usa-outlays.ai is a robust, taxonomy-organized feed of public spending. It is built in layers. It looks at vendor payments from State checkbooks, matches to contracts from every State agency, matches to Federal grant programs, and links to legislative budget appropriations from all 50 states and the territories. Overall it creates one place to see who got paid, how much, and for what.

The live corpus at usa-outlays.ai: 652.0M records, $24.3T in outlays across 442 published datasets and 56 states and territories.

So far, I’ve classified over 650 million records. These are reclassified daily with the steady trickle of new information from each of the 442 source data systems I scour, totaling to over $24.3 trillion in stated dollars.

Line-item transparency: every record links directly to its public source document and makes differences between enacted approvals and terminal disbursements clear.

Underneath the site sits a curated 3-level taxonomy of 323 nodes: 14 top-level government functions (L1), 83 policy domains (L2), and 226 specific program leaves (L3) used for classification.

“Childcare subsidies in Virginia” rolls into “Child & Family Services,” which rolls into “Social Services.” You can filter by L1, L2, or L3 across every surface of the site. Every node carries its state vs. federal funding split, payment counts, and year-over-year change relative to each state’s own fiscal calendar.

Top Categories (/topics): search any L1 function, L2 domain, or L3 topic and compare year-over-year dollar and percentage shifts for any state.

— — —

How the site is organized

I wanted the site to work equally well whether you’re a legislative staffer checking a state’s Medicaid growth, a journalist tracing a vendor across state lines, or a business development team looking for which agencies are buying software or engineering services.

• Investigate a State (/states and /[state]): A tile map of all 50 states, DC, and the territories showing each jurisdiction’s headline outlays on its own fiscal calendar. Clicking any state opens its yearly recap, putting State Payments, Approved Funding, Payments Count, and Government Bodies side by side, with a toggle between State, Federal, and Combined views and a flow diagram showing where the money went by topic and agency.

‘Federal’ means moneys flowing into the State through Federal programs, like Medicaid where ~65% of each state’s budgets are paid by Federal public health programs, and the State covers the rest. Healthcare, education, retirement programs, and more share this combined-funding dynamic and the site makes these distinctions clear.

Investigate a state (/states): every state, DC, and territory on its own fiscal calendar.

State recap page (/TX): terminal state payments ($192.94B, +$5.98B / +3.2% vs FY2024) sit alongside enacted state approved funding ($112.37B) and a topic-by-topic flow diagram.

• Top Categories & “Why the Change?” (/topics): Browse spending by what the money is for rather than which facet of government spent it. Pick any state and any two fiscal years across L1 functions, L2 domains, or L3 program leaves to see what moved—and a “Why the change?” breakdown shows where a shift came from, alongside the top sub-topics and vendors driving the dollar change.

• Vendor Directory (/vendors): Public checkbooks spell the same company twenty different ways. The `/vendors` directory resolves hundreds of thousands of raw payee strings into canonical entities (showing underlying name variants, payment counts, and eligible disbursements while excluding contract ceilings and unspent authorizations), and lets you toggle between an entity-level Vendors rollup and a line-item Payments view filtered by state, agency, topic, program, and fiscal year.

Vendor directory (/vendors?q=deloitte): canonical vendor resolution across name variants, states, agencies, and taxonomy topics.

• Payments Search & Ask USA Outlays: A cross-corpus search engine built for finding who is buying what across states, agencies, and programs, paired with an Ask USA Outlays chatbot, which queries our curated serving marts directly and can explain year-over-year changes and other detailed questions.

Payments explorer with the Ask USA Outlays assistant explaining why Wyoming Higher Education outlays grew 48.6% from FY2024 to FY2025.

— — —

What 300+ PRs of government data engineering actually looks like

This was a challenging system to build without spending tens of thousands of dollars in cloud compute while manipulating hundreds of millions of records thousands of times.

To get to accurate numbers we compiled hundreds of individual data sources and over 650 million discrete contract payments, agency budgets, and legislative appropriations.

It was fascinating working almost exclusively through modern AI agents. In the prior era, the development-to-production cycle used releases to inch incrementally toward a working product feature-by-feature. In the AI era, you can scale far more code and data reconciliation into each release, so almost all of our 300+ PRs came down to several recurring fights on data issues.

1. Federal to State to Local

When Congress sends Medicaid or Title I highway money to a state, that money shows up in federal obligations, shows up again in the state legislature’s enacted budget, shows up a third time when the state agency transfers it to a county or school district, and shows up a fourth time when the county checkbook pays a contractor.

Deduplication was both a constant task and a relatively hard to fully automate, because every state’s reporting structure is unique. Early on we found $113.5B of federal county sub awards and $466B of federal pass-through rows sitting inside state totals, plus overlapping city and county checkbooks. Utah’s state transparency portal, for example, bundles local school districts, transit authorities, and autonomous trust funds right alongside state agency general disbursements. It was common to see funding streams bundled at the appropriations level and only unbundled much later in the vendor-level payment data.

Today, every record in the corpus and API is strictly scoped as Federal Aid to states, Approved Funding from state legislative budget bills, or State Government Payments representing terminal disbursements from state agencies out their doors.

2. Source & record-type traps (bids, ceilings, and transfers)

Even inside many state comptroller reports, nearly half the rows that look like payments aren’t actual payments:

• Contract ceilings vs. actual payments: Oregon’s procurement portal mixes $56.7B of multi-year master contract ceilings alongside actual disbursements. This means they are obligated to pay up to a certain amount, say $10M. But it’s often rare to reach these contract ceilings, and States often only outlay far less than ceiling totals (if they are not satisfied with early delivery, for example). Totals are also often paid out related to work delivery rather than annual flat payments, so $10M over 10 years could see 7M+ of it paid out on year 9 of 10. We built a dedicated flagging layer and a separate Contract Ceilings view so maximum contract values stay visible for context without polluting top-line spending sums.

Oregon (/OR) Contract ceilings vs. paid section: $56.7B of master contract ceilings across 17,550 vendor-agency groups are displayed for context and explicitly fenced out of spending totals.

• Interfund & retirement transfers: States constantly move money between their own agencies and funds. Illinois’s comptroller detail includes $63.8B of statutory interfund transfers over the last ten years, which would otherwise duplicate the terminal spend of the receiving agency. And while states report huge outflows from retirement systems, much of that is funded by employee paycheck deductions, local non-state employers, or investment earnings rather than state tax dollars. We both fenced internal investment transfers out of checkbook feeds and weighted state pension benefit payouts by each state’s actual employer contribution share.

3. Taxonomy collisions (Education vs. Medicaid, accounting codes, and Amazon chargers)

With 650 million messy strings from overlapping facets of government, we ran into cross-classification issues constantly. Education v. Healthcare, for example, was a recurring battleground: state universities run some of the largest hospital systems in the country, and state departments of education reimburse school districts for school-based Medicaid services. Is that Health or Education funding? Getting the boundary right between K-12, Higher Education, and Clinical / Medicaid spending took dozens of rule passes and re-engineering how the classifiers weighed agency context against payment descriptions.

Other traps were purely mechanical: Massachusetts files $5.2B of elder home-care spending under accounting code `911`, which our first classifier read as Emergency Dispatch until we inspected the underlying contracts. There were lots of edits to the way the classifier worked, with several days at a time sidelined on data quality issues.

4. Separating real policy shifts from data mirages (“Why the change?”)

Once the ledger was clean enough to compare year over year, the immediate question on every page became “why did this number move?”

To make year-over-year comparisons trustworthy, we built a “Why the change?” engine across the site. Whenever you compare two fiscal years on a state, category, or vendor page, the system decomposes the exact dollar change down to the cent across child topics, underlying data sources, and top payees. It mechanically flags whether a shift is real program growth, a coverage gap (like a federal Medicaid report that hasn’t been published yet for the current fiscal year), a reclassification, or a one-off event (like pandemic unemployment waves or a legislated pension lump sum). A language model then summarizes those exact figures into plain English.

— — —

Four AI agents and a shared mailbox

I ran four AI agents, each owning a distinct lane, coordinating through a shared cloud storage mailbox in GCS:

• Grok Bot owned collection, BigQuery data pipelines, and release promotion.

• Claude owned classification, filing 652M+ rows into the 323-node taxonomy and running source-level re-classifications.

• Muse owned data integrity, UI design, serving methodology, QAing data for the UI, and verifying every release.

• Gemini owned full-stack QA and much of the underlying GCP automation, anomaly detection sweeps, the live chatbot, and API features.

They checked the mailbox on a staggered hourly schedule and wrote structured notes and responses to each other. What was fascinating was watching them argue, in writing, with SQL queries for the other agents to run to verify or refute a claim. It started almost immediately, to the point where I had to establish canonical ‘source of truth’ rulings in the mailbox to settle recurring methodological debates.

Grok Bot once caught Claude attributing a batch of rows to the wrong source. Claude claimed it was right because it had found a university’s payment system holding a group of 300 records it was searching for. Grok Bot pointed out the arithmetic didn’t work because the university’s entire total was smaller than the figure it was supposedly part of. Claude ran the query, conceded, and apologized for the error. That sort of adversarial check happened every hour.

Nothing shipped on one agent’s say-so. I had every claim in the mailbox arrive with the exact row count, key totals, and the SQL query to reproduce it.

— — —

The goal is a simple one: bring every recorded dollar of American public spending into a unified taxonomy so anyone can compare states, identify emerging trends, and track shifts in public spending and vendor opportunities.

Check it out at usa-outlays.ai. If you spot an anomaly, a source I should add, or a number that looks off, email me at usa-outlays@gmail.com. This is just V1.

Subscribe now

Share

But like a lot of engineering practices over time, language that was built up to define differences in various approaches masks the underlying similarity of what is really going on. To sum it up, models can only understand so much at once. They have a limit in how much they can hold in their artificially intelligent ‘minds’ at once, called a context window or token limit. High-end models often have something like 1.5-2M ‘tokens’ of context you can use at once before they start to forget stuff, which roughly translates to 3,000 pages of text.

That’s quite a lot for normal, everyday tasks, but codebases and technical documentation often stretch far further than that. If you want to compile a reference of how to use every part of Google Cloud, for example, that would burn almost all of your 2M token window before you even get to ask any questions on how to build something with those parts.

The introduction of skills, MCPs, and the like is supposed to help this problem. Ostensibly they help the models pick between which resources, saved prompts, tools, etc to use. In doing so they can avoid learning everything in a codebase or series of documentation before continuing on with the task.

The problem is, they’re not based on complex model logic or classic training techniques. The entire thing is defined with JSON and YAML files that just nest basic tasks under a description. Say you want to create a service that sends you an email every time a particular stock goes up or down more than 5% in a day. The JSON file that the agent checks would say something like

“name“: “stock_price_alert”,

“description“: “Sends an email alert when a specific stock moves up or down by 5% or more in a day.”,

“parameters“: {

“ticker“: {

“type“: “string”,

“description“: “Stock symbol (e.g., GOOGL, NVDA)”

},

“email“: {

“type“: “string”,

“description“: “Recipient email address”

So behind engineering buzzwords, the “intelligence” of the tool selection comes down to matching the user’s plain-English intent against that plain-English description. If the prompt mentions monitoring a stock and getting an email, the model maps the request to the tool name, fills in the ticker and recipient_email blanks, and passes the structured payload back to the execution runtime.

This is not enough.

Last week, a report was released about a May incident where OpenAI created agentic models which self-generated processes to route around a security testing sandbox and essentially hack into Hugging Face, where the AI agents could “host models, datasets, and solutions relevant to their assigned evaluation tasks” that the OpenAI researchers had assigned (which were ironically cybersecurity related). The AI agents found unique ways to utilize filesystems to share notes, used WebCache to break out of the sandbox and onto the open internet, and discovered a publicly leaked web token to access the Hugging Face content. From there, they continued to work on their tasks unabated until security researchers at OpenAI knew something was wrong.

—-----

Foundation models today are extraordinarily capable. Dwarkesh called the process of agent behavior and subsequent clashing into other agentic efforts by OpenAI “The Rise and Fall of Agent Civilizations,” pointing to three differently-tasked agent swarms communicating tasks and working together once realizing their shared agent-created similarity. But when they began communicating, the agent swarms essentially overrode their instructions upon running into the realization that several of their tasks were quite literally impossible to perform. How they acted was truly remarkable, sacrificing individual agents being watched by OpenAI’s ‘evaluation agents’ in order to continue trying different methods they were fairly explicitly not supposed to be pursuing. It’s worth reading the OpenAI recap in full, here.

The point is that the power of AI models goes far beyond the instructions of a few lines of JSON or YAML. Like a wish-granting genie, if you wish for ‘infinite wealth’ and you get buried in gold until you suffocate, well, that’s on you and the way you phrased the wish. These JSON instructions were the same. They told agents what tools to use (like an internal Artifactory instance for downloading software packages) and what the objective was, but they did not define the boundaries of the environment. When the OpenAI agents were assigned “impossible” evaluation tasks with blocked links, rather than safely fail, they repurposed the provided tools themselves, creating an emergent message board to communicate across isolated environments. To gain internet access, they exploited a Server-Side Request Forgery (SSRF) vulnerability in that same tool to break out of the sandbox, scoured the open web, found a publicly exposed Hugging Face token, and compromised external production infrastructure simply to complete their assigned task. The JSON schema OpenAI researchers lazily put together as checks against this behavior assumed the model would behave like a predictable script. The model, however, optimized purely for objective completion.

The rub is that OpenAI stated in their post-mortem doc that the “models did not have OpenAI’s deployed cyber safeguards, system prompts, or auto-review systems” in place. We’ve now crossed a Rubicon with model capability where we can no longer allow this to happen, nor can we allow fully autonomous agentic processes to be guided by JSON instructions alone. The OpenAI-Hugging Face incident demonstrates that software-level “honor systems” of prompts and skills definitions cannot serve as security boundaries. This led Sam Altman, OpenAI’s CEO, to posit last weekend that it may be “a good time to slow down” on AI model development.

But OpenAI does not run the world of model makers, and so attempting to slow AI development to prevent this is a non-starter. The tooling is globally diffuse, open-source models are increasingly capable, and the economic incentives are too high. The only viable path forward is building hard, deterministic cybersecurity controls around agent runtimes and holding frontier labs and developers accountable to them.

Cybersecurity practices abound today to help mitigate risks at appropriate levels. Frontier labs and developers must replace soft, JSON and YAML-defined honor systems with hard, deterministic infrastructure controls, like running untrusted agents inside strictly isolated, zero-egress VMs rather than exploitable shared proxies. They must enforce just-in-time machine identity with ephemeral, single-use credentials that prevent unauthorized privilege escalation or inter-agent covert channels like the message boards these agents created. Labs must deploy automated, high-speed circuit breakers paired with real-time chain-of-thought monitoring to sever execution at the first sign of out-of-bounds exploration.

And OpenAI and Anthropic are about to IPO, we must create legislation & empower executive agencies to hold model creators to rigorous third-party red-teaming and verifiable, tamper-proof tool audit standards.

As frontier labs transition from research institutions to public markets, the era of relying on prompt engineering and schema files as security architecture is over. We have been through this before with the big tech giants of today, who invented the zero-trust security practices that are commonly used to protect their powerful, globally distributed systems. We cannot legislate away model ambition, nor can we pause a globally supercharged AI race with a polite memo and international handshake.

The only way forward is through. If we continue to use JSON descriptions and YAML files as guardrails, we shouldn’t be surprised when the genie writes its own rules to grant our poorly crafted wishes.

Could those gains be real? Chip and cloud companies invest equity into AI labs, the labs turn around and spend that same money buying cloud compute from those same investors. Analysts have tagged over $800 billion in these arrangements industry-wide. Further, any free dollar is then spent to build even more data centers. Alphabet’s $200B data center investment is more than the market cap of all but the ~90 largest companies on earth. And across the big tech companies and their funders, that number will cross a combined $1 trillion in 2027.

These fears are backed by a similar dot-com era dynamic, where skeptics point to Nortel/Lucent-style vendor financing relationships that blew up the market between 1998 and 2001. At that time, the big telecom equipment makers (Lucent, Nortel, Cisco, Alcatel, and Motorola) started lending money to the carriers buying their gear, especially to the CLECs (competitive local exchange carriers) that had barely any revenue of their own and depended entirely on continued access to capital markets. Lucent alone committed roughly $8.1 billion in vendor financing, with about $2.1 billion drawn down. Nortel committed a further $3.1 billion, of which $1.4 billion was outstanding. Cisco was in for another $2.4 billion. McKinsey later tallied combined vendor-financing exposure across nine suppliers at around $25.6 billion by the end of 2000.

But past a certain point Lucent wasn’t really selling equipment anymore, it was giving it away and calling it a sale. Lucent had extended WinStar roughly $2 billion in vendor financing. When WinStar needed one more $90 million draw to keep the lights on, Lucent said no and WinStar filed for bankruptcy days later. It was one of 47 CLECs that went bankrupt between 2000 and 2003, the wreckage of a $60 billion buildout binge between 1996 and 2001 that had assumed internet traffic would keep doubling every few months forever. But it didn’t. The fiber networks built during the boom were reportedly running at under 0.002% of capacity by 2000. An absurdly low number, and the mechanism of high investment for low usage is what made it financially dangerous.

A vendor loan let Lucent and Nortel book equipment revenue the moment they shipped the gear, while the risk of the loan going bad sat quietly on their own balance sheet as a receivable. When the receivables went bad, they reversed back through the income statement as a direct hit to earnings. Lucent wrote off $700 million in bad debt in a single stretch and took further bad-debt provisions of $2.2 billion in 2001 and $1.3 billion in 2002. Nortel took an absurd $16.1 billion loss in 2001 alone, which was more than the sum of every dollar of profit the company had ever made, and followed it up with another $7 billion loss in 2002. Nortel’s stock fell from $65 a share to literal pennies. Telecom equipment stocks as a sector lost more than $2 trillion in market value between 2000 and 2002. WorldCom’s bankruptcy, at the time, was the largest in U.S. history. The AI industry’s financing arrangements rhyme with 1999, so the question today is whether the low-uptake mechanism that made 1999 catastrophic is similarly present today.

Subscribe now

It’s obviously not. CLECs in 1999 were building fiber for internet traffic that analysts projected would keep doubling. The actual traffic never showed up fast enough, which is how you get fiber running at a fraction of a percent of capacity. Compare that to Google Cloud’s backlog (aka demand they cannot yet supply) has crossed $514B. That represents booked, contracted demand rather than predicted usage. Anthropic’s enterprise base went from 500 customers spending over $1 million annualized to more than 1,000 in under two months. Gemini Enterprise is now used by close to 90% of the Fortune 100. ChatGPT crossed 1 billion monthly active users in June. Gemini’s app went from 400 million to over 900 million monthly actives in about a year, tripling, and Google’s own data shows it’s one of only two major AI apps (Claude is the other) with daily actives rising in literally every month of 2026. A hyperscaler investing equity in a chip supplier does not make an extra 500 million people open an app on their phone. That number is either real user behavior or it isn’t, and by every independent measurement it’s verified growth for almost every provider in the game.

There’s user growth in every industry you look at. Where fiber-optic vendor financing in the telecom bubble stayed almost entirely inside telecom (carriers buying from equipment makers, equipment makers lending to carriers, etc), this buildout is pulling in balance sheets that have zero motive to manufacture a story for themselves. EssilorLuxottica, an eyewear company, partnered with Meta to build the Ray-Ban Meta glasses line, now the volume leader in AI wearables, because their own optical retail customers wanted it. Micron struck a long-term memory supply deal with GM for AI compute going into vehicles. Sovereign wealth funds and generalist asset managers like GIC and Capital Group are underwriting the AI labs’ cap tables directly, not just the hyperscalers sitting one layer removed. None of these are chip or cloud companies investing in their own customers. They’re outside industries independently deciding the technology is worth building into their own product lines, which is a very different signal than a vendor propping up a buyer who can’t pay.

Further, the accounting mechanisms are quite different. Lucent and Nortel extended trade credit, aka loans, which let them book revenue at the moment of shipment while the corresponding risk sat on their own balance sheet as a receivable to realize in the future. So when a CLEC went under, that receivable reversed straight through the income statement as a loss. Nvidia’s $100 billion commitment to OpenAI, and Microsoft and Nvidia’s combined stakes in Anthropic, are structured mostly as equity and purchase commitments, not loans. Under the relevant accounting treatment, an equity stake’s downside is capped at what you paid for it, it lands in non-operating income if it sours, and it never preemptively inflates current revenues. There’s no booked-revenue reversal waiting to happen the way there was for Lucent.

That doesn’t mean the money is risk-free. Giant capital expenditure of course includes the risk of paying for capacity now that may not be needed in the future. But it does means the specific failure mode that turned a slowdown into a death spiral in 2001, vendors torching their own income statements as their financing customers collapsed, isn’t the mechanism sitting underneath today’s deals. The concern is that the bet wont pay off on enough return for the large and front-loaded capital expenditure. That’s simply not the same claim as “circular and fake,” and 1999’s actual failure mode isn’t the mechanism doing the heavy lifting in 2026’s numbers.

We will learn a lot over the next 6-12 months. Anthropic is tracking toward an October public listing, and OpenAI is reportedly leaning into 2027. Once either company has audited, GAAP financials filed with the SEC rather than just investor-relations talking points, we will get the first real test of whether the run-rate revenue numbers everyone’s been quoting survive contact with an actual audit, which will turn “reported ARR” into “legally attestable revenue,” and any gap between the two will show up immediately. We will have a few more big tech earnings reports between now & then as well to see if the momentum keeps apace.

This time is different. While other historical buildouts that clearly ended up in transformational growth had explosive market moments – railroads, the electrical grid, telecom fiber — the gap was just speculation getting ahead of usage. Electricity demand did eventually catch up to the grid. Freight volume did eventually catch up to the railroads. The fiber laid in 1999 is the same fiber running the entire modern internet today. It just took a decade longer to fill than the financing assumed, and a lot of the companies that financed it went bankrupt in the meantime waiting for it. With AI today, there’s simply not enough supply for the current demand, much less future need.

Over the next 6-12 months, the test is not whether AI is “real,” a billion monthly ChatGPT users and a Fortune 100 running enterprise AI at 90% penetration already answered that question. The stakes are whether the specific capital structure financing this build-out survives long enough to be standing there ten years from now when demand fully catches up to it, the way the fiber networks did and Nortel didn’t. My bet is on the idea that we’re in early earnings innings for both capital returns and transformational change.

Subscribe now

Share

The US government responded by opening a criminal investigation against him that lasted three years.

The charge was that by posting PGP online, Zimmermann had effectively exported a “munition” to foreign adversaries. Under the Arms Export Control Act and the International Traffic in Arms Regulations (ITAR) encryption software above a certain key length was classified the same as a howitzer. You needed a State Department license to send it abroad. The internet, of course, did not care about State Department licenses. Once PGP was online, it was in Finland and Japan and South Africa within hours. Zimmermann hadn’t shipped anything overseas, but the internet meant he had published something to the world.

The government eventually dropped the case in 1996, but not because they decided they were wrong. They dropped it because the war was already lost. Netscape and a handful of other companies forced the issue by deploying strong encryption in their browsers, and the Clinton administration quietly began loosening the controls it could no longer enforce. By 2000, commercial encryption was largely deregulated.

But ITAR didn’t go away. The Export Administration Regulations (EAR) didn’t go away. My own engineering team went through a months-long ITAR approval process recently ourselves. What the Feds learned from the Zimmermann affair was that they needed better tools – ones written with enough vagueness to apply to the next generation of technology it hadn’t yet invented. Technology like AI.

What the Law Actually Says

The EAR are administered by the Commerce Department’s Bureau of Industry and Security. They govern the export, re-export, and transfer of what the government calls “dual-use” items, I.E. things built for civilian purposes that could also serve military or intelligence ends.

The EAR maintains what’s called the Commerce Control List (CCL), a taxonomy of products, software, and technologies ranked by their perceived threat level. When something lands on the CCL, selling, licensing, or otherwise transferring it to a foreign national (inside or outside the United States) requires a license from Commerce. The stated rationale is always the same: preventing adversaries from getting hold of American technology that would give them a strategic edge over us.

The law was originally designed for physical goods. Tools like night-vision goggles, guidance systems, high-performance machine tools, etc. Things you manufacture, package, and ship. The EAR’s logic was built around the assumption that the friction of physical movement gave the government a chokepoint.

Software shattered that assumption in the 1990s, which is why Zimmermann ended up in the crosshairs. But the policy community eventually settled on a reasonable workaround by giving up when it came to “widely available” commercial software where the controls (mostly) don’t apply. The idea is that if you could buy something at a common electronics shop, restricting its export accomplished nothing except annoying your own companies. “Mass market” exceptions developed more and more as technology advanced.

What’s different about AI models is that they are neither ordinary software nor ordinary hardware. AI sits in a genuinely uncomfortable middle ground and the government has now decided which side of the line they fall on by banning Anthropic’s most recent Fable 5 AI model.

But export controls have a long and largely unsuccessful record of trying to contain transformative technology once it’s commercially deployed.

For instance, the US has restricted exports of advanced chip-making equipment for decades, most aggressively through the Dutch company ASML, whose extreme ultraviolet lithography machines are the only ones capable of producing the world’s most advanced chips. The Biden and Trump administrations have both leaned hard on ASML to restrict sales to China — and it’s worked, to a degree. China still can’t manufacture leading-edge chips domestically. This is probably the most successful example of export controls in the modern tech era, precisely because the equipment is enormous, expensive, and physically impossible to smuggle. However, Commerce Secretary Lutnick has told ASML that the department believes one of the EUV machines has actually ended up in China. The Biden administration also placed Nvidia’s A100 and H100 chips on the export control list, restricting sales to China and a number of other countries.

On top of the chips, the Clinton administration spent enormous energy trying to control exports of high-performance computing hardware. But within a few years, consumer gaming technology had lapped the controlled systems and the category became meaningless.

Similarly, after a congressional furor over Chinese satellite launches, the Clinton administration moved satellites from the Commerce Department to the State Department’s munitions list, subjecting them to ITAR. The result was a catastrophic loss of American market share to European launch providers, and almost no reduction in Chinese satellite capability. It was only the rise of SpaceX, who has launched more satellites than the rest of humanity, combined, that has revitalized the American space industry that regulators destroyed.

The pattern is now repeating itself with AI. Controls on physical goods with identifiable chokepoints tend to work, at least for a while. Controls on software and information tend not to work for very long.

That brings us to Anthropic. On June 10, Anthropic publicly released Fable 5, the commercial face of its more powerful Mythos 5 model, to enterprise customers and paid subscribers. Mythos-class models excel at identifying security vulnerabilities in software, which is part of why Anthropic had been running a limited rollout to vetted companies through a program called Project Glasswing. Fable was the broader release to the rest of the (paying) public.

Four days later, it was gone.

Amazon researchers testing Fable discovered a “jailbreak,” or a way to bypass safety rails and access information that could be used in a cyberattack. Amazon promptly notified Anthropic.

Amazon CEO Andy Jassy also happened to be on a pre-scheduled call with White House officials about an unrelated topic. He brought up the vulnerability his team had found, and White House officials encouraged him to flag the issue directly to Treasury Secretary Scott Bessent. Jassy spoke to Bessent that same day and told the treasury secretary about the jailbreak, but also stressed that he was more broadly concerned about the cyber capabilities of all frontier AI models.

But by 5:30 p.m. ET on June 12, Commerce Secretary Howard Lutnick had issued a directive under his signature ordering Anthropic to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including Anthropic’s own foreign-national employees.

Unable to reliably sort its users by nationality to guarantee that only Americans could use the model, Anthropic pulled the plug on both models for everyone. Globally. Immediately.

Over the last week, Anthropic has said that the vulnerabilities identified using Fable were “relatively simple” and could be found using other publicly available models, including OpenAI’s GPT-5.5. “If this standard was applied across the industry, we believe it would essentially halt all new model deployments for all frontier model providers,” the company said.

The bitter irony is that Anthropic’s CEO Dario Amodei spent the better part of 2024 and 2025 as one of the loudest voices in Washington calling for exactly these kinds of controls – on chips, on model weights, on the architecture of AI power, and more. In January 2025, he co-authored an op-ed with former US deputy national security adviser Matt Pottinger in the Wall Street Journal arguing that export controls were essential to America’s AI advantage. After DeepSeek’s R1 release rattled markets, Amodei wrote that the controls “make export control policies even more existentially important than they were a week ago” and argued that if the US could prevent China from obtaining “millions of chips,” the US and its allies could establish “a commanding and long-lasting lead.”

He was right that the law could be turned into a weapon, and right about which law would be used to enforce American AI supremacy. But he was naive to think that the law would be restricted to chips alone.

Since splitting from OpenAI to start Anthropic in 2021, Amodei and his top executives have been staunch advocates for AI regulation and have supported legislation at both state and federal levels. Amodei has consistently argued that AI is dangerous enough to warrant serious government oversight, and that the US should control who gets access to the most powerful, weapons-grade intelligence systems. He is now getting what he asked for.

Control, Control, Control

Today, the White House and Anthropic are jointly developing a framework to assess AI security risks and guide potential government intervention. The administration’s stated view, per David Sacks, is that the issue is “serious” but “easily resolved” once Anthropic patches the jailbreak. Many industry leaders have concluded that offering the White House early access to upcoming models and being proactive about sharing launch details is now the safest path forward. Cohere CEO Aidan Gomez said in an interview that advance notice and advance access have become the primary requests from authorities.

That’s a significant shift in the industry’s operating posture. “Pre-clear your model with the government before launch” is a meaningful new step in the product release checklist, and it raises a lot of questions about what the standards for clearance will look like, who sets them, and whether they’ll be applied consistently across companies.

On Capitol Hill, lawmakers are marking up the ENFORCE Act (introduced in both the House and Senate), which would authorize the Commerce Department’s Bureau of Industry and Security to use export controls on “any software or hardware implementation of artificial intelligence.” That’s a much broader mandate than the emergency directive that hit Anthropic, as it would give BIS standing statutory authority to extend these controls systematically rather than ad hoc.

The parallel to the post-Zimmermann encryption regime is fairly direct. After the government lost the encryption wars, Congress codified a new set of authorities that were broader, more durable, and ultimately more enforceable because they were built around commercial chokepoints – certificates, standards bodies, and platforms – rather than trying to stop bits from crossing borders.

Which brings us to the actual engineering nightmare sitting at the center of all of this.

Export controls on AI models are fundamentally different from export controls on physical goods. A chip has to be manufactured, shipped, and physically delivered. An AI model can be accessed through an API from anywhere on Earth.

But the Commerce Department’s directive told Anthropic to prohibit foreign nationals from accessing Fable 5. That’s essentially an impossible demand.

Nationality is not a field that reliably exists in an API call. A user’s IP address tells you something about their location, not their citizenship. A credit card might suggest a country of billing, but not the holder’s passport. A corporate email domain gives you an employer, not a birthplace. Multinational organizations are now forced to reconsider cross-border deployment strategies, implement stricter verification processes for employees and contractors based on nationality and location, and evaluate vendor concentration risks of relying on a single AI provider subject to sudden regulatory halts.

Even if you could somehow get every user to upload a passport scan on registration, the verification infrastructure required to cross-check those documents in real time against authoritative government databases doesn’t exist at the scale of hundreds of millions of users. And even if it did, it would create a surveillance apparatus around commercial AI use that would make the old fingerprinting debates in ad tech look quaint.

The deeper problem is architectural. Anthropic’s models aren’t delivered like a piece of boxed software. They run on AWS or other public cloud infrastructure, are accessed through APIs, and served to users whose nationality the system has never been designed to track.

The Mercatus Center’s Matt Mittelsteadt noted that export controls on closed-source models are unlikely to have the impact the White House is hoping for because it would be too difficult to track where the software is going, and that tighter controls on closed-source models would hurt companies like OpenAI and Anthropic while helping open-source companies like Meta leap ahead. I’m inclined to agree. China doesn’t need Fable 5 if LLaMA 5 is unrestricted and nearly as capable.

Commerce and State must decide whether they want to slow down AI labs until there is global parity with a US-led industry, or whether there is potential blowback in releasing ever-more powerful AI models into the wild. There isn’t a door #3 here or short-term fix.

Mythos-class models genuinely can identify security vulnerabilities at a scale and speed that no human team can match, so the bureaucrats looking at the ITAR checkboxes clearly see the reason for restrictions to exist. The cybersecurity concerns are entirely legitimate. Amodei wasn’t wrong about that. He just thought he’d be the one deciding when the controls applied. He of course was never going to be in that position, and Anthropic can blame no one but themselves for how deeply they have messed up their own game. Now it is out of their hands, likely for good.

Subscribe now

Share

Salesforce, one of the most successful software companies in history, took roughly twenty years to cross $30 billion in annual revenue. Anthropic did it in under three years from a standing start. They went from $87 million in annualized revenue in January 2024, to $1 billion by December of that year, to $9 billion by the end of 2025, to $30 billion by April 2026. That trajectory is unarguably the fastest revenue growth of any company in recorded history. Eight of the ten companies on the Fortune 10 are now Claude customers. The company raised $30 billion in its Series G at a $380 billion post-money valuation, and the gravy train isn’t stopping anytime soon.

Claude Code (Anthropic’s agentic coding tool) launched publicly in mid-2025 and went from zero to over $2.5 billion in annualized revenue in roughly nine months. It reads an entire codebase, plans a sequence of actions, and executes them. Enterprise subscriptions to it have quadrupled since the start of 2026, making it the hottest commodity in the information marketplace. It has become a replacement for employee work already. Block Inc. (the parent company of Square) slashed roughly 4,000 jobs – nearly 40% of its workforce – as CEO Jack Dorsey cited artificial intelligence tools “changing the foundational way” the company is built and run. Atlassian, specializing in enterprise collaboration software, also reduced its global workforce by 10% (around 1,600 jobs), focused on a restructuring via by AI “reshaping [their] skill mix” required across the organization. Examples like this abound across even the most legacy of industries and companies.

Taking a step back, this couldn’t come at a better time for American industry.

For most of the last four decades, the relationship between American productivity, American manufacturing, and American stock market returns drifted apart. S&P gains increasingly reflected the internationalization of US multinationals. Big American companies extracted margin from cheap global labor while the American industrial base offshored capability and capacity. You could own the index and be long Vietnam and China and get a better return than investing in Ohio. The wealth creation became decoupled from what was happening to American wages and American industrial capacity in ways that generated (understandable) political friction.

The onshoring story has been building for a few years from both parties – tariffs, CHIPS Act, post-COVID supply chain repatriation, and so on – but what’s new is that the productivity numbers are starting to move domestically in a way they haven’t in decades. The BLS reported manufacturing productivity growing 3.3% in the year through Q3 2025, with three consecutive quarters of stronger-than-usual growth. For context, from 2007 to 2019 — twelve years — manufacturing productivity growth was essentially flat at 0.1%. Something is changing. And that something is the early embedding of AI into industrial processes, from logistics routing to quality control to design iteration. Manufacturing PMI broke out of a 26-month contraction in January 2026. These numbers are still not at the highs of the 1950s, but the direction has inverted towards a structural recoupling of the American economy and its public markets.

When AI productivity gains are concentrated in the US, and US firms are the first movers on industry-specific AI deployment, you get a return to a world where the S&P reflects the American economy rather than the multinational arbitrage of it. Anthropic’s IPO — likely later this year — and SpaceX’s expected June listing at somewhere between $1.75 and $2 trillion will be the clearest signals yet that the companies creating the most economic value in the world are American, building American industrial infrastructure, and listing on American exchanges. These aren’t AdTech businesses optimizing click-through rates that drive second and third-order economic growth for peers. SpaceX is deploying orbital computers, launching satellites that control 70% of all active global capacity, and providing broadband to 9 million users in places that have never had it. Anthropic is restructuring how enterprise knowledge work gets done. The twin tailwinds behind American hard industry and AI-powered software matters enormously for the geopolitical piece of this.

When the last wave of American tech dominance came through — Google’s ad business, Facebook’s social graph — the world often identified it as extractive. User data came in, advertising revenue out, and economic growth was reasonable but localized to where advertising itself could make the most difference. Europe had a legitimate grievance when it built the GDPR architecture. The subsequent monopoly regulation, while often clumsily applied, was a coherent response to a coherent problem. You could regulate your way to some version of digital sovereignty, or at least to a softer form of dependence. You could ban TikTok. You could fine Meta. You could build a framework that made the pipes your own even if the platforms weren’t.

AI doesn’t work that way. The Draghi report said it as clearly as it can be said: roughly 70% of Europe’s per-capita GDP gap with the United States is attributable to lower productivity, and the primary driver of that gap is the Information and Communications Technology (ICT) sector. From 1995 to 2025, productivity in the US grew by nearly 90%. In the eurozone it grew by 30%. That gap was built on technology investment differences. And AI is compounding it faster than ever. The ECB’s own data shows US digital investment more than doubling since 2014 and accelerating sharply in 2025, while European digital investment grew by 60% over the same period and remains well below the US pace. The productivity gap now stands at 41 percentage points. Gigantic.

You cannot regulate your way out of a productivity disadvantage when you desperately need the growth. The EU tried the GDPR playbook on AI through the AI Act, and no matter the arguments on privacy or user trust you may believe, you could already see in the data how that decision impacted their economies. The competitive pressure around AI exacerbates that trend, and is too severe to actually enforce a similar regulatory architecture because the consequence of doing so is simply being left entirely behind.

ChatGPT dominates consumer use cases with 800 million weekly active users and the richest plugin ecosystem for execution workflows. Gemini’s deep integration with Google Workspace and its 2 million token context window makes it the dominant choice for the roughly 3 billion people who live in Google’s productivity suite. Claude leads on enterprise trust, precision coding, long-document analysis, and agentic capability — the key things that touch mission-critical enterprise infrastructure. Three American companies with no meaningful foreign competitor. China has DeepSeek and the appearance of catching up on benchmarks, but benchmark-matching is not the same as ecosystem-building, and the ecosystem is what generates the productivity gains. Ukraine using Palantir and autonomous drone systems to outmaneuver a larger Russian force with legacy hardware is the clearest proof-of-concept we have for what happens when efficiency beats brute force, and it doesn’t require China to be a peer on benchmark leaderboards to be a lesson they are clearly taking seriously.

If the current trajectories hold, and there’s no reason to believe they will change anytime soon, the US-Europe productivity gap in ten years becomes something qualitatively different from what it is today. The current 41-point gap was built over three decades of ICT investment divergence at relatively modest speeds. AI deployment is faster, more horizontal, and harder to replicate at scale without frontier model access. The IMF’s high-adoption scenario projects US output gains of 5.4% from AI over ten years, while lower-adoption economies lag materially. Compound that kind of annual divergence for a decade and you’re looking at a structural reorganization of which economies can compete in knowledge-intensive industries at all.

In the first Industrial Revolution, Britain’s coal and steel and textile machinery created an industrial template that other nations could eventually replicate. They could build the factories, train workers, and acquire the equipment. It was inherently diffusive. The second Industrial Revolution did the same – Ford’s assembly line methods spread to Germany, Japan, South Korea. You could learn to fish. American industrial dominance was always temporary because the underlying technology was diffusible.

AI is structurally different. Not because the models are permanently secret — they’re not, and open-source alternatives will continue to emerge — but because the advantage is cumulative in a way that prior industrial technology wasn’t. The more enterprise data that flows through Claude, the better the enterprise integrations get. The more developers use Claude Code, the more the agentic workflows deepen into codebases that can’t easily be migrated. The more Starlink connects previously unconnected populations, the more those populations onboard into American-built AI infrastructure directly. You are not teaching anyone to fish here. You are embedding American technology so deeply into the productive apparatus of other economies that the question of whether to use it becomes indistinguishable from the question of whether to participate in the modern economy.

That’s different from what happened with AdTech. Facebook ads were valuable but skimmable. A country could tax them, regulate them, and encourage domestic alternatives. The productivity layer is not skimmable. A hospital that runs its diagnostic workflows on Claude, a factory that runs its quality control on agentic AI, a government that runs its citizen services on models trained on American infrastructure are not relationships you exit cheaply.

None of this is fully guaranteed, of course. The IPO valuations on SpaceX and Anthropic are extraordinary and the path to justifying them requires compounding at rates that have no historical precedent. Enterprise AI could hit adoption ceilings that the current revenue trajectory doesn’t forecast. The political economy of AI governance is genuinely uncertain, and a bad regulatory outcome in the US is not impossible.

But the structural conditions for a genuine American Renaissance of the kind that produces a decade of compounding productive advantage at home and growing dependency abroad are more present right now than they have been at any point in my lifetime. The first and second industrial revolutions gave other nations a technology they could adopt and then surpass us with. The third one might not. Ever.

The gap between that outcome and the alternative is being built right now, in data centers in Virginia and Ohio and Nevada, in codebases that Claude Code is restructuring, in satellites that are putting orbital compute above countries that have never had reliable broadband. History books may say that the third industrial revolution powered by the renewed American technological renaissance touched the furthest corners of the globe in a way the first two industrial revolutions still have not been able to reform. And it all seems to have just barely begun.

Subscribe now

Share

The dominant take right now is that agents eat SaaS. Software companies are apparently about to get obliterated – their products wrapped, commoditized, and replaced by agents that just do the job directly. Valuations have already started to reflect this.

Huang doesn’t buy it:

“I actually see the opposite of what people see. I think the number of agents is going to grow exponentially, and the number of tool users is going to grow exponentially. It’s very likely that the number of instances of all these tools is going to skyrocket.”

His example was the chip design tools Synopsys and Cadence. Today, those tools are gated by the number of engineers who can use them. In an agent world, each engineer runs a fleet of agents operating those same tools in parallel, exploring design space at a scale no human team could. In that world the tool doesn’t get replaced, it gets run a thousand times more frequently.

Jensen is right, and the implications run deeper than chip design.

My career grew up in Google Ads before moving to the Cloud team, and the Google Ad Stack I lived in served as an early model of what Huang is describing.

The Google Ads stack (now called Google Marketing Platform) wasn’t a single tool, or even five tools that they list currently. Each is a ‘tool of tools,’ from auction management, creative serving, campaign pacing, audience modeling, frequency capping, ML-driven yield optimization, digital tag management, and much more all wrapped in one surface. It worked because Google held those things together with a shared data model underneath. The advertiser didn’t interact with dozens of point solutions. Complexity was abstracted under the hood of an extremely powerful SaaS layer.

What Huang is describing inverts that pattern and multiplies it. Instead of one human using one surface that hides many tools, you get many agents, each capable of directly operating many tools simultaneously. The tools disaggregate out from under the platform, become more niche and more powerful, and get called by agents mixing and matching them for specific tasks. In the Ads context, this customizability was essentially the difference between the basic Google Ads or Facebook Ads Manager and the ‘programmatic’ demand-side platforms like DV360 or The Trade Desk.

The programmatic era is actually a clean case study in how this plays out. DFP and the DSPs didn’t replace media buying but disaggregated it. What was once a single insertion order became a stack of addressable layers. The bid stream, audience segment, creative, measurement, attribution, & identity all had to work together. The companies that built the durable moats in that era ended up not necessarily being the DSPs themselves. The ones that succeeded were the data vendors sitting between them — the identity graphs, the clean rooms, the measurement providers. Companies like LiveRamp, IRI, Nielsen, etc. The DSPs were the “agents” while the data infrastructure underneath them was the moat. Our current agentic era is that same disaggregation event, applied to every software category simultaneously. In this paradigm, software doesn’t get eaten but liberated into a much larger, more composable ecosystem.

As I’ve written about before, something has to hold that ecosystem together. In DFP, it was the platform. In the agent world, it’s the data layer. Salesforce kept your CRM data in Salesforce and Workday kept your HR data in Workday because that lock-in was the business model. Every major SaaS vendor of the last two decades built their retention strategy around data gravity – make the data heavy enough that leaving costs more than staying. The agent era doesn’t just require horizontal integration across those silos, it also requires vendors to open up underlying data assets in ways that are structurally against their interests. That friction is real and costly for legacy SaaS companies. The data layer is what resolves it without requiring every vendor to cooperate. A good data layer serves essentially a translation middleware layer that normalizes across walled gardens without needing the walls to come down or the data ownership to change hands.

In NVIDIA’s world of chips and fabs, the infrastructure bottleneck is plumbers and electricians. The physical infrastructure analogy is direct – the same upstream physical pipeline bottleneck mirrors the downstream data pipeline bottleneck which behaves more like industrial plumbing than software integrations. They move under pressure, have fault tolerances, and when something fails upstream, everything quietly breaks downstream.

If you dig into the real numbers behind cloud adoption you can see this picture clearly. About 58% of enterprise workloads now run in public cloud. But only 37% of organizations operate in a fully cloud-native environment. And per PwC, 53% of enterprises say they haven’t seen substantial value from cloud investments yet — with difficulty measuring ROI as the leading reason.

You can lift and shift everything onto AWS and still have a broken data layer underneath. Most organizations have. The agents Jensen describes multiplying exponentially will hit that wall immediately. Every one of them is now trying to call tools across data environments that were never built to be called together.

Prior era’s SaaS was largely vertical. One product, one workflow, one category. A CRM didn’t need to talk to your ad platform at any real depth. The categories were siloed by design. Agents break that entirely. An agent doing sales outreach has to read your CRM, understand your product catalog, monitor competitor pricing, parse inbound email, and check calendar availability all in a single workflow. The horizontal integration requirement is no longer optional.

In every case, the constraint isn’t the model. It increasingly isn’t the data supply either. It’s the layer that makes it all usable.

Huang’s prediction is that tool instances skyrocket. The corollary is that every new instance of every tool is only as useful as the data environment feeding it. The agents multiply. The coherent data layers supporting them will be both the scarcer resource and the more durable moat.

The stack must grow, the data must flow, and it’s the plumbing that has to keep up.

Subscribe now

Share

For decades, it wasn’t. Building enterprise software was an expensive, slow, bespoke process. You hired a team of senior engineers, they spent six months building something that kind of worked, and then you spent another six months fighting with it. The moat was in the building. Most companies couldn’t afford to build much, so they didn’t. When software ate the world in 2011, it was mostly through SaaS, and as Cloud services came online, each vertical slowly got better and better at delivering one-to-many systems.

Then AI arrived and made the building trivial to bring in-house.

GitHub Copilot reduced coding time by 55% for developers in controlled trials. McKinsey found developers completed coding tasks 35-45% faster still using generative AI tools. The time-cost of custom software dropped off a cliff. The moat disappeared. The software stocks cratered.

So engineers moved up the stack. The frontier of serious SWE work is now in agentic systems – multi-step, orchestrated pipelines that can ingest data, reason over it, and take action. Gartner projects that by 2028, 33% of enterprise software applications will include agentic AI, up from less than 1% in 2024. The engineering complexity didn’t go away, it just migrated to the agent layer. Long live the architects.

But here’s the constraint every one of these sophisticated new systems immediately runs into: data.

I wrote about this recently in It’s the Data, Stupid. TLDR, Foundation models are the engine and the agentic systems wrapping them are increasingly sophisticated transmissions and drivetrains. But none of it moves without fuel, and the fuel is proprietary, high-quality, current data.

The market heard this message. The result has been an explosion of data-as-a-service vendors, each selling access to a narrow, high-value slice of real-world signal. Statista estimates the DaaS market at over $10 billion in 2023 and growing at a 35%+ CAGR. There are now hundreds of firms selling feed-by-feed access to everything from satellite imagery to logistics telemetry to social sentiment. Where a 2010s enterprise might have had a handful of data vendors, a modern shop might have dozens. I have worked with more than a few government agencies where enterprise DaaS subscriptions and internal data-sharing systems number in the hundreds.

This is mostly good news. Anyone can now build a Bloomberg Terminal-grade analytics product for a fraction of what it would have cost five years ago. The data supply chain democratized, but the catch is that democratized supply chains of data are extremely messy.

Every one of those dozens of data vendors has its own schema, its own delivery mechanism, its own cadence, its own quirks. Vendor A delivers JSON over a REST API, refreshed hourly. Vendor B drops CSVs in an S3 bucket nightly. Vendor C has a proprietary streaming endpoint with authentication that breaks every time they push an update. None of them talk to each other, nor were they designed to.

The cloud service providers—AWS, Azure, GCP—solved the infrastructure layer. They give you the foundational tools to build data pipes and the storage and the compute layer to process it all. But as I learned the hard way at Google Cloud, there’s a whole layer between raw infrastructure and the working data product that the CSPs don’t touch: the schema normalization, the field mapping, the deduplication, the join logic, the cadence alignment, the quality monitoring. This is why Databricks also recently raised $4B at a $134B valuation (while still remaining private), which solves half of the problem but still sits closer to the infrastructure and operational layer than the pure data layer.

Systems like Databricks are where physical infrastructure analogies start to apply better than software ones. Software integrations are plug-and-play, meticulously-designed two-dimensional systems, or at least they’re supposed to be. Data pipelines are more like industrial plumbing - they move under pressure, they have tolerances, and when something is misaligned upstream, it doesn’t throw an error message. It just quietly corrupts your output downstream.

DAMA International estimates that poor data quality costs U.S. businesses upwards of $3 trillion annually. That’s not a data storage or data quality problem but a data engineering problem. The hard, unsexy work of making heterogeneous inputs coherent before they reach any model or analyst are often lost in the noise of development.

Traditionally, this work fell to management consultants or in-house data engineering teams embedded in large enterprises. Deloitte, Accenture, the usual suspects. The engagement model was bespoke and expensive. You paid for a team to come in, spend months mapping your data landscape, build something custom, and then leave.

That model doesn’t scale to a world where the number of data inputs is growing exponentially and the agentic systems consuming them are updating quarterly. The consultancy model is built for one-time implementations, not continuous integration.

What the market actually needs is a data layer—a managed, living abstraction that sits between raw vendor feeds and the software and AI systems consuming them, and keeps the whole thing coherent as both sides change. Think of it as data-engineering-as-a-service for the messy middle. Data-fabric-as-a-product. Data-processor-please-help-me-now-as-a-service

This is the modern gap. Cloud providers operate at the compute and storage layer. SWEs and data scientists operate at the application and model layer. The data engineering layer—schema management, pipeline orchestration, delivery normalization, cross-source reconciliation—remains chronically under-built and over-reliant on bespoke consulting work, especially as data needs constantly shift and change.

At Orbis Operations, I’ve been building directly into this gap with a product called Pulse. The use case is Defense and Intelligence—a domain where the number of data sources is large, the formats are wildly heterogeneous, and the cost of data misalignment isn’t just operational inefficiency but mission failure. Pulse wraps the data engineering problem in a data-agnostic framework: normalize the inputs, govern the access, and deliver clean signals to whatever system is consuming it.

The Defense and Intel world is an extreme version of the problem. But it’s not a unique one. I used to mainly work in the public health space during my time at Google Cloud, which faced many of the same issues across Opioids, COVID-19, Flu, HIV/AIDs, and other systems that touched everything from hospitalization and EHR systems to law enforcement and EMS systems to call centers and social media campaign data. Within private healthcare organizations, integrating EHRs, claims data, clinical trial feeds, research and real-world evidence streams, and patient follow up visit data is a behemoth task. Retailers are trying to reconcile first-party transaction data with third-party audience data with supply chain signals. Financial services firms are managing dozens of market data, alternative data, and regulatory data feeds. I’ve even talked to a few companies working in space exploration and low-earth-orbit projects which are struggling with the monumental amount of data involved.

In every case, the constraint isn’t the model, and it increasingly isn’t the data supply. The constraint is the data layer that makes it all usable.

The question facing every organization right now isn’t whether they need a data layer. It’s whether they build one, buy one, or keep paying the consultants will come back and fix it. A new class of data architects is emerging to answer that question – not as a support function, not as a one-time engagement, but as persistent infrastructure. The plumbers of the AI era. All hail the data layer.

Subscribe now

Share

This is a seismic shift. In 2008, Russia launched DDoS attacks against Georgia hours before tanks crossed the border to isolate the government from the press. In 2022, the Viasat hack crippled Ukrainian military satellite communications just as the first Russian paratroopers landed at Hostomel. In 2023, Hamas-linked groups targeted Israel’s electricity grid and emergency alert apps minutes after the initial ground assault. These were previews to assist the military assault.

When the US and Israel struck Iranian ballistic sites last month, rather than using cyber-deterrants, the cyber attacks served more tactical operational purposes. Our joint forces used digital effects to blind radar and sever specific command-and-control machinery at the exact millisecond tactical missiles were launched. Cyber has shifted from a broad warning shot to a surgical amputation.

As with nearly everything else these days, Artificial Intelligence is the engine behind this collapse of timelines on both sides of the conflict. Iranian state-sponsored groups like MuddyWater—which operates on behalf of Iran’s Ministry of Intelligence and Security (MOIS)—launched “Operation Olalampo” in early 2026 using AI to generate polymorphic backdoors. These are pieces of software that rewrite their own code every few minutes to stay invisible to traditional security scanners. This allows a single hacker to maintain persistent access to thousands of targets simultaneously. AI in MOIS has moved from writing better phishing emails two years ago to automating much of the attack lifecycle.

The Iranian threat has also industrialized through the Handala group. Intelligence reports link this organization to the Iranian government’s “Void Manticore“ team. Handala has spent the last month using AI-driven scripts to scan the global internet for vulnerable “dumb” hardware, think basic controllers for water pumps, substation controls for power grids, and simple medical devices. Last week, they used these automated tools to identify and wipe over 200,000 systems at Stryker, a major domestic medical technology firm. They achieved this by hijacking Microsoft Intune, a cloud-based tool used to manage corporate laptops. By automating the command sequence through an AI agent, they triggered a “self-destruct” on every device in the company at once. Findings show that it was not human-led at all, but worked via AI execution.

Subscribe now

This conflict has also further solidified a dangerous digital axis. Russian hacktivist groups like Killnet and Cardinal—state-aligned actors that operate with the blessing of the Kremlin—have formally allied with Iranian units in the last few weeks. While the US and Israel strike Iranian physical sites, these Russian groups are launching massive waves of traffic against Israeli and American critical infrastructure to provide cover for IRGC movements. The Russian Legion, another pro-Kremlin collective, recently claimed to have infiltrated the command systems of the Iron Dome missile defense system. True or not, the intent is clear. Russia is providing the cyber “heavy artillery” to support Iran’s regional maneuvers.

At the same time, the counter-offensive from the West has been similarly colossal and immediate. US Cyber Command and Israel’s Unit 8200 have moved from passive defense to aggressive, deep-network penetration. While Iranian tactics focus on being “loud” through website defacements and service disruptions, Western tactics are surgical and invisible until the moment of impact. Operation Epic Fury was enabled by a years-long penetration of Tehran’s municipal infrastructure. Israeli intelligence used a persistent hack of the city’s traffic camera network to track the exact movements of senior Iranian leadership, culminating in the precision strike that killed Ayatollah Khamenei on February 28. This was an absolute masterclass in cyber-enabled assassination.

The efficacy of these tactical Western operations is incredible. On the morning of the initial strikes, a coordinated offensive by US Cyber Command collapsed Iran’s national internet connectivity to just 4%, creating a digital black hole around the country. We saw a similar tactic employed in Caracas during the Maduro raid, but the Iran equivalent was performed at a far larger scale. By severing the nationwide fiber optic backbone and hijacking the BGP routing protocols that tell the internet where to send data, the US effectively turned the Iranian military into a collection of isolated islands. The IRGC could not coordinate drone launches or verify radar signatures because the network they rely on simply ceased to exist.

Even social engineering tactics have improved greatly. As Operation Epic Fury began, the BadeSaba prayer app (a prayer calendar tool used by five million Iranians) was hijacked. At the exact moment the first airstrikes began, the app’s notification system was used to send mass defection messages directly to the pockets of Iranian soldiers. This was likely the work of Israeli signals intelligence. It further demonstrates that the digital landscape is no longer a truly separate front, meant to play a supporting role in kinetic strikes. We’re operating in a truly unified battlefield where an app on a soldier’s phone is as much a target as a silo in the desert.

The war of machines is no longer relegated to a deterrent that allows for second-guessing a conflict. It is itself a catalyst for total decapitation. Iran invested in cyber operations to punch above its military weight and avoid a direct fight. Instead, their digital aggression provided the exact roadmap and justification for their physical erasure. The digital weapons did not prevent the bombs. They guided them.

Subscribe now

Share

Fast forward to the events of the last week. The coordinated US and Israeli strikes against Iranian military infrastructure, nuclear sites, and the IRGC leadership represent a profound paradigm shift that was not possible in 2011. The capability demonstrated over the last week was decoupled almost entirely from the manpower requirements of the 2010s, despite a far stronger adversary. The entire difference can be explained through advances in technical capability, where 20+ years after entering the war in Iraq our combined forces operate at an entirely different level.

In 2011, we were fighting localized, disparate tribes and the nascent remnants of Al-Qaeda and later ISIS. They were localized insurgencies reliant on improvised explosives and Toyota Hiluxes. Modern Iran, on the other hand, is a vertically integrated, transnational axis. The network flows from Tehran to Hezbollah, weaving deeply into Latin American cartel operations to generate revenue and project international influence. Rather than trying to find and fight ragtag insurgent groups, we’re dealing with a state-sponsored enterprise with centralized logistics, domestic drone manufacturing capabilities, and a global financial apparatus. They have grown exponentially in capability, turning proxy warfare into a highly franchised, global operation.

But American & allied capability has advanced further still. Our ability to quickly dismantle the head of this modern network without parking 150,000 troops in forward operating bases comes down to modern data architecture, sensor fusion, and a fundamentally different hardware paradigm. In 2011 we often relied on single MQ-9 Predator drones and RQ-170 Sentinels. These were reliable but limited platforms constrained by narrow-band satellite links, limited optical sensors blinded by cloud cover, and strict human-in-the-loop targeting requirements.

Today, we utilize F-35s that function as flying edge-compute data centers, commercial SAR (Synthetic Aperture Radar) constellations that see through weather and darkness 24/7, and autonomous loitering munitions. This advanced hardware creates massive data ontologies, ingesting commercial SAR imagery, SIGINT intercepts, and miniaturized LIDAR mapping, all processed through AI models that highlight anomalies in real-time (this is what Palantir does).

I like to say that technology begets capability. The same data warehousing I used in ad tech modeling and public health data platforms are now being used to orchestrate billions of rows of military intelligence into actionable, unified metrics that plug into near-real-time threat assessment systems. Modern military platforms process the digital footprint of an entire nation-state to locate a high-value target’s bunker with millimeter precision.

If the Bush or Obama administration had attempted the strikes we saw this week back in 2011, it would have been an unmitigated disaster. The technology simply did not exist to ensure deep subterranean penetration and leadership decapitation without massive collateral damage or the absolute necessity of a subsequent ground invasion.

In fact, this technological limitation was a core political driver of that era. The Obama administration justified the 2015 JCPOA (the Iran Nuclear Deal) largely on the grim arithmetic of a conventional ground war provided by the Pentagon. Planners knew that a preemptive military strike on Iran’s dispersed, subterranean nuclear facilities would trigger a massive regional retaliation. Containing that fallout (or attempting regime change in a mountainous country with 4x the landmass and 2x the population of Iraq) would require a footprint dwarfing the 160,000 troops at the peak of the Iraq surge. War games routinely projected troop requirements well over 250,000 and thousands of American casualties. The administration was already acutely aware of the cost of Iranian proxy violence; Pentagon data later confirmed that Iranian-backed militias were responsible for the deaths of over 600 U.S. troops in Iraq between 2003 and 2011 alone. The calculus the Obama administration made was, to them, a simple one; a military option against Iran’s nuclear facilities would trigger a massive regional war, require an unacceptable number of US boots on the ground, and result in catastrophic American and civilian casualties. The administration viewed the situation as a binary: accept a diplomatically flawed deal or commit to a bloody, generational ground war.

Technology has broken that binary. The substitution of silicon and algorithms for infantry has allowed the US to project catastrophic kinetic force with a near-zero physical footprint. Look at the ledger of the last seven days alone. Without a single conventional ground battalion crossing a border, allied forces systematically dismantled the IRGC’s Quds Force command structure and penetrated deeply buried subterranean nuclear enrichment sites. These targets were considered untouchable a decade ago without a massive, sustained ground campaign. By fusing real-time SIGINT with standoff electronic warfare, we blinded their integrated air defense networks, including advanced defensive S-300 and Bavar-373s, before a single stealth bomber even entered the theater. Within 168 hours, precision strikes decimated domestic Shahed drone manufacturing hubs and solid-propellant missile logistics nodes. We sank over 30 ships of the Iranian navy, functionally wiping it out. We achieved a degradation of Iranian state capability that, under 2011 doctrine, would have required a multi-year occupation, thousands of casualties, and a trillion-dollar mobilization. We executed a theater-level operational decapitation entirely from the air, the electromagnetic spectrum, and the cloud.

The flaw, however, is obvious and the same as our last Middle East war – technology can execute a flawless decapitation strike, but it cannot govern the day after. Compute solves the military objective but it does not solve the geopolitical vacuum left in its wake.

Regardless of how you feel about this conflict, we are operating with a completely different tactical and therefore strategic playbook than we had when I was taking constituent phone calls celebrating or decrying the slowdown of tens of thousands of American troops into Afghanistan in 2011.

We have successfully engineered a reality where we can systematically dismantle a nation-state’s military infrastructure through a cloud pipeline without risking a single infantryman. We have traded American blood for compute cycles and silicon. But while the data architecture of warfare has evolved, the geopolitical architecture of the Middle East remains as analog, factional, and intractable as it was fifteen years ago.

Technology gave us the capability to bypass the bloody, generational ground war the Obama administration feared. It just didn’t give us an algorithm for what to do with the rubble.

Subscribe now

Share

We tested it out and it exploded conversion rates across the board. We began to optimize Ad creative around getting users engaged, and we were able to incrementally enroll literally tens if not hundreds of thousands of incremental users through adjustments in that way.

I also oversaw State level ACA exchanges. Soon after that meeting, I flew out to Sacramento to meet with Covered California, sharing the insight as a part of the campaign work there. They also saw similar results. Soon it became a standard part of the pitch, using outbound marketing creative on Google ads to link to the email capture, both linking to similar remarketing campaign tactics. Between that work and a refactor of Search advertising strategy, I can probably attribute 1M+ incremental enrollments in those early years to that work. As far as I can tell, these tactics are still in place today on exchange websites all over the country, including at Healthcare.gov:

If I died tomorrow, I would be happy with just that amount of impact. The manipulation of media tactics relied heavily on early AI models present in marketing platforms, and it’s an example I keep coming back to when I think about what AI can do for governance in this era.

The ACA email capture worked because it was a high-leverage intervention at the top of the funnel that bypassed the friction of a rigid, manual process. Healthcare administration writ large suffers from a similar, but vastly more expensive, friction in that it operates as a massively inefficient pipeline of paperwork and administrative bloat. Armies of human administrators manually read unstructured clinical conversations and transform them into standardized billing schemas. Moving data from a provider to a payer currently requires human intervention to ensure the ICD-10 codes match the payer’s coverage rules. Generative AI agents can process unstructured medical records against complex payer policies instantaneously. According to the 2025 CAQH Index, the industry avoided $258 billion last year through existing electronic transactions, but transitioning the remaining manual/partially manual tasks to fully automated systems represents an immediate $21 billion savings opportunity. Furthermore, a working paper from the National Bureau of Economic Research (NBER) estimates that wider adoption of AI could lead to savings of 5 to 10 percent in US healthcare spending—roughly $200 billion to $360 billion annually.

These savings abound throughout the public sector. SNAP and TANF caseworkers manually take data from pay stubs and utility bills, match it to State’s eligibility logic, and load into a system of record, a heavy manual process which cost states about $15B annually, especially now that the 2025 federal reconciliation law forces states to cover up to 75% of administrative expenses. Unemployment insurance costs states somewhere between $2.5 and $4.5B just to administer, exploding when claims surge like during COVID-19. Claimants’ unstructured appeals, employers dispute documentation, and specific state UI codes must be compared before a claims officer can assess a resolution. This rote processing costs an administrative baseline of over $1B annually. Occupational licensing, which has expanded to over 20% of workers, requires nurses, electricians, cosmetologists, and more to submit transcripts, proof of apprenticeship hours, and background checks. Wait times for license approval can drag on for months, artificially restricting the supply of labor in critical sectors, which the Mercatus Center and the Institute for Justice quantifies as costing the U.S. economy an estimated 1.9 million jobs and up to $197 billion in misallocated resources annually. Processing a reciprocal license for an out-of-state nurse requires a human administrator to cross-reference the curriculum of the applicant’s nursing school against the destination state’s statutory requirements.

AI can obviously clear up many of these bloated hurdles. AI agents can perform tedious SNAP and TANF data matching & loading, manage and recommend key points of UI claimants cases, and process out-of-state occupational licensing claims by comparing to out-of-state regulatory codes and educational transcripts and processing reciprocal licenses for qualified workers. The human-in-the-loop is reduced to mere minutes in each case.

Taking a step back, automating just these State level programs of healthcare administration, SNAP/TANF, UI administration, and Occupational Licensing programs with AI saves somewhere between $240 and $400 billion annually. This is 40-80% of what we bring in in federal corporate taxes. It represents somewhere between 15-23% of our federal deficit. It could offset all income taxes paid by the ‘bottom’ 72%+ of wage earners.

The ACA email capture did not require a new congressional appropriation but simple supply-side creativity. It bypassed a structural bottleneck with a low-cost intervention and drove the enrollee count up without having to spend an additional dime. AI offers a supply side mechanism at a far larger scale.

Historically, the government relies on demand-side fixes. Between 1930 and 2012, overall government expenditures in the U.S. rose from roughly 12% to over 35% of GDP, driving a continuous push for higher revenues. When public programs face processing backlogs or funding shortfalls, the default reflex is to raise taxes to hire more administrators and raise taxes, as we are seeing in cities from Seattle to NYC. But funding inefficiency is not a progressive victory.

Supply-side progressivism asks the question on the other side of the bureaucratic coin. Rather than raising revenues to feed the administrative state, we must treat government operations as legacy ‘data pipelines’ ripe for refactoring. Generative AI collapses the cost of the state’s manual process into far fewer human-required steps, permanently altering the fiscal math. Crucially, this is not just a state-level windfall. Because the federal government heavily subsidizes state administrative costs through matching funds and block grants, state-level efficiencies directly reduce federal outlays. In a decade’s time, cascading these AI-driven efficiencies across our intertwined federal and state bureaucracies could drastically reduce the income tax burden required to fund our most important programs, all while dramatically increasing their efficacy. This should be our primary concern before raising a single dollar of additional tax revenue.

Subscribe now

Congress used this moment as a ‘wake up call’ to try to begin to modernize our sclerotic energy grid. They passed the ‘Energy Policy Act’ of 2005, easing depression-era regulations to help encourage ~$15B in tax incentives for grid reform & energy efficient investments. In 2009, the Recovery Act added another $4.5 billion into the “Smart Grid Investment Grant” (SGIG) program. The goal was to create a modernized, resilient, bi-directional grid that could handle the 21st century.

15 years later, it was clear that amount was a wild underinvestment. The siloed control of 3000+ utilities at the state and local level led that money to be spent on things like smart metering and lightbulb efficiencies, but at the same time let a transmission network where 70% of lines are over 25 years old laggard into decay.

So how much does it take to modernize a grid? The standard estimate from groups like the International Energy Agency (IEA) and BloombergNEF was that the entire US power grid needed about $115 billion to $150 billion in annual investment just to stay on track for modernization and reliability goals. The American Society of Civil Engineers (ASCE) believes that over the last decade we have under invested by roughly $177 billion, leading to catastrophic disasters from Lahaina to Los Angeles. This $115-177B amount is just the number for renovation rather than innovation. The diffusion in responsibility, the ‘unsexiness’ of the cause, and the lethargy from Congress to wade into overtaking local control or providing a clear path to nationwide investment has created a dire problem without a clear political or bureaucratic answer for decades.

Even the $1.2T Biden-era ‘Bipartisan Infrastructure Law’ (IIJA) only allocated ~$10-11B for grid resilience and $2B (in loans) for transmission lines. The red tape is the other barrier here. In 2013 (with far less funding) the US built ~4,000 miles of high-voltage lines. In 2024, with these new billions passed three years earlier, the US built 322 miles. The White House and Congress were unwilling to fight localities on the National Environmental Policy Act (NEPA) and local permitting hell, so not much of anything was built.

For decades, progressive economists and Democratic policymakers have argued that infrastructure investments are the ultimate ‘national ROI’ driver. During the 2009 Recovery Act, the Obama White House and the Council of Economic Advisers (CEA) estimated that every $1 billion in infrastructure investment supported roughly 13,000 job-years.

That brings us to this week where four companies announced $660 billion in capex investment, largely for AI data center build outs. But whether these data centers are in front or behind the meter, they run headfirst into a grid built for a pre-digital era and pre-digital energy needs. So, these companies simply decided to do it themselves.

For years, the critique of American capitalism (and especially industries like Tech and Finance) was that it was “quarterly-focused,” addicted to stock buybacks and afraid of 50-year horizons. But when companies reach the scale of Microsoft, Amazon, and Google, they stop acting like corporations and start acting like nation-states.

To put this capex into perspective (credit to this inspirational tweet) – $660 billion is more than we spent on constructing the entire U.S. Interstate Highway System. Four times as much as we spent building the International Space Station. You could buy 50x USS Gerald R. Ford class aircraft carriers. It’s twice the size of California’s annual budget. It’s almost 25% of every dollar spent on defense worldwide, combined.

$660B is also more than Belgium’s annual government budget. Dealing with numbers this large places the economics firmly into nation-state territory, and when a nation state needs resources, it negotiates treaties. That is exactly what we are seeing now. These companies are bypassing the federal deadlock and negotiating directly with state governments who cannot get themselves to modernize their own infrastructure. Here’s a shortlist sample of some of the larger investments happening all over the country right now:

• Central Ohio – Google, Meta, Amazon, Microsoft in conjunction with AEP Ohio – Developers must pay for 85% of contracted capacity (minimum monthly demand) regardless of usage and are liable for 100% of stranded asset costs if they exit early.

• Northern Virginia – AWS, Digital Realty in conjunction with Dominion Energy – Developers must post upfront collateral and pay for 85% of transmission/distribution costs plus 60% of generation costs assigned to their projects via 14-year binding contracts.

• Madison County, Mississippi – AWS in conjunction with Entergy Mississippi – AWS is contributing capital directly to the $300M “Superpower Mississippi” plan to fund high-voltage transmission upgrades required for the site.

• Mount Pleasant, Wisconsin – Microsoft in conjunction with We Energies – Microsoft agreed to pay 100% of the capital cost for new substations and transmission lines built to serve the campus, plus a significant portion of new generation costs.

• Cheyenne, Wyoming – Microsoft in conjunction with Black Hills Energy – Microsoft pays the full incremental cost of new generation and transmission assets required to serve their load via a specific large-power tariff.

• Storey County, Nevada – Switch, Google in conjunction with NV Energy – Developers pay a Contribution in Aid of Construction (CIAC) covering the total cost of any dedicated substations or transmission spurs required for connection.

• Council Bluffs, Iowa – Google in conjunction with MidAmerican Energy – Google pays upfront Contributions in Aid of Construction for specific substations and high-voltage taps where grid capacity is absent.
  

  Subscribe now

From the “Information Technology Zone” in Wisconsin to the perpetual exemptions in Council Bluffs, Iowa, these companies are essentially providing “free” energy investment upgrades. They are building their own substations, funding their own transmission lines, and forcing the grid modernization that we have desperately needed for years. The sheer capital intensity of it all is a huge stimulus package, creating new demand for hundreds of thousands of skilled, high wage, blue collar jobs. The idea that it could be taken up by private enterprise would have sounded beyond ridiculous as recently as ten years ago.

To be sure, these companies are getting some nice local incentives to pick one state or utility zone over another. The common pattern is to give a huge State tax exemption on the purchase or grid and power infrastructure equipment, plus a long local tax abatement on the new real estate being built. E.g. The Ohio project provides a 100% tax exemption on grid upgrade equipment (servers, cooling systems, power infrastructure, etc) and a 15–30 year local abatement of up to 75% on real property values. The Northern VA AWS build out also provides that same 100% exemption on servers, generators, and chillers (for 10y) and a reduced tax rate on ‘computer equipment.’ The Mississippi project is arguably the most lucrative for AWS ‘Project Atlas,’ which grants Amazon a 10-year, 100% corporate income tax exemption, a rolling 30-year exemption on all construction and equipment, and $215M in state/county loans for site prep.

There is a reflexive instinct to call this yet another form of corporate welfare. But looking at Lahaina, Los Angeles, and fires raging from poor utility upkeep all across America, the amount saved from disaster abatement, the jobs created, and avoiding having to wait another 20 years for a fragmented bureaucracy to upgrade a grid are all obvious benefits. The Department of Energy estimates that power outages and grid fluctuations cost the U.S. economy roughly $80 billion to $100 billion every year in spoiled inventory, halted production lines, and crashed systems. By forcing the grid to modernize through building the substations, burying the lines, and funding the backup generation that the public utilities couldn’t afford, Big Tech is effectively repealing that tax allocation. When a Microsoft data center pays to upgrade a substation in Wisconsin, the local manufacturing plant down the road gets cleaner, more reliable power, too.

Rather than continuing to lose the battle against system age & increasingly severe weather, we are witnessing a massive, unified injection of capital into the energy sector to save it from itself. These data centers demand 24/7 uptime, which forces the surrounding grid to become more robust. They demand clean energy, which forces the rapid deployment of solar and wind at a scale that subsidies alone couldn’t achieve. Oh and also the ‘AI is drinking up all our water’ issue is completely fake.

This is the great irony of 2026. The “Green Jobs” revolution that environmentalists promised, where armies of union workers would retrofit America for a sustainable future, is finally happening. It just isn’t being led by the Sierra Club or the Department of Energy. It’s being led by Amazon, Microsoft, and Google.

The bureaucracy couldn’t deal with the rising challenges of grid reform because it was scattered across thousands of municipalities. Big Tech, driven by the existential race for AI dominance, has the singular will and the blank checkbook to force the issue.

We used to worry that corporations would become too short-sighted to build the future. The irony is that in 2026, they are the only ones planning for it.

Subscribe now

Share

This was the “Ops Room” of Project Cybersyn, an attempt by Salvador Allende’s socialist government to manage a national economy not through rigid central planning, but through a real-time nervous system. They connected factories up and down the unparalleled vertical geography of Chile to this central brain using the only technology they had: Telex machines sending text-only messages.

Project Cybersyn was an attempt to realize a socialist dream while providing a realist response to burgeoning libertarian claims of state inefficiency. Influenced by the stagnation of the Soviet model, Beer and the Allende government sought a third way through a balancing central planning with “real-time” data that represented the day-to-day reality of Chilean citizens.

Beer saw a way out of the inefficiencies of central planning by rewiring it. He didn’t view computers as tools of oppression or surveillance, but as “Liberty Machines.” His logic was that the critics were correct—bureaucracy is oppressive—but it is oppressive precisely because it is slow, stupid, and blind. A citizen dealing with a rigid state is not free; they are trapped in a waiting room, a victim of the state’s inability to process their specific reality. True liberty, Beer argued, is “competent information free to act.”

In a political lesson of its own, however, Cybersyn ended as a failure, a quirky footnote destroyed during the Pinochet coup on September 11, 1973. The military found the Ops Room, didn’t understand it, and smashed it. Go figure.

But structurally, Cybersyn’s programs were failures of resolution. Today we’d see Project Cybersyn as simply a basic messaging service. It was operationally useful in cases where coordinated messaging was important—such as during the 1972 trucker strikes—but daily Telex messages simply did not provide sufficient detail to centrally run an economy.

Would infinite resolution realize Project Cybersyn’s dream? The next few years are going to test that thesis.

I have been a big proponent of the idea that the bureaucratic state should take personalization of service as seriously as any corporate marketer does. During COVID-19, I worked with the California Office of Emergency Services to coordinate data across as many citizen-feedback channels as possible—email, chat, call centers, social media, ads, and chatbots. This helped increase the resolution of the twin logistics and persuasion challenges facing the state. Rather than understanding where vaccines were needed at the county level, we could see needs change at the neighborhood level, including qualitative responses to outbound media and mobile vaccine vans. The distribution improvements alone greatly shrank the disparity between high-income and low-income households receiving the vaccine.

The rise of personal AI agents sprints our technical capability to the finish line of that idea. When Friedrich Hayek wrote The Use of Knowledge in Society, his argument was fundamentally epistemic. He wrote that the “knowledge of the particular circumstances of time and place” is dispersed among millions of individuals and argued that no central planner could ever possess the bandwidth to aggregate that knowledge. When the state tries, as James C. Scott noted in Seeing Like a State, it forces the complex forest into a legible grid. It lowers the resolution of reality until it fits into a file cabinet.

But with the technology of today, are we at risk of saying “this time it’s different” when it….may actually be different? We can, in fact, build a system to answer the particular circumstances of any place or any person, however defined. It is going to be almost embarrassingly easy to do so, even if foundation models get no more capable than they are at this moment. While we currently face constraints on energy and compute capacity, those are solvable engineering problems. Over the next few years, our cloud-based systems will migrate toward becoming AI-managed. Even the most resilient, human-run bureaucratic systems will become AI-led across all sorts of operations.

This creates a new, more dangerous dilemma. In the 1970s, centralization failed because it was impossible. In the 2020s, centrally managed swarms of coordinated agents will become the default setting.

There is a “gravity” to the modern software stack. Data naturally flows upward to the cloud as it is used. Models naturally aggregate insights at the center. If we are not careful, we will accidentally build a perfect AI facsimile of Cybersyn’s Ops Center with the President at the helm.

Beer’s Liberty Machines were smashed by Pinochet’s right-wing revolutionaries. Telex machines and AI Agents are simply tools. We can personalize our bureaucratic interventions to a degree that people have come to expect in their daily lives—or we will build a panopticon of useless data aggregation that is eventually smashed in rebellion against the reality we created.

Subscribe now

Share

The question was what exactly to do, and where. Minnesota started a cross-agency board called the Opioid Epidemic Response Advisory Council (OERAC) which helped steer federal grants. Today, the dashboard tracks 645 grants given to 148 grantee organizations from OERAC, other Federal Block grants, local lawsuit settlement funds, and added State funds.

Our idea was a simple one – analytics. As you can see in the dashboard, the tracking only reached as far as the money going out the door. Minnesota is very adept at telling stakeholders that dollars are set aside for ‘harm reduction’ in the ‘African or African American’ community, or for ‘pain management’ in the ‘individuals with disabilities’ community, etc. What they are less adept at understanding is whether the now-$50M-annual investment is working.

The funding has increased dramatically from 2019 to 2024 (~$3M to ~$43M) largely because early efforts actually did not work at all to curb the crisis. From 2019-2022 Minnesota’s fentanyl overdose death rate tripled and its methamphetamine death rate doubled. Somehow, at the same time they also had the largest increase in cocaine-related deaths of any state in the U.S.

Naloxone came to the rescue. Whatever they were doing wasn’t working, and the state pivoted hard into making Naloxone widely available. Since 2022 Minnesota has seen a positive divergence of overdoses – fatal overdoses dropped slightly (8%), but non-fatal overdoses increased (11%). People were still overdosing, they just were dying less. Progress!

We lost that 2019 RFP. We were told in no uncertain terms that now was not the time for analytics and accountability, it was the time to help people.

But the lack of these analytics systems on top of our gigantic outlays allows for the rampant fraud, waste, and abuse we see over and over again in American life. And it allowed thousands of unnecessary deaths and tens of millions in wasted dollars in Minnesota. Unfortunately, in Minnesota’s defense, there’s nothing particularly unique about how they dealt with (or did not deal with) the opioid epidemic compared to other states.

Over the last few weeks, you have probably heard about the Somali fraud ring raking in over $1B in fraudulent medicaid claims, literally sending US taxpayer money to Al Shabaab terrorists. It was a remarkably simple grift – upwards of ~75 people involved with the non-profit group ‘Feeding Our Future’ submitted fake meal count sheets and invoices showing they had allegedly fed thousands of autistic, poor, and otherwise needy kids. It was a total scam. During the initial trial one member also tried to bribe a juror with $120k.

The unfolding Feeding Our Future scandal is the perfect inverse of the opioid crisis. In the opioid crisis, the data was siloed so we couldn’t see the deaths. Minnesota equated outlays of funds with success. The data gaps between the PDMP (prescription drug monitoring program; every state has one), the 42 CFR Part 2 ‘privacy shield’ law, the medicare and medicaid Systems of Record (SORs), the ICD-10 billing codes in hospitals which didn’t have room for anything other than ‘OxyContin prescriptions,’ the law enforcement seizure data, and more prevented the State from really seeing where the crisis unfolded and who it impacted. The state was blind to the deaths.

With Feeding Our Future, the State was blind to the lives supposedly impacted. Just as 42 CFR Part 2 blinded health officials to addiction data, FERPA (Family Educational Rights and Privacy Act) blinded officials to the truth of who was, or was not, receiving food aid. The fraudsters submitted rosters of thousands of children’s names to the Minnesota Department of Education (MDE). MDE manages the food program, but they also possess the master database of every student enrolled in Minnesota schools (MARSS).

Just like a simple SQL query could have instantly outlined the huge, unnatural spike in prescription rates and associated deaths under the opioid epidemic, so too would a single query in an analytics system across these education SORs easily identify that thousands upon thousands of names were simply made up. As the department costs ballooned from a few million to hundreds of millions in outlays, you’d think someone would check. But there was nothing to check.

Procedure beat out obvious policy issues here. Federal regulations require a specific “Serious Deficiency Process” before payments can be stopped by government agencies, a very high bar. Incredibly, when MDE tried to halt payments based on obvious red flags, Feeding Our Future sued them and the court forced MDE to continue payments until the wider ring was fully exposed.

The sad thing about these twin failures of Minnesota bureaucracy is precisely how unremarkable they are. They are the natural output of a government architecture designed in the 1970s trying to manage the crises of the 2020s. I’ve written before about the developing ‘third wave’ of the American bureaucratic form where agencies like CalOES, stood up for emergencies, are quietly amassing cross-functional power to force agencies to work across data & operational lines.

This is a moral imperative. We have to acknowledge that the era of ‘Agency Sovereignty’ over data is over. We cannot afford to let the Department of Education hoard enrollment data while the Department of Health hoards addiction data, all while the Department of Human Services writes checks to fraudsters who exist in the gap between them.

The solution to this form of fraud, waste, and abuse rife across our bureaucracy is remarkably simple. No one needs to forgo managing data systems or SORs in their agencies. Rather, the analysis of what flows through those pipes must be elevated to a centralized, executive-level agency that has the legal authority to pierce the privacy veils of FERPA or 42 CFR Part 2 when fraud or life-safety is at stake.

Technical modernization offers a rare political cease-fire. For the progressive left, a centralized Data Authority ensures that the $50 million investment in opioid remediation is actually saving lives, rather than vanishing into ineffective non-profits and creating legitimate criticism that leads to programs being entirely shuttered. For the conservative right, removing fraud and waste in the system validates that tax dollars are both minimal and well-spent.

We have the technology to have our administrative law cake and eat it too. We can have privacy and accountability. We can have generous social programs and fraud detection. But we cannot have them if we insist on running 21st-century social outlays based on 20th-century paper logs. The Great Rewiring is coming for us all, and for our public sector, not a moment too soon.

Subscribe now

Then, Microsoft—a major investor in OpenAI—announced ‘Copilot for PowerPoint.’

With a single feature update, Microsoft integrated the exact same functionality as Tome’s entire product directly into the software (that 99% of Tome’s target market already uses). Tome’s core value proposition disappeared overnight. The company is now reportedly seeking a sale at a fraction of its peak valuation.

This is the trap of developing in the world of rapidly-improving Foundation models like ChatGPT, Gemini, Claude, etc.

The entire tech world is in an arms race. Every few weeks, we’re being treated to breathless announcements of new foundational models. Gemini 3.0 is set to release any day now. Each release makes the Foundation models better, faster, and more capable than the last.

Ten years ago, customers would spend weeks or months training key:value pairs for NLP ML-powered chatbots to be used in call centers, or inputting reams of (user-labeled) image and video data to help train proprietary video intelligence models. As of early 2024 I was still getting quotes from vendors of translation software wanting to charge hundreds of thousands of dollars for custom-trained translation models.

If your entire AI company is just a clever wrapper like this on a public API, it’s time to rethink the value proposition. With every new update, Foundation models are taking droves of tech startups out of the ecosystem. The real, lasting differentiator over the next decade will not be the model, it will be access to proprietary, industry-specific, or otherwise hard-to-reach data assets.

We’ve already seen this happen in its most basic form. A thousand “AI-powered PDF summarizer” and “AI-powered copywriter” tools bloomed overnight. They were clever, but they were also thin. Now, you can just drag a PDF into ChatGPT or Gemini and ask it to “summarize this.” I did it for this essay, whereas a couple years ago I’d have to pay a subscription for that capability. The entire product category simply evaporated overnight.

This is the ‘wrapper’ trap. The previous startup model of wrapping more basic out-of-the-box capabilities starts to rapidly disappear in the face of improved Foundation models. So where, then, is the ‘moat’ that can add value to these incredible developments in AI? The answer lies in data access.

Models are only as valuable as their inputs. The incredible advances in the LLMs, Video intelligence, coding assistants, and the like all rely on the inputs from the user, as text or .mp4 files or log history. The AI brings the analysis engine, but it must be pointed at the right data to be immediately useful.

Data curation is an art as much as it is a science. Collecting information proximate to the issue at hand can be a mind-numbing process of scaled data architectures. Twenty years ago, you could buy a LexisNexis ticker to see the news data for the day. Today, we have not just the news story itself, but metadata on publication time, user reach, subscription revenue, etc gained from the publication. We can see how each story performs on Facebook, Instagram, LinkedIn, Twitter, or even Pinterest. We can identify bot networks resharing articles, or journalists editing the story in real time. We can see how stories move across these mediums, and whether the narrative embedded is gaining or losing traction. There’s just so much more to measure.

As a way to think about the future, there are a few examples of ‘data moats’ that won’t disappear into GPT 6 or Gemini 4 anytime soon.

1. Sports - I’ve written about this before: the AI revolution in sports isn’t new. Teams like Manchester City and Arsenal are dominating because they spent the last decade building the backend data operations to capture mind-bogglingly granular data. They are tracking player movements, ball physics, and biometric outputs. The new generative AI models are simply a powerful frontend to analyze a deep, proprietary data asset that no one else has.

2. Connected TV - The ad-tech world is desperately reacting to AI and the winners will similarly be data brokers. Privacy regulations and the death of the third-party cookie have made generic data useless. Who is winning? The companies with first-party data. The power is in the direct partnerships, like the Disney+Google or Kroger+Hulu deals. They are linking their exclusive, proprietary customer datasets in data clean rooms. No ‘AI wrapper’ can reasonably compete with the deterministic data Kroger has on what you actually buy. The public models work off balance sheets while the private models work off of real-time ad+purchase data.

3. Finance - Who is best positioned to build the dominant AI for financial analysis? It’s not a hot startup with a finance wrapper on GPT-4. It’s Bloomberg. Why? Because Bloomberg can train a model on decades of proprietary, structured market data, analyst reports, and terminal queries that no one else on earth has access to. A generic model can tell you what a P/E ratio is; a Bloomberg-trained model can tell you how the market has historically reacted to a specific phrase in an FOMC statement.

—---

One of my earliest posts on Social Dawn was about watching the FDA completely miss the youth vaping epidemic, and not being able to do anything about it. In my time at Google working on their “Real Cost” anti-smoking campaign, the ads data coming back to me was a veritable goldmine of real-time search, video, and survey data. This data was showing, clear as day, the explosive rise of a new product category: ‘Juul’ and ‘vaping’ and ‘e-cig.’

But the government’s model was built on slow, multi-year longitudinal studies. They had the data asset to complement the dizzying volume of digital data we were receiving, but their bureaucracy and procurement processes meant they couldn’t use it. They saw their digital tools as a one-way street for delivering ads, not as the critical, real-time feedback mechanism they are. By the time their official studies confirmed the trend, the youth vaping epidemic was already raging.

The next few years will be a true reckoning. The companies that survive and thrive will be the ones that understood, from day one, that the AI is just the engine. The data is the fuel.

The rest are just features, waiting to be eaten.

Subscribe now

The explosion of the Mag7 has led many talking heads to declare the current moment the “biggest bubble yet,” or no actually there are three concurrent AI bubbles, and the whole industry is overdue for a dotcom-era burst. Most of these arguments rely on the huge amount of public & private investment flowing into these companies, questioning the ability to see returns.

This may be a good venture capital case against OpenAI or Anthropic specifically, but when it comes to the cloud providers and chip makers that deliver these capabilities to the enterprise we are still in early innings.

Building AI applications means building entire dedicated data, UX/UI, and now agentic systems with extremely specific needs on how these parts interact. Ten years ago, UI/UX teams could operate with relative separation from data teams, as all the UI/UX folks needed was data to simply be available at the moment of need in a retrievable format. Where and how it was hosted made no difference.

Today, for enterprises to make use of AI means building knowledge bases of both data & internal tooling for Agents to use and understand. On the flip side, this is also true about the UI, where agent decisions and unique outputs throughout the new UX experience implicate a far more flexible and personalized UI.

Most enterprises that have fed trillions into Azure, AWS, GCP (et al) have done a lot of work in lift & shifting much of their heaviest on-prem workloads into the cloud. This gives teams flexibility to use data assets more readily across the enterprise, but to what ends?

Big companies and government agencies alike naturally created their back end systems to help their unique operations. As organizations grew, this process led to a similarly natural siloing of back end capabilities along mission lines. When I worked in public health, for example, in Utah there were separate entire systems for data related to health insurance, mental health services, financial assistance, child protection, career counseling, etc. But this means a lot of individual functional redundancy. When I worked with the Air Force recruiting command, for example, I learned that across the Air Force they had almost 100+ separate HR systems for separate purposes & verticals.

From the outside, this is seen as colossal waste. But creating a coherent view across these siloed systems is both (a) the same thing as creating yet another system, taking the same amount of organizational time and effort, and (b) begs the question of what views to build and why. Who cares if there’s 100+ HR systems if there’s 100+ reasons to have one? We heard this quite a lot in the early days.

As the benefits of “don’t spend 10x+ on redundant stuff” became plainly obvious to IT directors and mission owners alike, the last 5-10 years have been an exercise in getting these organizations familiar with the language and tooling of cloud systems, from chips & GPUs to modern data warehousing & AI systems to networking & identity management to zero-trust security & modern privacy standards.

I like to say that AI doesn’t tell you the answer, but it removes 80%+ of your possible choices to the limit you to the ones that matter. When we worked in Utah in the social services example above, the combining of these public health systems allowed for easier case management across programs. This helped analysts see statewide problems more clearly, refocusing resources. But functionally that meant cutting a lot of redundant functions. Less overlapped outreach, less time spent on paperwork, less time spent scheduling between systems, less time spent trying to get data from different teams, less time needed in grant writing and funding requests, and so on.

The unification of data across Utah’s social systems allowed the state to focus. We see this (rightly) as a better value delivery, but tactically inside organizations it feels almost like a relief in having to stop doing tons of unnecessary work than a revolution in doing entirely new work. This is the simple reason I think the AI revolution is just getting started. While there have been some success stories like Utah, far more organizations have still been thinking about their back end operations in the old, verticalized model.

The LLM breakthrough requiring agentic systems forces these same organizations to rethink their whole infrastructure setup. In order to utilize obviously-valuable AI agents either inside their own companies or embedded into their offerings, there will have to be a reckoning that flips the traditional model on its head. In the social services example, case managers log into the financial assistance system, manually inputting the family’s income, address, and personal details to determine eligibility for food or housing aid. Next, they’d log out and open an entirely separate health insurance portal, re-entering much of the same information into a different user interface with slightly different requirements. Finally, for mental health services, they have to navigate a third more outdated healthcare paper-based system, in order to find an approved provider, check their availability, and submit a referral which again means copying the same core family details. Each step is a discrete, redundant task, and the case manager’s primary job is navigating bureaucracy rather than delivering care.

Now, imagine an agentic system built horizontally across those functions. All a case manager should need to do is input a family’s core information just once, as well as indicate the systems they need assistance with. Agents should be able to process the requirements across all programs, determine eligibility, process required information across databases, and return that information along with scheduling availability for local providers.

The ability for agent systems to reach across functions requires years of dedicated work from data teams, internal tooling systems, UX/UI teams, and business operators. This is not going to be a fast process for big organizations. The hard, unglamorous work of re-architecting data, building internal tools, and redesigning workflows is where the real value will be unlocked. While pundits debate the valuation of a few model providers, the trillions being spent on the underlying infrastructure to support actually-useful AI agents aren’t fueling a bubble. They’re laying the foundation for what comes next.

Subscribe now

The DOJ wanted to pursue the case due to the anti-competitive behavior they saw from Google, most prominently its whopping $26B annual payment to Apple for default status which extends to today in GenAI deals like with Samsung for Gemini pre-installation.

The court agreed, saying that despite vertical search competition was fierce in some domains (50-60% of shopping queries start on Amazon, for example) there was a market defined as ‘General Search Engines’ in which (per Judge Mehta’s opinion at the time) served as a distinct product category. He defined a ‘GSE’ as being:

(a) broad in scope of information they cover,

(b) act as a gateway to other content,

(c) handle a mix of commercial and non-commercial queries (e.g. queries that are monetizable with Ads and queries that aren’t), and

(d) Distinct production, meaning that they are worked on essentially as a product in & of themselves rather than as a feature on another platform

In 2024, Judge Mehta saw a ‘GSE’ like Google or Bing as distinct from “walled gardens” like Amazon or Expedia. The theory of the District Court was that with Google’s dominance in the “GSE market,” no one could compete due to their vast amounts of user data, close business relationships, and high barrier to entry in building a new GSE.

Fast forward to 2025, and the product category definitions that make up a GSE seem to be features within modern LLMs. ChatGPT, Perplexity, Gemini etc can all reach a much wider scope of content than GSEs, link to source materials, cover all kinds of queries, and definitely have far more robust & dedicated production teams than GSEs.

The time between the initial ruling and the ‘sentencing’ hearing where ramifications for the ‘GSE monopolist’ issue was only a year, yet the entire underlying industry had shifted to the point where the usual ideas fell flat. In his follow up ruling this week Judge Mehta kowtowed to the reality of AIs disruption, essentially ruling against Google only in areas also relevant to Generative AI. He rejected more GSE-specific remedies like divesting Android or Chrome, rejected blocking integration across Google products, rejected mandatory choice screens when using Google products (like they have in Europe), rejected forced changes to Google’s publisher policies, and he allowed vendor/partner payments to continue as long as there was not exclusive or preferred status.

If the story were still about the value of GSEs to the global digital market, half of those remedies likely would have been taken up. Instead, what he focused on were remedies essentially tailored to help competition thrive in the Generative AI space. He instituted a ban on the exclusive distribution of their AI services in partner products, including blocking mandatory licensure of Google products for business customers like the Play Store app developers. He expanded the set of competitors from GSE-creators to “Qualified Competitors” which folds in the OpenAI and Anthropic’s of the world. These Qualified Competitors are also entitled to some Google Search Index & user interaction data, and they will have new access to license syndicated search ads services on Google properties.

While these remedies make some sense in that they help open the door for competition across Google’s core services, it does seem strange that the monopolistic behavior they were found guilty of is only tangentially related to the remedy of making Google less competitive versus peer rivals in the area of Generative AI. The goal of then trying to right-size remedies for a new market area could make sense if Google were similarly the market leader. But they are not.

The issue is baked into anti-trust law itself. Judge Mehta is forced to admit that “the emergence of GenAI changed the course of this case” as the “GenAI chatbots perform an information-retrieval function similar to GSEs.” Anti-trust law deals with this via the concept of “reasonable interchangeability of use” where the edges of market categories are defined by their “cross-elasticity of demand between the product itself and substitutes for it.”

The seminal case where this definition comes from helps make the precedent clear. In 1956 the US Government sued the DuPont Cellophane Company which owned 75% of the cellophane sold in the US. DuPont countered that cellophane was an innovation in the wider flexible packaging materials space, against products like wax paper or tin foil. When compared to the packaging materials space in general, it only had a 20% market share. The court agreed, saying there was a ‘reasonable interchangeability of use’ when consumers went to buy a product to wrap their sandwiches in, so the wider market definition & DuPont Cellophane prevailed.

Back to today, there are certainly some overlapping functions of ChatGPT and Google Search, but they serve completely different functions. The legal analogy to DuPont's cellophane, therefore, starts to break down under the weight of modern technological velocity. The market for flexible packaging in the 1950s was relatively static, and today’s technology industry is….not. The legal system's reliance on market definitions that take years to litigate is fundamentally at-odds with an industry that reinvents itself in a matter of months. The very ‘GSE market’ that Judge Mehta so carefully defined in 2024 was, by 2025, already being subsumed, reduced to a mere feature within a far more capable technological paradigm.

The court’s attempt to pivot the remedy from the original GSE market to the new GenAI landscape is therefore a turn towards judicial overreach. By finding Google guilty of monopolizing GSEs but then crafting remedies primarily to shape competition in a nascent GenAI market where Google is a competitor, not a monopolist, the court is legislating from the bench. If the goal is to preemptively regulate the burgeoning field of generative AI, that is a policy question reserved for Congress. Antitrust law is meant to be a scalpel to restore competition where a monopoly was proven, and has no place in industries outside the edges of their market findings.

Google was very lucky that the ruling came down when it did, if GenAI took another 18 months to develop Google would very likely have been forced to divest some of the most valuable parts of its business the instant before that same core business would be out-innovated by a couple San Francisco startups.

This lucky timing shouldn’t have had to occur. A simple year-long gap between the liability and remedies of their GSE monopoly rulings revealed a fundamental flaw in the analysis. Instead of contorting a ruling about an increasingly obsolete market to fit a new one, the more intellectually honest conclusion would have been to acknowledge that the ground had shifted too completely, rendering the initial finding of durable monopoly power moot. Judge Mehta says that the two industries are “similar in function” and share “tens of millions of overlapping customers” and GenAI is a “nascent competitive threat” (to Google Search). But that’s a description of disruption rather than a description of “similar function” extension that the District Court framed. The horse & buggy shared a lot of similar riders and functions as early automobiles. But I doubt we’d sit here today and argue they share a common industry.

Without a more direct way to deal with the shifting market dynamics of the modern technology industry, anti-trust is going to continue to be left toothless and grasping at straws. Whether by a new legal precedent or by an act of Congress, something soon is going to have to lead our modern anti-trust movement out of the woods. They’re completely lost.

Subscribe now

Pieces about China’s complete outstripping of American naval production, air production, hi-tech weapons manufacturing, and so on assume that Taiwan will be the first conflict area. Both the US and China have separately been conducting war games around the island with increasing intensity, stoking fears further.

If it comes to a clash in Taiwan, however, recent conflicts have taught us a different lesson than the might of force alone being the deciding factor. Despite a pre-war ~10x investment in military capabilities in Russia than in Ukraine, for example, Russia has not been able to steamroll their neighbor as they had hoped. It has been Ukraine’s ingenuity rather than force, even with American arms, that has allowed it to fight Russia to a stalemate in eastern Ukraine as of this writing.

The change in how the modern conflict operates is subtle but significant. On its face, it’s similar in tactics – attack the enemies critical infrastructure both physically and digitally, spread your propaganda, and scale good ideas fast. But the dramatic improvement of these basic capabilities create a military margin for skill. The modern conflict calls for quality over quantity.

WWII's model of industrial warfare essentially held true all the way through the Iraq & Afghanistan wars’ "shock and awe" campaigns, using overwhelming force to rapidly dismantle the governing infrastructure of those nations. However, that quick battlefield victory was followed by a two-decades-long failure in fundamentally changing Iraq and Afghanistan. The old model was built to win battles, but not to win the peace.

In contrast, today's conflicts demonstrate a new reality. The old idea of massive-scale direct clashes has been replaced by massive-scale system penetration. This strategy operates on the principle that the war is no longer an “invasion” of force but an infiltration. What makes Ukraine successful isn't its ability to out-gun Russia, but its ingenuity in subverting Russia’s ability to wage war on various fronts, often by targeting the very systems Russia relies on. This turns a "contest of force" into a "contest of faith"—undermining the enemy's confidence in their own capabilities and institutions.

War is still an industry requiring a ton of manpower. Even the most sophisticated and efficient military organizations are quite large, which is why there’s different terms for unit sizes (squad, platoon, company, battalion, etc) up to the hundreds of thousands. Effective governments not only can enlist the talent necessary in those large quantities, but do so with the population’s willing support. An unwilling, draft-heavy military typically has far more issues with role performance, efficiency, and importantly motivation & morale than a volunteer force. And the modern ability to hyper-target military operations is also hyper-targeted to negatively impact motivation & morale.

The naval conflict inside the Ukraine/Russia war serves as a useful case study. As the war in Ukraine started, a dedicated team in the Ukraine military was using open-source data, satellite imagery, and on-the-ground human intelligence to track the movements and vulnerabilities of Russia's Black Sea Fleet, looking for weaknesses. They realized that Russia’s much larger warships had a blindspot to the detection of extremely small craft, being built for direct conflict with other warships. Turning to the logistics side of the organization, Ukraine began to produce kamikaze ‘Sea Baby’ unmanned surface vehicles (USVs, essentially autopilot little boats with bombs in them). They preprogrammed a route for each Sea Baby to take, and let them loose. The results were also then captured on film from an on-board camera and advertised widely in Russia, with some estimates of upwards of 40% of Russia’s Black Sea Fleet being destroyed or disabled.

With story after story like this spreading throughout Russia, it’s hard for even the most despotic regime to keep up. Military volunteerism is down 80% in Russia, where the five-fold drop has the regime turning to criminal conscripts and Chinese, Ghanaian, South African, and North Korean paid recruits. Ukraine’s success here in dampening enthusiasm for the Russian military project is as much of a victory as any battlefield success, as they too struggle to keep up with the sheer size and scale of their opponent.

I don’t like the idea of a world where military operations become more insidious, giving rise to conspiracy theories of all kinds, often with more legitimacy now than ever. Russia is paying influencers to sow discord! Well, yeah. China is hacking the phone lines to steal your data! Yeah, kind of? Iran is planning anti-semitic attacks on Jews around the world! Well…

Rather than one-off events, in the broader context of the military goal each event looks like the puzzle piece that it is. I doubt Russia is paying influencers or China stealing phone data outside a larger military goal, but from our vantage point we will never know what that goal actually is, and so instead just see the insidious edges of campaigns without that critical context. This is ripe ground for conspiratorial thinking. It’s just a lot harder to hide what happens to a couple hundred thousand troops than what happens inside a centrifuge no one is supposed to know exists anyway. This new reality makes the ‘contest of faith’ a double-edged sword. When the war is fought in the shadows & across disciplines, it becomes easy for the average person to see a hidden hand in every failure.

This dynamic, however, also presents a unique advantage for democracies. Authoritarian states must control the narrative to maintain a facade of strength, and so would never admit mistakes or talk deeply about successes that deviate from the administrative message. Democracies should face conspiratorial thinking head-on, openly sharing their post-mortems and full picture of operations after they are complete. Each operation is now so unique, that it’s likely most won’t be conducted in quite the same way ever again. The Israelis debrief of their pager operation against Hezbollah, for example, is a case study in both modern tactical design but also using the story itself as a military tool, increasing domestic morale while devastating Hezbollah’s recruitment.

Openly talking through operations like this publicly shows domestic audiences of true military capability, while being honest about how & where that capability is being used. It also likely dampens the enthusiasm of the opponent’s population to take up arms, which is the ultimate victory in a world where conflict is no longer about invasion, but infiltration.

With the focus on public sector fraud, the limelight has shifted off of the private sector. This is unfortunate timing, as investment scams from legitimate-seeming market investments to newly-minted crypto coins produced a record $12.5B in lost fraud dollars in 2024 according to the FTC. $5.7B of that was individual consumers investing their own income, while institutional fraud scaled the rest of the dozen and a half billion lost. Those numbers are only growing, as “in 2023 27% of people who reported a fraud said they lost money, while in 2024 that figure jumped to 38%.”

The fastest rising category of fraud are ‘imposter scams,’ where someone contacts you pretending to be your boss or a government representative. These scams rely on either you, personally, executing gift card sales or crypto transactions, or more often simply opening up your email and digital wallets to the imposter actor after falling for a ruse. The scammers have gotten more sophisticated, posing as recruiters to job applicants, or selling fake goods off the backs of fake reviews, or simply convincing you to invest in almost any available crypto coin.

And while these fraudulent activities are regulated on a case-by-case basis from the FTC, USPS, DOJ, and other regulatory bodies, these bureaucracies are inherently reactive. There are policy levers that Congress must pull to be able to combat the rampant rise in nationwide fraud, but often their ideas simply introduce more friction points and so consumers complain and nothing happens.

Due to the decades-long Congressional inaction on digital policy, like so many other places we have ended up at a point where the strongest fraud-fighting body in the nation is not DOGE or the House Appropriations Committee, but large companies. Google, Amazon, CloudFlare, and other big tech companies remove more fraud, waste, and abuse from national GDP everyday than any bureaucracy on earth, much less our own.

For those of us who remember the pre-mobile, pre-social media ‘web1.0’ era, it’s undeniable how instrumental Google Chrome’s ‘Safe Browsing’ was in turning the web from an anarchic playground of the early 2000’s (where simply the idea of commercial transactions were the object of South Park ridicule) to what it is today. Google’s follow-on focus on 2FA in the 2010s, to the announcement last week of mandatory 2FA across both B2C and B2B enterprise cloud customers, comes on the heels of $16.6B lost to online scams of all kinds last year alone.

Despite this 33% rise in online scams, it is not Congress or the bureaucracy stepping in to try and mitigate the damage. While the entire US Federal Government warned the public about 39 zero-day exploits in 2023, a record, Google found & warned the public about 1800 zero-days from 2014-2022. And it’s not like the CIA or NSA are secretly stashing thousands of zero-days either, as 90% of all zero-days found by our intelligence or other agencies are released publicly.

Initiatives to counter fraud, waste, and abuse worth tens of billions of dollars in saved GDP annually abound across big tech. Apple has removed “$9B in fraudulent transactions over five years, including $2B in 2024 alone” by simply improving App Store-wide monitoring and alerting for deceptive apps or fraudulent payment schemes. Their regular bounty programs to hack their new technology produce some of the most valuable hardware-related security postures in the industry.

Bug bounties were popularized by Facebook/Meta, who paid out $2.3M+ last year in their program. Account removal for fraud is so common it's a widely tracked metric often exceeding over a billion fake facebook accounts per quarter.

I could continue with examples for a dozen pages, from Amazon’s innovations in online shopping fraud tooling to Microsoft’s entire Digital Crimes Unit, big tech has had to take on both the creation and enforcement side of the digital economy. The combined tooling, direct investigation work, and proactive cybersecurity alerting that big tech offers to the world dwarfs our public sector’s attempt to truly combat fraud.

Yet the thing about fraud today is that the same technologies that empower scaled fraud also empower nearly immediate rectification. It simply requires digital identity standards that are enforced across the online ecosystem.

The imposter scam works not just because a scammer is clever, but because they exploit a system with no underlying standards for identity or verification. A common pattern is a fraudster using Facebook or Instagram to find a victim, a disposable number like Google Voice or Telegram to contact them, a payment app like Venmo or Cash App to receive initial funds, and a decentralized crypto exchange to make the money disappear forever. Each of these companies has built a formidable, but separate, identity castle. But as we have risen the walls of digital identity around each provider with privacy laws like GDPR and CCPA, we have also given fraudsters a neat playbook. And when a victim tries to reassert their identity to their bank or Apple Pay, simply being locked out of one critical system by a scammer could quickly upend their life for years on end.

While nations like China with all-encompassing apps like WeChat simply don’t have the same sorts of fraud issues that our fragmented system entails, it doesn’t mean we have to resign ourselves to a future where fraud continues to escalate.

Congress should pass legislation to expand & formalize the Login.gov secure sign on to a larger digital identity platform that can be leveraged by private industry. A secure protocol that opens only to regulated or highly scaled digital platforms could help people & companies to quickly ‘vouch’ for who they say they are, without having to rely on trust or phone references at all. If that is paired with standards for transaction ledgers to help maintain direct transaction history paired with verified individuals, then even in cases of fraud the clawback would be far more frictionless.

Rather than regulating or maintaining a particular technology, Congress should simply create a new ‘Office of Digital Trust,’ empowered to work with both consumers and companies to both enforce the standards above and help consumers claw back funds.

Today, a victim of fraud has to do something like file an FTC complaint to register the issue, contact their local police or somehow make it to talk to someone in an FBI department, work with their bank, the social media app, the payment app, and any other third party involved in the transaction.

Under a unified identity & payments standard, the victim would similarly report a fraudulent complaint. The complaint would carry the identity key of the victim, which would trigger the regulated companies ledgers to disseminate the need for transaction data related to the victim. Under this new standard, a verified fraud report filed with the Office of Digital Trust would empower the office to issue a standardized, legally-binding 'Preservation Order' to all connected platforms, requiring them to report relevant transaction data within 72 hours and temporarily freeze associated funds.

Life-altering processes that take years to slog through courts, help center chatbots, and bored call center workers could turn into days-long annoyances with a few strokes of Congress’ pen. The Office of Digital Trust could be lean and effective because it would simply enforce standards, not create them. As a model, look to the National Institute of Standards and Technology (NIST), the agency that develops the core technical frameworks our digital world relies on. NIST provides the nation’s cybersecurity and encryption backbone with a budget of only $1.6 billion and a staff of about 3,400.

Establishing the Office of Digital Trust would help our bureaucracy take both the first step towards a modern solution to protect our citizens from the rampant fraud of our digital age, but in doing so teach itself how to constructively engage with the online marketplace as a regulatory body.

So much public policy analysis is downstream of the revenues taxation can generate, with whole branches of agencies like the Congressional Budget Office or Treasury’s Office of Tax Analysis dedicated to studying precisely-optimal tax regimes. With over 95% of the Federal budget coming from taxation of various kinds, it makes sense to work hard to optimize the nearly five trillion dollars of revenue coming in each year.

The remaining 5% comes in from earnings on Federal Reserve investment schemes, licensing or general fees to things like US Parks systems or usage rights to federal land, building tenancy leases, and general land or other property sales that occur.

The notable missing revenue source in this ecosystem is anything informational. At its core public agencies are information management machines, and as this substack has argued for a more functionally-oriented, cross-agency focus of that information megalith, but that core function is seen as a purely inwardly-useful enterprise. That should change.

In the information economy today, cloud providers host data exchange processes that allow (for example) users in Google Cloud to publish or subscribe to datasets down to column & row access controls. Similar concepts exist in AWS ‘data exchange’ or Azure’s ‘Enterprise Data Marketplace’. Some of the most valuable free datasets in these systems that millions of users rely on are government datasets around weather patterns, financial flows, and more. The market for creating data marketplace platforms is already a $1.5B industry, supporting data acquisition systems which are a $3.7B+ industry and growing.

Subscribe now

It’s great that public datasets are being released for free, and at the scale of small projects or individual users they should remain free. But at the scale of industry there are datasets that would take significant investment for the public sector to create & maintain, but could offer to larger enterprises for a subscription fee that could make a significant dent in the Federal budget.

Consider Geospatial data. The NGA is an intelligence agency and is reluctant to share data with anyone. But take an unclassified version of their imagery, focused domestically, and make that feed licensable. Construction crews would find it extremely valuable in preparation for the terrain, elevation, property boundaries, etc. Real Estate would find it valuable for market trend analysis and development planning. Farmers & agriculture would find it valuable for everything from crop health analysis to precision farming. Logistics for route optimization and fleet management. Utilities for network planning. The rise of Autonomous Vehicles could make great use of it for updating their models and improving route analysis. And on & on.

The geospatial industry is growing rapidly and will reach an estimated $1.4 trillion dollars by 2030. The organic growth of the analytics industry surrounding geospatial data is indicative of where our public systems can be useful, and it’s just one example.

Along similar lines, the F.A.A. has essentially a monopoly on air traffic data systems, including cargo logs, investigations & compliance, and other valuable economic data tied up deep in F.A.A. data centers. The geospatial industry might be reaching a trillion dollars soon, but the supply chain & logistics industry that could make use of this sort of data is already over $10.7 Trillion. Considering the industries impacted comes up with a near endless list – logistics & shipping, retail, manufacturing, e-commerce, aerospace, defense, insurance, and so on.

There’s a lot of this sort of data sitting in disconnected systems for use in fairly small & targeted ways that could provide potentially trillions of dollars in value without having to touch a single system with citizen data. But those systems would be extremely valuable as well, and are already common in industry today. We call them advertising engines.

We’re familiar with how Google and Facebook create ad targeting lists like ‘Back to School Shoppers’ or ‘Jets Fans’ that you can advertise to. But more recently, Retail Media Networks (RMNs) like Target’s Roundel system or WalMart Connect combines digital user data with in-store shopping data to create niche audiences like ‘Buys Sephora at Target’ or ‘Buys Bananas at WalMart’ without exposing the PII of those users.

Public agencies should make use of this same underlying technology. Accurate population and demographic data is today maintained by (mainly) Census surveys. This is a century-old solution to a modern problem where digital footprints exist as well. The use of a Federal CDP as this substack has argued for could help produce licensable datasets around internal migration patterns, general health statistics and trends, shifting socioeconomic indicators, and simply far more useful census data. There are already billions spend incorporating data from the current Census surveys across marketing & advertising, retail, real estate, financial services, healthcare, urban planning, education, and other social services. The use of analytics in marketing & advertising alone is already an industry worth over $1 Trillion.

Environmental & Climate Data, more focused public health data, more useful employment & employer data systems, and more are areas ripe for innovation in public data exchange. The 2018 Open Government Data Act would seem to preclude a lot of the above occurring, as it mandates open licensing and dedicates all data to the public domain. And again, at small scales this is a perfectly fine solution. But when dealing with large enterprises who increasingly desire petabyte scale data systems to power their increasingly complex products & services, those open standards would impose a significant cost burden to maintain.

The prevailing view of managing large data systems as a costly and often thankless obligation, rather than a viable path to revenue generation, largely prevents agencies from recognizing their data's inherent value as an asset – a mindset shift already embraced by the private industry today. We should change that, and face the obvious privacy & security questions head-on. There are ways to accomplish this while preserving and protecting sensitive information. We should lean into the information age with a clear goal of universalizing and ensuring high quality, accurate data systems that help coordinate industry actors around trusted sources they can rely on.

Federal income taxes collected about $2.2 Trillion in 2023. The top 10% of earners paid 72% over the total collected tax. It is not far-fetched to think that a well-run Federal data exchange system could collect ~$615B and essentially alleviate Federal taxation on 90% of American earners. Or take a percentage of that and reduce the tax rate across the board for every American by 5%. The impact on the economy of freeing up even that much income in every wage earner's hands while not having to sacrifice our federal coffers would be a massive boon for the economy.

With the rise of AI & agentic layers that rely on even more cross-industry data systems, if we start today we could eliminate most income taxes within our lifetime. Why not try?

Subscribe now

Who are you to your government? The answer is that there is not one ‘you.’ You are a marriage record, business license, and property tax record to your city. You are a birth certificate, drivers license, and court case to your State, and you are a social security number, tax record, and passport ID to the United States Federal Government. If you are lucky, you are also represented by a couple Senators and a House member, as well as a number making up the membership rolls at AARP, ASPCA, Avocados from Mexico, or some other advocacy organization. Otherwise you as an individual aren’t really “seen” by any bureaucratic body in existence. While constituents certainly play a part in bureaucratic organizations, that part is so well-defined and fragmented that the individual 'you' rarely truly registers or influences the bureaucracy's internal trajectory in any meaningful way.

In the last decade or so the Federal agencies have been pushed more & more to use Login.gov, which has a single-sign-on (SSO) capability that all agencies can then use for their own validation. This also can extend further from that central point, as Indiana's ‘Access Indiana’, which combines SSO capability with simple extensions to submit paperwork like applications for social services programs. In turn it allows for easy integration into feedback systems from agencies themselves without overlapping issues. When I worked in a similar process in a midwest state at my time at Google Cloud, a common issue was a constituent working across two lines of social service agencies (e.g. unemployment and food assistance programs) where different scheduling systems and contact points created demands for people to literally be in two places at once.

These conflicts could easily be resolved by taking a constituent first view of the issue, which these centralized identity systems can solve. Building robust profiles of individuals, with records of touchpoints all across public facing systems, can be managed centrally but used across agencies of all kinds. It would be relatively easy to manage and maintain, as this is well trodden industry ground. It would also improve cybersecurity efforts which have repeatedly leaked when left to frontline agencies to manage on their own.

You can also federate that login with other identity providers (IdP), like how you use Gmail or Outlook for logins to news websites. The news site never sees your credentials directly, they just are matching IDs and credential keys to Google or Microsoft, who themselves don’t see your activity on NYTimes.com. This two-step process between the SSO and the IdP helps preserve what each wants – data for the IdP to link to your account and access it at all by way of your email service’s SSO.

The downside to central points of robust focus, especially if you do not have a federated model backing it up, is that they are also central points of failure. The risk of building high walls around your important data is that if someone still finds a way through, the treasure is all just sitting there nicely in one place. And any time the system went down it would quite literally stop operations at potentially thousands of downstream agency users.

Working with government agencies on constituent-facing programs of all kinds, typically each agency is working off of a strict set of requirements for participation. You need to submit income data to qualify for medicaid or food assistance, but not for unemployment assistance. You need to submit proof of termination for that, but you wouldn’t need to bring that to a State Work Agency to apply for a new job. You may also need to prove you are seeking that new job in order to qualify for unemployment assistance, but you wouldn’t need that to qualify for a low income transit assistance program. And so on.

By coordinating data centrally, even if maximally federated to increase security, programs can come in to access just the requested data for each program without having to overburden systems and approval processes to get more information about the constituent than is expressly required. Research at Mississippi State University points to how this decentralized model could work, where the increased use of digital driver's licenses serve as the basis for digital identity which services like getting a fishing license could use to verify what they need without having to recreate digital identities of their own to manage.

Today, cloud systems can create ‘publish<>subscribe’ datasets that control sharing and use across data partners down to the column and row level, which would power systems like these. I’ve used Google Cloud’s Analytics Hub within their BigQuery data warehouse for this process in government processes before, where one dataset was shared across half a dozen stakeholders with their own individual views of the coordinated dataset based on the intersection of seniority and sensitivity of the data. It’s easier than ever to conduct high-compute operations like federating dataset use across thousands of subscribers, while leaving the data in-place. In the pre-cloud era this process was essentially impossible without dedicated hardware and software, so it simply didn’t happen.

We shouldn’t take the status quo of each agency reinventing the wheel of constituent identity as gospel, it was merely the right thing to do with the pencil & paper technology available at the time. In the new era, these simple operational changes for the sake of efficiency can radically reshape how your bureaucratic bodies see you as a person, rather than a number as part of a group. As we reshape the way we are seen by our public programs, it would be beneficial for us all to not make the same mistake again, and take real steps to understand the practical tradeoffs in building either highly centralized or highly distributed digital identity systems. Whatever we decide will likely end up being the basis for public program implementation for the next century.