed in clinical language. He gave the anti-psychiatry movement its intellectual spine.

3. Ken Kesey’s “One Flew Over the Cuckoo’s Nest” arrived in 1962. It made the psychiatric ward a villain.

4. John F. Kennedy signed the Community Mental Health Act in 1963. It federalized mental health treatment. Shifting away from long-term institutionalization toward community-based care. The community centers it promised were never fully funded. The hospital beds were closed anyway.

5. Ronald Reagan, as California governor in the late 1960s, saw deinstitutionalization as a cost-cutting opportunity and signed the Lanterman-Petris-Short Act in 1967. Making it nearly impossible to hold someone involuntarily. Other states, on bipartisan basis, followed.

6. Patients discharged from state hospitals in the early 1970s flowed onto sidewalks. It was not a housing market, it was an oversupply of disfunction.

7. By the early 1980s, the NY subway system had become a de facto psychiatric ward without staff, treatment, or walls.

8. Prison replaced hospital as the default container for untreated mental illness.

Subscribe now

Real people are dying. Real families are receiving flags. Real communities in Tel Aviv, in Tehran, in Kyiv, in Gaza, in small American towns that still send their sons and daughters to serve, are absorbing losses that no spreadsheet captures. Whatever strategic logic sits behind these decisions, it sits on top of human suffering that we should not normalize or bracket away.

The purpose of my recent essays is not to celebrate conflict. It is to illuminate the decision calculus that serious people actually use when managing it. Because the alternative, pretending these decisions are made emotionally or randomly, leaves us less equipped to understand what is happening and less equipped to plan around it.

The US is Always in Conflict

The United States has been in armed conflict for roughly 230 of its 250 years of existence.

We have not formally declared war since World War II. Conflict is a feature, not a bug, in the American system.

Since the mid-70s, we have stopped sending conscripts into these engagements. We send professionals. People who have spent careers studying, training, and refining the application of deturance and force. Our current military has been in combat longer than most countries have maintained standing armies. The US military’s senior leadership has more accumulated combat hours than any institutional force in history.

The question is not whether the United States will be in conflict. The question is whether we manage that conflict intelligently or expensively. Whether we pay by the month or buy out the contract.

The System Behind the Conflict

To understand the Iran problem, you have to see the full network.

October 7, 2023 was not just a Hamas attack. It was the most visible node in a financial and military distribution system that runs: Hamas and Hezbollah in the Levant, Venezuelan financial infrastructure providing sanction evasion, Iranian central command funding and directing the entire network, and Russian military capability sustained in part by Iranian, China and Venezuelan weapons manufacturing.

They are one system.

Iran has funded Hezbollah at roughly $700 million per year. It has exported Shahed drones to Russia at $20,000 per unit. Ukraine burned nearly $100 million worth of missiles in a single night, battling a Russian barrage. Iran is not just fighting Israel and the United States. It is fighting Ukraine through Russia.

Iran has poured an estimated $20 billion into proxy militias since 2012. That is the investment portfolio of a country with a $7.9 billion annual defense budget. The leverage comes from the network. Sort of an offense layer inside BRICS.

The Houthi campaign in the Red Sea follows the same logic. Every Houthi drone intercepted by a US Navy destroyer costs 100 times what the drone costs to build. Iran does not need to win militarily. It needs to impose more cost. It has been doing that for 42 years.

A review of Chinese military doctrine suggests China has no interest in a kinetic war. They have not fought since Vietnam. They wish to create the threat of war, so the US spends itself into bankruptcy. Since the US values life in conflict so highly, the cost differential from asymmetric threats puts the US in an ever more difficult cost position.

It has been cheaper and cheaper to disrupt every year. In 1998, it cost 20B+ to build a nuke. Today, probably less than 1B. Lots of cheap drones burn up a lot of Patriot missiles. In the late 90s, we saw this coming and decided to lower our cost of conflict through precision and network-centric warfare. More missions like Venzualla, fewer like Gulf War I.

Two Options. Neither of Them Is Peace.

Most pundit analysis fails because it frames the choice as a conflict between restraint and conflict. That is not the actual choice. We really don’t have the option to do nothing. (Notionally, we could walk and let the Barbary pirates re-emerge)

The actual serious choice is between two different conflict strategies, one that pays indefinitely and one that pays acutely.

Option A is the status quo. Keep the containment. Maintain two carrier strike groups in the Gulf permanently. Continue the Houthi intercept campaign. Keep funding Ukrainian defense against Iranian-supplied Russian drones. Continue absorbing periodic escalation, then de-escalating, then absorbing the next one.

The numbers on this path are public. Brown University’s Costs of War Project puts the post-October 7 tab alone at $34 billion, nearly six times the typical annual US military aid to the region. The broader post-2001 Middle East total exceeds $8 trillion. Current deployment costs run approximately $30 million per day. US debt closed 2025 at 122.8 percent of GDP.

This is not a peace strategy. This is a subscription service for a war that never ends. And the interest compounds.

Option B is to accept the upfront cost, remove Iran’s capacity to fund and direct the network, and terminate the ongoing drain.

No more Houthi campaign. No more Red Sea escort missions at destroyer day rates. No more Shahed drones arriving in Kyiv. No more $700 million per year flowing to Hezbollah. No more nuclear brinksmanship every 18 months requiring carrier group deployments as deterrence signals. No more proxy militia funding spreading instability from Yemen to Venezuela to Ukraine.

The upfront cost is real and ugly. Perhaps $50 to $100 billion for the operation and its immediate aftermath. Real lives at risk. A spike that is visible, accountable, and politically costly.

But it is a finite number attached to a problem that gets resolved.

Option A, discounted over 20 years at a conservative run rate, lands somewhere between $250 and $400 billion in present value before counting oil shock drag, shipping disruption, and the economic cost of sustained regional instability. That number does not include the next major escalation, which, based on 42 years of pattern, is not a question of whether but when.

The NPV math is not close.

While Option B could become Option A as it did in Iraq, we have been developing the systems to make that less likely. In part, the strategy of Iraq was to drain the enemy so that the next generation in the Middle East would seek something like the Abraham Accords, instead of conflict. What took 150k troops to get Saddam, we did with 1,000 to get Marduro. Watch the trend not the single data point.

All involve real conflict. Option A and Option B both have human costs.

The Ukraine connection follows Iran

If you want to end the war in Ukraine without direct confrontation with a nuclear power, the rational path is to starve it of cash and weapons. Don’t defeat Russia militarily. Collapse its sanction evasion network. Cut the supply lines keeping its war machine functional.

Russia’s ability to sustain its drone campaign against Ukraine runs directly through Iranian manufacturing. Destroy that manufacturing capacity, and you reduce Ukrainian intercept costs. Increase the cost of conflict for Russia. Collapse the Venezuelan financial infrastructure that Iran uses to evade sanctions, you accelerate the economic pressure on Russia directly.

Modern conflict is less kinetic. It’s as much finance as it is bombs, electronic warfare, and media. Correlated as a network of precision focus.

October 7 opened an operational window. The subsequent unraveling of Hamas and Hezbollah, the exposure of their Iranian command relationships, and the documentation of the Venezuela-Iran financial corridor are not separate events. They are sequential nodes in a network becoming visible.

You do not have to believe the administration has executed this perfectly to see that the underlying strategic logic is coherent. Its not a strategy from the White House. It is a strategy born from 100 Lt. Colonels in the Pentagon thinking a lot. If your goal is to end Ukrainian suffering with minimal kinetics, the rational path runs through Tehran, not more US weapons to Kyiv.

What the Hawks and Doves miss

The public debate is almost entirely conducted in moral terms. Hawks say strength deters. Doves say restraint prevents escalation. Both frames miss the operational question.

The operational question is: what is the NPV of each strategy?

Do the one that costs less.

Forty-two years of “restraint” produced the current situation. Every American base in the Gulf under missile threat. Ukraine burning through Western munitions against Iranian-supplied Russian drones. Red Sea shipping disruption adding inflationary pressure to a global economy already strained by debt. Hezbollah reconstituting. Hamas leadership dispersed but not destroyed.

That is the fruit of the subscription model.

The people arguing for continued restraint need to price it honestly. Not just the moral cost of action, which is real, but the ongoing economic and strategic cost of inaction, which is also real. And it compounds.

Removing Iran’s capacity to fund the network solves the financing layer. It reduces the amplitude of every subsequent conflict. The ideology, the grievances, and the structural instability persist. They require a different toolkit. More Abraham accords, less bombs.

The Honest Reality

I am not here to tell you this is the right decision. I am not the one with access to the intelligence, the operational plans, or the follow-on reconstruction strategy. I did spend several years in a think tank thinking about these things. Reasonable people with full information might weigh the upfront costs differently.

The framing that treats military action as irrational and restraint as costless is wrong. Both options carry costs. Both options carry risks. The relevant question is which cost structure you prefer and over what time horizon.

For 42 years, the United States paid in installments. Small, steady, recurring costs that rarely generate headlines but compound into an extraordinary ongoing drain. Eight trillion dollars. Permanent carrier deployments. Proxy wars on three continents.

That is one strategy. It has the advantage of deniability and the disadvantage of never actually resolving anything.

The alternative accepts a large, visible, politically costly upfront payment in exchange for the possibility of actually closing the account.

Whether this administration executes that strategy competently is a separate question. Whether the underlying logic is coherent is not.

The subscription model has been running for four decades. The balance does not shrink. It grows.

Either choice has moral weight. The decsions is really more about arithmetic.

Carter Williams is managing partner of iSelect Fund and writes Creative Destruction on Substack. He spent the early part of his career at Boeing Phantom Works and has been tracking defense, technology, and innovation economics for 30 years.

Subscribe now

The Trump circus is real. It is also a distraction. I have nothing to add to the present politics. The turmoil is and will result in harm, but diminishes against the larger systems context.

Underneath the noise, a strategic architecture that has been building since the late 1990s is executing. It does not care who is president. Presidents are temporary. Systems are not.

Here is what the system just did…..

The Abraham Accords were a commercial and security realignment built on a simple observation. Arab governments had quietly stopped treating Palestinian statehood as a governing priority. What they cared about was Iran, economic modernization, and regime survival. The Accords said that out loud and formalized what was already operationally true.

UAE normalized in August 2020. Bahrain followed within days. Sudan and Morocco joined within months. Saudi Arabia began open coordination without formal normalization. An architecture that would have taken a generation under traditional diplomacy assembled in months.

Iran read this correctly. A normalized Sunni Arab-Israeli security architecture anchored eventually by Saudi Arabia would encircle the Shia crescent economically and diplomatically. The resistance axis Iran had spent 40 years building would become strategically irrelevant.

So Iran moved.

The Unforced Error

Hamas and Iran made a rational bet using an obsolete model.

They believed Arab street solidarity with Palestinians would shatter normalization. The 1973 playbook. Activate the street, embarrass the regimes, restore Palestinian solidarity as the mandatory Arab political position.

They also read American politics through a specific lens. A fractured Israeli government. A cautious Democratic administration. Progressive pressure on Biden. Ukraine fatigue exhausting American strategic bandwidth. They expected Biden to persist. They did not believe Trump would return.

That was System B thinking applied to a System C moment.

System B is the post-WWII rules-based order. Formal alliances. UN resolutions. Arab League solidarity. American electoral cycles as the binding constraint on action. Iran had navigated that system successfully for four decades.

System C doesn’t work that way. It operates through network logic, capability, logistics, financial flows, and intelligence architecture that has been building for 30 years largely invisible to actors running the older model. 9/11, Afghanistan and Iraq formalized the United States ability to manage the network centric domain.

October 7 was designed to blow up the Abraham Accords. It gave someone else permission to do something far larger.

Planetary Alignment

What followed was not a master plan.

It was a series of decisions that cascaded into structural outcomes no scenario planner fully anticipated. But the cascade was not random. It followed the grain of a strategic architecture that the national security establishment had been building, war-gaming, and pressure-testing for decades.

This is worth being precise about. The US national security apparatus does not operate on four-year electoral cycles the way political commentary assumes. Defense strategy, financial network analysis, alliance architecture, and military capability development operate on 20 to 30 year horizons. The people running these systems in 2024 were shaped by strategic frameworks developed in the late 1990s. They had been monitoring demographic cycles, institutional decay curves, and geopolitical stress indicators long before any politician named them.

They understood that certain windows open. That momentum creates optionality. That the job is not to manufacture the moment but to be positioned when it arrives.

October 7 opened a window.

The Godfather Principle

There is a moment in the Godfather where Vito Corleone, near the end of his life, explains to Michael what he built and why.

He did not build it by being the most aggressive actor in the room. He built it by understanding debts. Who owed what to whom. Which accounts had been accumulating for decades. And when a moment of general disorder arrived, which debts could be settled simultaneously while everyone else was focused on the immediate chaos.

The United States has carried a set of unresolved strategic accounts for decades.

Iran’s revolutionary government has been in open conflict with US interests since 1979. The sanctions architecture built to contain it was always permeable. Russia’s revisionism in Ukraine was enabled partly by its ability to route around Western financial pressure. Venezuela had become a manufacturing and logistics node for both, sitting inside the Western Hemisphere, producing Iranian drone designs and Russian munitions outside the sanctions perimeter. The Shia crescent had grown into a distributed militia network spanning four countries funded by Iranian oil revenue that Western sanctions repeatedly failed to fully interdict.

These were not new problems. They were accumulated debts.

A Fourth Turning does not create the solution to accumulated debts. It creates the permission structure to settle them.

Trump and Hegseth are ephemeral agents in this moment. Competent or incompetent at the retail level, that question matters less than it appears. What matters is that the national security establishment had the target set, the intelligence architecture, and the strategic logic prepared. The political moment provided cover for action that had been constrained by System B norms for years.

Plausible deniability is a structural feature, not a side effect. The next president inherits a cleaner balance sheet and attributes the ugliness to the previous administration. This is not cynical. It is how constitutional democracies manage necessary but uncomfortable transitions.

What Actually Happened

The sequence looks improvised from the outside. From a systems perspective it looks like supply chain destruction executed across multiple theaters simultaneously.

• Syria fell in December 2024 and severed the land bridge moving Iranian weapons to Hezbollah. This was not primarily a US operation. It was a downstream consequence of Hezbollah’s degradation, which removed the ground forces that had kept Assad viable since 2013.

• Venezuela was neutralized in January 2025. Operation Absolute Resolve dismantled what had become a significant node in the sanctions evasion architecture. CAVIM, the state weapons manufacturer, was producing Iranian drone designs and Shahed-style loitering munitions outside the sanctions perimeter. Russia had broken ground on a munitions factory there specifically to produce ammunition beyond the reach of Western sanctions. The shadow fleet oil corridor routing Iranian and Russian revenue through Venezuelan tankers collapsed simultaneously.

• The Iranian nuclear and missile program was struck in June 2025, described at the time as limited. It was not limited in its effects on the enrichment timeline or the ballistic missile production infrastructure.

• Nasrallah was killed, along with most of Hezbollah’s senior command. The organization that had not fought a real war since 2006 discovered its 18-year deterrence mythology was not matched by its actual military capability when tested at scale.

• Khamenei was killed on February 28, 2026, along with the IRGC commander, the defense minister, and the secretary of the Iranian Security Council, all in a single operation.

Each of these was a node. Taken together they represent the systematic dismantling of an adversary’s financial, logistical, manufacturing, and command infrastructure across five years. A cancer dies without food.

What This Demonstrated

The Venezuela operation was a live test of Russian and Iranian weapons systems against top-tier US electronic warfare, cyber operations, and precision airpower.

S-300VM batteries did not work. Buk systems did not work. Pantsir systems Russia rushed in during October 2024 did not work. Iranian Mohajer drones did not work.

Rinse and repeat in Iran right now.

Every customer for those systems watched the results in real time. Russia’s arms export revenue and its alliance credibility took simultaneous hits. China’s military planners, who have been building their own contested-space doctrine around the assumption that layered Russian-style air defense creates deterrence, now have a data set that challenges core procurement assumptions.

The Monroe Doctrine was reasserted without being named. The Western Hemisphere is not available as a staging ground for adversary manufacturing, logistics, or military positioning. That message was delivered kinetically and is unlikely to require repetition.

What Remains Uncertain

The punch list of accomplishments is real. The uncertainty ahead is equally real.

1. The Shia crescent is degraded but not eliminated. Iran’s revolutionary model has lost its founder-level theological authority with no clear succession path. The IRGC retains organizational capacity but lacks the velayat-e faqih legitimacy that gave it political cover for 35 years. Whoever emerges as supreme leader will be starting from a substantially lower base of authority, facing a population that visibly celebrated Khamenei’s death in the streets of Isfahan, Shiraz, and Karaj.

2. The question of who leads the Shia globally is genuinely open. Sistani in Najaf is the most respected living Shia cleric and his quietist model is the ideological opposite of the Iranian revolutionary framework. At 95, his succession question is equally unresolved. A little late for a longevity boost from Peter Attia. What fills the vacuum between Qom’s degraded authority and Najaf’s quietist tradition will shape Shia political identity for a generation. Will it be Kinetic any time soon? Does Pakistan/Afghanistan become the new center? What happens in Saudi’s Shia dominated eastern oil region?

3. Russia’s sanctions evasion network is nearly gone (The BRICs dark web). Venezuela offline. Syria severed. The shadow fleet under active interdiction. Iranian manufacturing nodes being destroyed today. Russian budget math at $60 oil does not support a sustained land war in Ukraine. US National Security now controls most of the oil domain. The question is whether the financial pressure reaches a capitulation threshold before Russian domestic politics produces a leadership transition that reframes the conflict.

4. BRICS is wobbling. The alternative financial architecture that was supposed to provide a dollar-independent settlement system has not delivered. Members are watching two of its most significant participants get systematically dismantled. Is the dollar’s reserve currency position more secure today than it was 18 months ago despite years of commentary predicting the opposite?

5. China is watching all of this and making calculations. Its weapons systems were tested by proxy and failed publicly. Its primary regional partners are collapsing. Its own export architecture is being sanctioned in real time. The internal debate between continued investment in conventional military capability versus accelerating toward ASI is being forced by external events faster than Beijing’s planning cycles anticipated.

The Frame is Network Centric and System Dynamics

The public conversation treats this as Trump foreign policy. Aggressive. Unpredictable. A departure from norms.

It is not a departure from anything. It is the execution of a strategic logic that has been developing inside the national security establishment since the late 1990s.

Trump is acting on the optionality he is given, that explains his departure from his policy.

Network-centric warfare doctrine, developed at the end of the Cold War, was always about more than military coordination. It was about understanding that modern power operates through networks of financial flows, information systems, logistics infrastructure, and command architecture. Disrupting those networks is more decisive than winning individual battles.

The financial network disruption ran parallel to the military operations and largely invisible to public commentary. Shadow fleet interdiction. SWIFT exclusions. Sanctions on precursor chemical suppliers. Cryptocurrency transaction tracking that collapsed the money movement layer underneath the militia funding networks. These were not improvised. They were coordinated across Treasury, State, the Federal Reserve, and allied financial institutions over years.

Precision military action creates a moment of punctuated attention. In the background, the financial architecture gets dismantled. By the time attention returns to the financial layer, the nodes are already gone.

This is not four-dimensional chess. Chess implies a single brilliant player seeing moves others cannot. What this actually represents is institutional momentum. Decades of war-gaming, scenario planning, and strategic coordination across military, diplomatic, financial, and intelligence institutions. Average intelligence applied consistently over long time horizons with respect for momentum beats genius applied episodically.

The three derivatives of innovation apply here as directly as they apply to markets. First derivative is the velocity of action. Second derivative is the convergence of multiple simultaneous operations. Third derivative is the inversion point where the adversary’s system stops recovering and begins collapsing. We appear to be approaching or past the third derivative threshold across multiple adversary systems simultaneously.

What’s past is prologue

Fourth Turnings end. First Turnings begin. Same routine played across 4 generations.

The national security establishment is not thinking about what to do with the current moment. It is thinking about what the first turning architecture looks like and what institutional frameworks need to be in place before the crisis energy dissipates.

That is where the ASI question enters. The transition from network-centric warfare built on human coordination to network-centric architecture built on machine coordination is the defining strategic question of the next 20 years. The debate over Anthropic is not 4D chess, it is irrelevant. The institutions that establish the frameworks for that transition during the current crisis window will shape the first turning the way the post-WWII institutions shaped the last one.

The System leads.

This is not publicly acknowledged. It does not need to be. The people doing this work understand that the best time to establish foundational frameworks is when everyone else is focused on the immediate chaos.

The plans being built today for the first turning were informed by people who spent their careers building the fourth turning strategy. They did not get credit for the strategy. They built it anyway. Because that is what institutional momentum looks like when it works.

The debt is being settled.

The next question is what gets built on the cleared ground.

Subscribe now

Twenty thousand views. Twelve hundred likes. The replies split predictably between people who trust the military and people who don’t, both sides operating from the same Hollywood mental model, the one where defense engineers are either reckless cowboys chasing contracts or cold bureaucrats indifferent to consequences.

I want to replace that mental model with a real one. Not because Warner’s politics bother me, but because the mental model he is feeding is expensive. It costs us clarity at exactly the moment we need it most. And if you are trying to understand where value gets created as artificial intelligence matures, operating from a bad map is a fatal error.

The Pentagon is decades past unsupervised autonomous weapons systems.

I was at McDonnell Douglas in 1989. What I watched being built there changes everything about how you should read this debate.

What We Were Already Doing in 1989

When I joined McDonnell Douglas as a young engineer, part of my technical education involved learning LISP and Prolog. Those are not languages most people have heard of today. In 1989 they were the two primary tools for building expert systems, symbolic AI that could encode human reasoning into software and execute it faster than any human could process the same information.

A Note to regular St. Louis readers: Bob Lozano was my instructor at WashU in 1989, teaching me LISP and Prolog. 10 years later, I was an early investor in his startup that pioneered fabric compute, the first attempt to commercialize what ultimately became AWS-like infrastructure. At McDonnell, we had built early forms of fabric compute in 1992 to run in-silico models for computational fluid dynamics and finite element analysis at night across 1000 HP 720 workstations. This eliminated, for the first time in aircraft development, the need to build physical destruct articles (500M-800M in cost). The ultimate correct compute architecture, though, was a combination of software and hardware, well beyond RISC. NVIDIA and AWS ultimately won that race. Jim Simon was working in parallel at Renaissance Technology, doing very similar work to develop quant investing.

In terms of AI we were not building toys. We were building weapons solutions for the F/A-18.

The system worked like this. The aircraft’s sensors gathered threat data continuously. The expert system assessed the threat environment, identified the optimal weapons response, and surfaced a recommendation to the pilot. The pilot did not click through menus or navigate options under stress at five hundred miles an hour. He accepted or rejected a plan with a single input, hands on throttle and stick, HOTAS, exactly the way a developer today accepts or rejects a Claude CLI plan before executing code. The tech behind this is still novel today, and I left a lot out of my description.

The human was in the loop. The cognitive compression was handled by the machine. And this was production software on a front-line fighter aircraft thirty-six years ago.

Phalanx had been doing something even more autonomous since 1980. The close-in weapons system detects an incoming missile, tracks it, and engages it without any human input at all. The threat envelope is too compressed for human reaction time. A human in that loop would mean a dead ship. Nobody called this a civil liberties crisis for four decades because the system was solving a physics problem, not a political one.

Frankly, when a Marine rifleman shoots at a threat, is that cognition or 1000s of hours of training programmed into muscle memory?

When Senator Warner tweets about autonomous weapons as if they arrived last Tuesday, he is describing a system architecture that has been operational on US Navy ships since Jimmy Carter was President.

The Lineage of AI is more Obscure

AI is not new.

Claude Shannon published his foundational work on information theory in 1948. Shannon was at Bell Labs, but the military was among the primary funders and consumers of that intellectual tradition. The core insight, that information can be quantified, that signal can be extracted from noise with mathematical precision, was immediately recognized as a doctrine for warfare. If information superiority creates force multiplication, and it does, then Shannon’s mathematics was a weapons system before it was anything else.

The United States military absorbed these principles and built network-centric warfare around them decades before any commercial enterprise understood what information superiority meant as an operational concept.

John Nash developed his equilibrium theory under RAND Corporation grants in the early 1950s. RAND existed to solve military problems. The military needed formal models for adversarial decision-making, for war games where each side’s optimal strategy depends on what the other side does. Nash gave them the mathematics. What we now call game theory was built to model conflict before it modeled anything else.

Jim Simons was dismissed from his NSA codebreaking post in 1968 after publishing a New York Times letter arguing America should invest its intellectual capital in innovation rather than war, then did exactly that, applying defense-funded mathematics to financial markets. Building Renaissance Technologies the most successful quantitative fund ever created. The market is just another information environment. The insight wasn’t new. The application domain was.

Anthropic named their AI system Claude. After Claude Shannon. The training architectures that make Claude possible use generative adversarial networks, which are Nash equilibria by construction. Reinforcement learning from human feedback descends from reward modeling that carries the same adversarial optimization logic Nash formalized for RAND.

The military funded the mathematics. Absorbed it into doctrine. Ran it as a competitive advantage for decades. And now an AI company named after the mathematician whose work the Pentagon operationalized in the 1950s is in a public contract dispute with that same institution over whether their terms of service permit the use of AI in weapons systems.

One might reasonably ask whether Anthropic has the military’s permission to commercialize its own intellectual descendants.

WarGames made WOPR the villain in 1983. Joshua plays tic-tac-toe to learn that nuclear war is unwinnable, and the movie treats that as a near-miss catastrophe caused by reckless military automation. What the movie didn’t tell you is that RAND had already formalized that exact logic twenty years earlier under defense contracts. Our nuclear response dynamic, all classified, derives from that game theory. The military wasn’t learning the lesson from Hollywood. Hollywood was dramatizing, badly, what defense strategists had already internalized decades before. The student lecturing the teacher about the dangers of the classroom.

Creative Destruction Applied to Violence

The Hollywood version of American military history runs roughly like this. Overwhelming force. Massive footprint. Collateral damage accepted as a cost of doing business. Dresden. Hiroshima. Vietnam.

The operational reality of the last forty years runs in the opposite direction.

The firebombing of Dresden in 1945 required over a thousand aircraft and killed somewhere between twenty-two thousand and twenty-five thousand people. It destroyed a city. In 2011, US special operations forces located, extracted, and killed Osama bin Laden in a seventy-nine-person operation that lasted thirty-eight minutes. No declaration of war. No collateral damage reported. Precise force applied to a precise target.

That trajectory is one of the most dramatic examples of creative destruction in modern institutional history, and almost nobody frames it that way. Information density replacing physical mass. Precision replacing scale. The same economic force that allowed Amazon to destroy brick-and-mortar retail with better logistics applied to the projection of force.

The Hellfire R9X carries no explosive warhead. Six blades deploy on impact. It is designed to kill one person in a confined space without the pressure wave that kills everyone around them. Passengers in the same vehicle have survived strikes. The precision is that exact. This is the opposite of a weapon of mass destruction. It is a weapon of mass reduction, minimizing harm to the point where a single human being can be removed from a single seat of a single car from ten thousand feet.

The people who built this capability did not do it because they are indifferent to civilian casualties. They did it because reducing civilian casualties was the engineering objective. Military doctrine. All signal. No noise. Precision is not a moral accident. It is a design requirement.

The defense establishment has been the most aggressive adopter of deflationary precision technology on earth. The footprint gets smaller. The accuracy gets higher. The collateral damage goes down. That is the forty-year arc. Warner’s tweet pretends it runs the other direction.

Where Failure Actually Comes From

There is a version of the Boeing 737 MAX story that gets told as a parable about corporate indifference to safety. Engineers cutting corners. Profits over lives. It is wrong in the ways that matter.

The engineers who flagged MCAS were overrun. The problem wasn’t engineering judgment. It was that engineering judgment was removed from the decision. The choice not to build a clean-sheet replacement aircraft because the capital cost and certification timeline were inconvenient for shareholders meant solving an aerodynamic problem caused by a repositioned engine through software rather than geometry. MCAS was an attempt to compensate for a human factors problem through autonomous correction. The concept was not wrong. The sensor redundancy was inadequate, and when that was raised, the people raising it were downstream of a financial decision that had already been made above them. In some ways, a decision needing a 1M token context window, where an average human runs 20k tokens.

Meanwhile, the customers filling those aircraft with pilots who had a hundred hours of flight time instead of eight thousand. The emerging market airlines standing up fleets with fresh crews who had never gained the muscle memory of flying military aircraft, were introducing a human failure mode that more sophisticated autonomy would have caught. The accident chain required both failures simultaneously. A better autonomous system might have broken it.

The lesson is not that autonomy is dangerous. The lesson is that when people optimizing for financial or political outcomes override those optimizing for system integrity, complex engineered systems fail. The engineers who fly on the planes they design have the right incentive structure. The executives who don’t fly commercial have a different one. Small context windows.

HAL 9000 is the cultural reference point everyone invokes when thinking of AI gone wrong. It is worth remembering that HAL didn’t malfunction. HAL was given a contradictory mission by administrators who concealed information from the crew. The AI executed its instructions. The failure was in the humans who wrote the mission parameters, the people above the system, optimizing for a political outcome rather than a coherent one.

The risk in autonomous weapons is not the engineers at Raytheon or the career officers at CENTCOM. The risk is a civilian chain of command that wants systems capable of lethal action precisely because they don’t push back in the situation room. Generals provide friction. They ask about rules of engagement, escalation pathways, second-order effects. An autonomous system directed to skip that friction is a capability with enormous political appeal to exactly the wrong kind of decision-maker.

TIA, the Total Information Awareness program after 9/11, was built for narrow defense purposes.

In a twist of history, TIA was built by Admiral Poindexter, ex-post Iran-Contra, and later evolved Robin Hanson’s work on prediction markets to build DARPA’s future map, which ultimately became PolyMarket. It is one architectural thread of network centric warfare pioneered by Tony Tether at DARPA in the late 90s.

The Patriot Act expanded TIA into dangerous realms. Politicians ordered NSA and others to reveal information about political opponents. The surveillance apparatus that civil libertarians rightly worry about was not built by rogue engineers pursuing power. It was built by engineers following orders from civilian leadership that had decided the political benefits outweighed the constitutional costs. The defense community was the instrument. The civilian political structure, trying to win an election, was the author.

Warner’s tweet performs oversight theater directed at the engineering community while the actual failure mode, political actors directing autonomous capabilities toward domestic or politically convenient targets, goes unaddressed because naming it would require naming specific politicians rather than a convenient institutional villain.

The Classified Room

Serious risk analysis requires two inputs: probability of failure and consequence of failure. You need both to make good decisions.

Anthropic has the best current read on probability. They are closer than anyone to understanding how these systems behave at the edge of their training distribution, where emergent behaviors appear, how close current architectures are to something that surprises their designers. That knowledge lives inside their safety research and their operational data. It is not in any public document.

The Department of Defense has the most complete picture of consequences. What does a misaligned AI system do at operational scale in a combat theater? What does the cascade look like when a system receives contradictory inputs in an ambiguous threat environment? What are the second and third-order effects of an autonomous engagement decision that turns out to be wrong? That knowledge lives in classified operational experience and decades of war game modeling.

Neither institution has both inputs. Neither can do the risk analysis alone. It is a match of perfection that would extend defense and commercial value. A public contract dispute does not close that gap. It advertises it. Every tweet in this thread tells our adversaries exactly where the friction points are in the American AI development ecosystem. China and Russia are not pausing their autonomous weapons programs to read Senator Warner’s mentions. Anthropic’s concern is an unenforceable edge condition whose only value is theater.

The correct governance model is a classified collaboration framework where Anthropic’s safety researchers and DoD’s operational analysts work through the probability-consequence matrix together. Not because accountability doesn’t matter. Because accountability that is structurally incapable of accessing the actual information is theater, not oversight. The public debate about autonomous weapons is happening in a room with no windows.

The Sound Barrier

The specific debate about Anthropic and the Pentagon is a small window onto something much bigger.

Human scientists have spent centuries trying to decode the universe with a three-pound biological brain that evolved to find food and avoid predators. We are the instrument attempting to measure something far beyond our instrument’s range. Our cognitive ceiling is not a limitation we chose. It is the ceiling evolution built when the marginal return on additional brain complexity stopped justifying the metabolic cost.

We hit that ceiling tens of thousands of years ago.

Everything we have built since then, language, mathematics, institutions, markets, the entire architecture of civilization, has been an attempt to extend our individual cognitive range through collective coordination. Shannon’s information theory was an attempt to formalize what the brain does when it extracts signal from noise. Nash’s equilibrium was an attempt to model what happens when multiple bounded intelligences make decisions simultaneously. The F/A-18 expert system running on LISP and Prolog in 1989 was an attempt to offload the weapons-selection cognitive load so a human pilot could focus on staying alive.

The trajectory is one continuous line.

The question of what happens when we build a system that is not bounded by the same evolutionary ceiling is the most important question of this century. Artificial superintelligence, if we get there, will be the first entity in the known history of the universe capable of looking at the fabric of physical reality without our perceptual constraints. Our minds literally cannot perceive certain structures in quantum mechanics. Not because we haven’t tried hard enough. Because the instrument, our brain, is the limit.

Breaking that ceiling is the sound barrier of human evolution. Everything shakes at the approach. The instrument is at its design limit. What waits on the other side is different physics.

The defense community has been living with this question longer than anyone else. They were running the first expert systems on military hardware while the rest of the world was playing Pong. They absorbed Shannon when Silicon Valley was still an orange grove. They operationalized Nash before it appeared in an economics textbook. They have been asking, seriously and at classified depth, what happens when the system exceeds the operator’s comprehension, for four decades.

That is not the community Senator Warner’s tweet was describing.

The question for the investor is not whether this transition happens. It is whether your model of the world is built for the curve that is coming or the one that just passed.

Beyond the Hunter-Gatherer Brain

Venture capital has always been described as pattern recognition. The best investors have seen enough cycles that they can feel a dominant design forming before it becomes obvious to everyone else. The worst investors copy the last cycle and call it a thesis.

Both approaches share a flaw. They treat the innovation curve as something you read with intuition rather than something you measure.

The forty-year arc I have been describing is not intuition. It is calculus. The innovation curve in any sector has three measurable derivatives. The first is velocity, the rate at which new architectural experiments are entering the market. The second is acceleration, whether that rate is increasing or decreasing. The third is the inflection, the moment when latent demand activates at scale and a dominant design begins to emerge. Rebecca Henderson, who shaped my thinking on this at MIT, showed that organizational architecture is the gating function on the first derivative. Clayton Christensen showed that latent demand activation is the gating function on the third. The second derivative is where the investor who can read both forces simultaneously has an edge nobody else has.

This is what network-centric investing actually means. Not a portfolio of bets. A measurement infrastructure. Over ten years our fund has built direct relationships with over six thousand food and agriculture startups, twelve hundred farmers operating at the frontier of precision agriculture, and eighty portfolio companies converting experimental architectures into scale. That network is not a source of deal flow. It is a sensor array. It reads the first derivative in real time across a sector that represents the largest unsolved productivity gap in the American economy.

The sectors with the worst total factor productivity growth in Bureau of Labor Statistics data, healthcare, food, agriculture, have something in common with the military in 1945. They are information-poor environments where physical mass is still being substituted for intelligence. Dresden instead of the Hellfire R9X. The transition is not a question of whether. It is a question of when the second derivative turns positive and the third becomes visible.

The data we have accumulated over a decade, across startups, farmers, and portfolio companies building in parallel, is the foundation for reading that curve. Not predicting it. Reading it. The way Shannon read signal from noise. The way the F/A-18 expert system read the threat environment and surfaced a weapons solution before the pilot’s conscious mind could process the same information.

The biological brain that evolved to hunt and gather is the instrument and the limit simultaneously. It can see patterns in its own experience. It cannot see patterns across six thousand simultaneous experiments happening in parallel without a system that aggregates, weights, and surfaces signal from that complexity. That is what AI does to the investor’s cognitive load. It is the HOTAS system for capital allocation. The decision remains human. The compression is handled by the machine.

This is not a metaphor. The people building expert systems at McDonnell Douglas in 1989 were asking the same question we are asking now. How do you extend human judgment beyond what human hardware can process in real time? The answer then was LISP and Prolog running on avionics hardware. The answer now is a network of agents reading convergence signals across a sector in the fluid phase, surfacing the acceleration curve before the dominant design becomes legible to anyone without the sensor array.

The sound barrier of human evolution is not a technological event. It is an information event. We do not transcend our evolutionary limits by building smarter machines. We transcend them by building systems that read the patterns our three-pound brains cannot hold simultaneously. The next generation of investing is not about picking better companies. It is about having a better map of where the curve is going before the curve itself knows.

The optionality is still cheap. The second derivative is positive but early. The third derivative is approaching.

Three questions worth sitting with this weekend:

If the military funded the mathematics that created modern AI, and AI is now being used to read innovation curves across every sector of the economy, what does it mean that most investors are still making decisions with tools Shannon would have recognized as pre-information-theoretic?

The Hellfire R9X replaced Dresden through forty years of precision engineering, shrinking the footprint while increasing the accuracy of each action. What happens to the industries that still operate like Dresden when the same deflationary precision force arrives at scale?

Every dominant design in history was invisible until the moment it became obvious. The people who could read the second derivative before the inflection point captured the transition. What would you need to see right now to know the third derivative is arriving in food, health, or agriculture before anyone else does?

Creative Destruction covers the innovation forces reshaping food, health, energy, and the broader economy. If you found this useful, share it with someone who would argue with it over dinner.

Subscribe now