In early September 2026, the Houthis took the Yemeni Red Sea port of Mocha. That put their military presence about 80 km from the Bab el-Mandeb Strait. On September 10, shipping counts dropped at points to 6 vessels. The change is that the Houthis now hold a more direct position from which to pressure shipping. Red Sea security costs remain high.

Political pressure is also building inside Europe. French student protests lasted until November 2. After classes resumed, 41 high schools were still affected by blockades. Spain’s housing decree was defeated in parliament. On October 5, the government announced an early general election for November 29. A Forsa poll on November 3 put the AfD in the lead at 26%, with the CDU/CSU down to 20%. Dissatisfaction with Merz reached 87%. The yen is trading between 157 and 158. Finance Minister Satsuki Katayama keeps warning that Japan will act against excessive or disorderly currency moves.

Subscribe now

Read more

The second round of Brazil’s presidential election ended on October 25, 2026. The final results from the Superior Electoral Court gave Lula of the Workers’ Party 51.79% of valid votes. Flávio Bolsonaro of the Liberal Party got 48.21%. Lula wins a fourth term. In the first round, Flávio led with 47.03% to Lula’s 45.16%. Lula is the first candidate since direct elections resumed to trail in round one and win in round two. He takes office on January 5, 2027. Flávio refused to concede. His party’s chairman acknowledged Lula’s win on October 26.

Subscribe now

Read more

• Equities re-valued the return on AI capital.

• Treasuries traded safe-haven demand and positioning.

• Crude raised the risk premium on geopolitical supply disruption.

• French bonds repriced fiscal credit.

Divergence Is the Evidence

Equities

The selling was a concentrated repricing of the AI supply chain, not an undifferentiated exit from the market. The Nasdaq Composite fell 694.00 points, or 2.60%, to 26,008.86. That was its biggest one-day drop since June. The Philadelphia Semiconductor Index fell 773.58 points, or 6.72%, to 10,731.98.

• Lam Research fell more than 11%.

• KLA fell more than 9%.

• SanDisk, SK Hynix, Intel and AMD fell more than 8%.

• Nvidia fell 4.40%.

• Meta fell 6.98%.

The Dow rose 0.02%. The S&P 500 fell 1.21%. The semiconductor index is down more than 26% from its June high, which is technical bear-market territory. It is still up about 51.5% for the year, against about 11.9% for the S&P 500.

The selling hit AI capex, chip equipment, and richly valued assets. The market is testing earlier growth expectations, not leaving US equities wholesale. A company does not have to wait for profits to fall before its share price adjusts. If expected future cash flows are cut, or the cost of capital needed to earn them rises, the valuation has to be recalculated in advance.

Treasuries

This was one of the largest one-day declines in yields since the regional bank crisis of March 2023. The 10-year yield fell 19.3 basis points to 4.952% and is back below 5%. The 2-year fell 18.8 basis points to 4.635%.

Guosen Securities’ multi-asset weekly of October 25 estimates that about 15.5 basis points of the 10-year decline came from term premium compression. The change in risk-neutral rates was relatively small. That is a model decomposition.

A Treasury rally can reflect safe-haven buying, a change in rate expectations and short covering all at once. Treasuries combine low credit risk, deep trading liquidity and use as collateral in secured funding. The rally reflects haven demand, rate expectations and position adjustment together. It cannot by itself show that recession has become the base case.

Currency

The yen strengthened, but the exchange rate alone cannot separate two channels. USD/JPY fell about 0.9% at one point to 156.69. It was near 156.84 late in New York. The day’s range was 158.25 to 156.69.

Lower Treasury yields erode part of the dollar’s yield advantage. That is an asset-allocation channel. Haven demand may also push investors to cut yen-funded positions. That is a carry-unwind channel. Both can operate at once.

Crude oil

The market raised the risk premium for supply disruption. December Brent rose $4.35, or 4.00%, to settle at $113.17 a barrel. That is the highest settlement since May 4. WTI rose 2.39% to $99.47.

On the afternoon of October 22, the Houthis issued a statement. They said they had carried out four military operations against targets inside Saudi Arabia over the previous 48 hours: the Saudi Ministry of Defense building, King Fahd International Airport, Aramco’s Yanbu facility and Abha International Airport. They said the third operation was a “direct hit” on Yanbu and started a fire. Saudi authorities did not immediately confirm the Houthi account of damage at Yanbu. The Saudi civil aviation authority confirmed only that King Fahd International Airport was attacked on the night of the 21st.

The probability of supply disruption rose.

European bonds

French fiscal risk is now in the price. The French 10-year yield touched 5.1415% intraday, the highest since July 2002. It eased to 5.06% late in the session. The French-German 10-year spread widened at one point to more than 180 basis points, the most since November 2011. It then narrowed to about 173.

French public debt reached 119.0% of GDP in the second quarter of 2026, the highest since 1946. Total debt was €3,595.5 billion, up €59.6 billion in one quarter. The French 5-year CDS spread broke above 100 basis points on October 22, the first time since October 2012. London Stock Exchange data show it had already risen from about 71.6 basis points at the start of the month to about 87.

The market now demands more compensation from the French government for fiscal risk.

Precious metals

Haven demand did not outweigh the other forces on price. London spot gold closed down about 2.79% at $4,025.33 an ounce. Spot silver closed down about 5.78% at $54.58. It touched $54.17 intraday, the lowest since November 2025.

Higher oil strengthens worries about inflation and rates staying high, which pressures non-yielding assets. Losses in tech stocks may also trigger risk-budget cuts and demand for cash.

Subscribe now

These prices do not contradict each other.

• Tech stocks care about future returns on capital.

• Treasuries care about the rate path and liquidity.

• Crude cares about physical supply risk.

• French bonds care about fiscal solvency.

• Gold responds to real rates, the dollar, and liquidity all at once.

The divergence shows that different assets are repricing different risks.

Mechanism and Inference

Thread one: AI capex moves from growth story to return test

The backdrop to the tech selloff on October 22 is that Meta kept raising its capital commitments while Tesla cut its capex outlook.

Meta added, and its cash flow and long-term contractual obligations came under pressure together. Third-quarter capex was $34.91 billion. Operating cash flow was $32.79 billion. Free cash flow turned negative at -$2.12 billion, against positive $10.63 billion a year earlier. Guidance for 2026 capital expenditure is $135 billion to $145 billion.

At the end of the third quarter, Meta’s obligations stood at about $355.7 billion of leases not yet commenced, about $454.5 billion of non-cancellable contractual commitments and up to about $14.7 billion of conditional cloud capacity purchase obligations. The total was about $825 billion, running as far out as 2036. About $121.4 billion of those contracts fall due in 2027. Three months earlier, the figure was $81.7 billion.

The leases and contracts cover data centers, power, servers, network infrastructure, and existing operations, so they cannot all be classed as pure AI investment. The key question is how much room the company keeps to allocate future funds, and how long capital takes to pay back.

The negative free cash flow is a stress signal. It is not proof of a solvency problem. Meta still holds about $82.45 billion in cash and marketable securities. Credit markets are more sensitive. According to ICE Data Services, Meta’s 5-year CDS spread rose by as much as 8.9 basis points intraday to 119.3. That is the highest intraday level since the contract began trading.

Tesla subtracted, which exposes a gap between the pace of capex and growth expectations. It cut its 2026 capex guidance from “at least $25 billion” to “more than $20 billion” (filing wording). Third-quarter capex was about $4.73 billion. Operating cash flow was about $6.26 billion. Free cash flow turned positive at $1.53 billion, against a market expectation of about $4.9 billion of net outflow. Strip out working-capital changes and free cash flow is close to a loss. Positive cash flow has to be separated into operating improvement versus changes in spending pace and payment timing.

Credit markets reinforced the doubt about returns. Morgan Stanley research shows weekly average CDS trading volume tied to Oracle, SpaceX, Alphabet, Amazon, Meta, Broadcom and Nvidia above $9 billion in September. In July it was under $2 billion. AI high-yield spreads widened about 55 basis points in one week to 377. Non-AI borrowers widened only about 2 basis points over the same period. Hyperscaler free cash flow may turn negative around 2027 is a forecast.

The mechanism is a negative feedback loop running through return on capital, valuation, and funding cost. Companies raise capex to earn higher future cash flow. If the new cash flow will not cover the cost of capital and the investment risk, valuations must fall. Lower valuations raise the cost of equity. Wider credit spreads raise the cost of debt. Both raise the hurdle for the next round of investment. If AI revenue and operating cash flow do not improve together, funding pressure can turn around and limit investment. Share-price pressure and wider CDS spreads are the same worry about returns, showing up in equity and credit.

This chain cannot be stretched into “AI companies will issue less debt, so Treasuries must rise.” Corporate bond supply affects the corporate credit market. Treasury supply directly affects Treasury supply and demand and the term premium. The two can connect indirectly through portfolio rebalancing. AI capital returns and funding constraints are the main axis of the repricing in tech stocks and related credit that day. They are not the confirmed main cause of the 19.3 basis point fall in Treasury yields.

Leverage and liquidity are amplifiers. Low volatility and easy funding encourage investors to add leverage, holding credit assets or other strategies for steady relative returns. When prices move against them, losses tighten risk limits and margin requirements rise. Lenders may raise collateral haircuts. Investors must post more collateral or sell assets. Falling prices cut collateral value and trigger another round of position cuts. Leverage built up in quiet markets can become forced selling in volatile ones. That explains why prices sometimes move far more than the initial news can account for. It does not prove large-scale deleveraging occurred on October 22. Confirming it would take options positioning, funding haircuts, margin changes, credit spreads and actual fund flows. So far, I don’t have the actual data.

Subscribe now

Thread two: Middle East risk, and crude pricing the odds of disruption

On October 21, the Financial Times, citing three people familiar with the plans, reported that the Israeli military had prepared a multi-day strike for the weekend around October 10. Targets allegedly included Iranian nuclear sites, government facilities, and civilian infrastructure. The operation was reportedly not carried out because Trump opposed restarting hostilities before the November 3 midterms. On October 22, Netanyahu publicly denied the report and called it false.

The US government’s public position is that it will not attack Iran before the election. White House deputy press secretary Anna Kelly restated this on October 21. She also denied that the US and Israel plan another joint strike on Iran’s leadership before the election. So political restraint on military action before the election is a public policy signal. Whether things escalate after the election remains a risk scenario.

Oil pricing does not depend on that political chain being proven. The Houthi statement of October 22 supplied a separate supply-risk signal. However, if the damage at Yanbu is eventually verified, risk to the Red Sea, the Bab el-Mandeb, and related export and shipping routes can change the market’s view of future supply.

Oil also feeds other assets through inflation and demand. Higher energy costs raise transport and production costs and cut household real purchasing power. That may raise expectations of persistent inflation and high rates. Prices that are too high also suppress consumer and industrial demand. That demand destruction eventually weakens the rally in crude.

Brent’s rise on October 22 reflects geopolitical threat combined with a higher probability of future supply disruption.

Thread three: The yen, where rate differentials, allocation and carry all act

Japan’s Ministry of Finance reported on October 22 for the week to October 17. Japanese investors were net sellers of foreign long-term bonds by ¥161.6 billion. That was the fifth straight week of net selling, the longest streak in four years. They also sold ¥156.9 billion of foreign short-term bills on a net basis. Japan’s 10-year yield is up more than 80 basis points this year, which makes domestic bonds more attractive in relative terms.

But ¥161.6 billion was the smallest week of the five-week run. The adjustment is continuing, and the pace of selling slowed that week.

Two channels are at work. The first is rate differentials and asset allocation. Japanese institutions holding dollar bonds compare more than nominal yields. They also have to account for currency hedging costs. As US yields fall and Japanese domestic yields rise, the hedged yield advantage on some dollar assets narrows. Long-term investors may gradually reduce overseas bond holdings. This usually plays out over weeks, months or even quarters.

The main force behind net selling of foreign bonds this year has been Japanese domestic banks. Life insurers manage about $2.5 trillion in assets, and overseas holdings are a large share. Whether they keep adding domestic bonds is a key indicator of whether repatriation can grow.

The second is the carry-unwind channel. The yen has long been a low-yield funding currency. Investors borrow yen to buy higher-yielding assets. When the yen rises, or risk assets fall sharply, the risk-reward of the carry trade worsens. Investors may sell risk assets and buy back yen to repay funding. If positions are crowded, the process can reinforce itself.

The more reliable reading for that day: the yen was supported at once by lower US yields, haven demand, and better domestic Japanese yields.

Subscribe now

Thread four: French bonds, where fiscal credit risk is priced, and a negative feedback loop is forming

French public debt reached 119.0% of GDP in the second quarter of 2026, the highest since 1946. The 2025 deficit was 5.1% of GDP. The government expects 5.4% in 2026. The French 10-year yield rose intraday to 5.1415%. The French-German spread was at one point above 180 basis points. The French 5-year CDS spread broke above 100 basis points intraday, the first time since October 2012. These moves reflect a higher risk premium on France’s ability to repair its finances.

Comparing France with Germany is a more direct way to see whether France’s own risk premium is widening. Non-residents hold 57.5% of France’s tradable central government bonds. That foreign ownership may amplify price sensitivity to changes in fiscal credibility.

The mechanism is a debt feedback loop. Deficits stay high, so the government must keep issuing. The market demands higher yields, so new issuance and refinancing of old debt cost more. Higher interest spending squeezes the room for fiscal adjustment.

French public debt interest spending is about €79 billion in 2026. It may rise to €91 billion in 2027, an increase of about €12 billion in a year. Higher rates do not reprice all existing debt at once. If funding costs stay high, though, interest spending gradually erodes fiscal space.

Debt sustainability also depends on nominal growth, the effective financing cost and the primary balance. If the effective interest rate stays above nominal growth and the government cannot run a large enough primary surplus, the debt-to-GDP ratio faces continued upward pressure. High debt does not automatically mean default. The danger is a debt trajectory that loses a credible path to stabilization.

A selloff in French bonds may lead investors to adjust European allocations.

The Fed: A Pause Is Not a Pivot

On October 28, the FOMC voted 10 to 2 to hold rates. Governor Lisa Cook and Cleveland Fed President Beth Hammack argued for another 25 basis point hike. The committee said economic activity is expanding at a strong pace, capital investment is solid, and inflation remains elevated.

The conflict is between a supply shock and demand management. Higher energy prices can push inflation expectations up, but a rate hike cannot add oil supply. High oil prices also erode household real purchasing power, reduce corporate profits, and may slow future demand. The Fed can influence borrowing costs and aggregate demand. It cannot directly repair energy facilities or remove geopolitical shipping risk.

The dollar rate curve is an important anchor for cross-asset pricing. It is not a universal variable that explains every asset.

• Tech stocks are driven by both future cash flows and the discount rate.

• Treasuries depend on expected short rates and the term premium.

• Gold depends on real rates and the dollar.

• The yen responds to US rates, risk aversion, and funding positions together.

• Crude responds more directly to physical supply and demand, inventories and geopolitical risk.

  Subscribe now

The Risk Is Whether Funding Feedback Connects

On October 22, 2026, the AI investment case began to move from a growth story to a test of capital returns. At the same time, high debt, energy supply risk, and monetary policy constraints re-entered asset pricing.

This was a multi-shock, cross-asset repricing. The dangerous stage is when risks in different markets start reinforcing each other through the same funding mechanisms. One asset falling on its own does not get there.

• If AI capital returns fall short of expectations, corporate valuations and credit risk may come under pressure together.

• If geopolitical risk causes sustained physical supply disruption, energy inflation may extend the high-rate environment.

• If France’s fiscal repair loses a credible path, a rising sovereign risk premium may erode its fiscal space.

• If leveraged investors are forced to cut positions by price swings, funding contraction may amplify a local shock further.

The next phase depends on whether return on capital, real funding cost, supply elasticity and debt sustainability start to worsen together. If those variables keep diverging, markets can adjust through relative price changes. If they begin to amplify each other through leverage, collateral, and funding conditions, the risk escalates from a local valuation reset to broader financial stress.

Subscribe now

Read more

Even though this publication might die eventually, I will try my best to feed you with my authentic perspective.

Europe faces a contradiction that is getting harder to avoid. Total employment has not collapsed under AI, but hiring growth is slowing. AI has not destroyed enough jobs to prove the “machines replace humans” prediction, yet entry-level roles and young people’s career starts already feel the pressure. Pensions, healthcare and defence compete for the same fiscal resources while the main funding base of social protection, labour income, is squeezed at both ends by demographics and technological change.

They already appear in 2026 official data and research reports.

The European Commission’s Employment and Social Developments in Europe 2026, published in October, puts the EU employment rate at 76.1% in 2025, unemployment at a historic low of 6%, and real wages 1.4% above their pre-pandemic level. The same report expects the EU to add about 440,000 fewer jobs in 2026 than previously projected, because of the energy shock linked to the Middle East conflict and slower growth. The labour market is still resilient; the marginal momentum of hiring is weakening.

• https://employment-social-affairs.ec.europa.eu/european-employment-and-social-outlook-2026_en

Information and communication, finance and insurance, and public administration rank higher. Accommodation and food services, administrative support and some low-skilled occupations rank lower, with less predictable hours and weaker protection. Those are also the sectors where labour shortages are most pronounced.

The European Training Foundation’s 2026 review reads the evidence more precisely. AI’s impact on work is currently mainly job transformation, not full occupational disappearance. Between 40% and 60% of jobs are expected to undergo substantial task change; only a small share of occupations can be fully automated. The review also flags a structural risk. Entry-level jobs are more exposed than senior ones. Routine cognitive tasks traditionally handed to junior staff are being automated or assisted by AI, and the bottom rung of the career ladder is weakening.

AI’s most immediate effect shows up in job quality rather than job numbers. Some workers get more engaging tasks. Others face higher work intensity, less room to use their skills and more psychosocial pressure.

The OECD’s social protection financing report, published in September 2026, states the funding problem without decoration. Labour income is the main source of social protection funding in most countries, and about half of social protection spending is financed through contributions that mainly target labour income. That base is under pressure from two directions at once. Population ageing shrinks the working-age population, so the total wage bill shrinks. Labour-substituting technology has not yet caused net job destruction, but it may raise capital productivity and push up capital’s share of total income.

Two figures from the same report sharpen the picture. Social protection spending averages 23% of GDP across the OECD and the EU. Self-employed workers are 14% of EU employment but contribute only about 4% of social security contributions.

Europe’s problem therefore has three layers.

• Growth and high-quality job creation are insufficient.

• AI may change how income splits between wages and capital returns.

• The funding base of the traditional welfare system is squeezed by demographics and technology together.

Welfare is not only a fairness issue or a fiscal burden. It has an economic function. When wages and employment cannot distribute enough purchasing power, public services, social insurance and income transfers sustain basic consumption. They share economic risk, and they help people share in productivity gains. That is the link between Europe’s welfare crisis and the AI revolution.

Subscribe now

Read more

75% human writing draft + 25% AI editing, grammar checking, etc.

“There is no truth. There is only perception.”

As a slogan, the line has force. As an analysis, it should not be the end point. Facts do not disappear because positions differ. The risk is what happens next. Once a fact enters an interpretive frame, each stakeholder gives it a different meaning. Those meanings shape judgment and behavior.

Inside companies, this mechanism runs every day in the competition for positions.

Position, power, and year-end bonus are tightly linked. The number of promotable roles is limited. Visible work, work that is easy to show as a contribution, builds reputation faster. Important work that is quiet and hard to quantify can be ignored. Your colleagues may not tell you everything. When interests conflict, they may share information selectively. They may mislead you. Your promotion may cost them the same opportunity.

Your superior tells you to wait and build seniority. You believe that continued effort will eventually pay off. Then another colleague is promoted first.

Subscribe now

Seniority is not worthless. But seniority does not convert into promotion on its own. Your contribution has to be seen. It has to be understood. It has to be recognized by the people who hold decision rights. Often, career reward depends on more than how much value you created. It depends on who can define that value, and who holds the power to distribute the return.

None of this means workplace competition is necessarily deceptive. It means that when interests conflict, another person’s promise is not an institution. And personal effort is not a guaranteed payout.

Financial markets work the same way.

A narrative can attract attention. It cannot replace contract terms, formal approvals, cash flow, and property rights. A grand story can get investors to bet early. It cannot guarantee that the future follows the story.

I have written this newsletter for some time. Many paying subscribers have stopped renewing. Paying readers are now in the single digits. This is a real commercial result, and I have no reason to avoid it. My readers are smart and clever. The most I can say is that I did not offer something they could not get elsewhere.

So I have to admit this. An article can break down mechanisms and reconstruct cause and effect, and still lack commercial value. Readers do not pay continuously just because an author is careful and avoids exaggeration. They need to see what the analysis gives them that others do not. That could be better judgment, a clearer causal chain, or a more usable decision framework.

This is where I have to look at myself. The problem is not only how to write more accurately. It is also how to make the judgments sharper and the arguments more penetrating. And it is how to make readers understand why continued reading is worth paying for. That is what I will improve.

My earlier software career gave me another way of looking at the world. I analyzed large datasets, traced workflows, and diagnosed system failures. I rebuilt causal chains from inputs, rules, dependencies, and abnormal outputs. I believe data and mechanism can help build a basic factual framework about reality. They let us avoid simply accepting the explanation of an authority figure.

Society is not a fully predictable program. Data cannot cover all of reality. But mechanism analysis forces a few questions.

• How was this result produced?

• Which conditions cannot be missing?

• Who has the power to change the rules?

• Which link is most likely to fail?

The most dangerous thing in the market is a free opinion.

Subscribe now

Narrative Describes Reality and Also Controls Behavior

In the White House, in government departments and in corporate offices, narrative is more than a way to communicate. It is a management tool. It shapes expectations.

A company announces a strategic transformation period. Employees reassess

• what work is worth doing,

• what risks are acceptable,

• and which kinds of dissent will look out of place.

Management describes cost cuts as an efficiency reform. Employees may read

• their own losses as a necessary sacrifice.

• Extra effort gets defined as a shared mission.

• The cost an individual carries can be repackaged as a sense of responsibility.

Politics is the same.

“America First” places policy inside a frame of national interest first. “America is back” stresses that America is again playing its role and exerting its influence. The two slogans come from different political contexts. Both show the same thing. Political narrative first supplies an interpretive frame. Then it guides how the public understands policy, assigns responsibility, and judges success or failure.

Narrative works best not because it lies. It works because it decides what people notice first and how they act afterward.

The media industry shows the commercial root of this mechanism.

In 1833, Benjamin Day founded the New York Sun. He sold it for one cent to widen the readership. Then he sold the gathered audience’s attention to advertisers. He found that a newspaper’s commercial value did not have to come mainly from the price readers paid. If enough people were attracted, attention itself could be the product.

The model also changed the incentives of news production. To win readers, crime, scandal and sensational stories gradually gained the upper hand. In 1835, the Sun even ran a false report that humanoid creatures lived on the moon. The sensation helped circulation grow sharply.

The point of this history is not only that early media faked stories. The point is how commercial competition rewarded it. When revenue depends on attracting and holding attention, content that triggers fear, anger, curiosity, and group identity can outcompete explanations that are accurate, restrained, and complicated.

The internet did not change this basic logic. It raised the efficiency of it by a large margin.

In You Are the Product, John Lanchester goes further. Social media sells users’ attention. It also turns behavioral data, social ties, and psychological reactions into commercial value. The platform watches what people look at, whom they interact with, and what they react to. Then algorithms decide which content gets more exposure.

Users think they are browsing friends’ lives and looking for useful information. In fact, the order of information is also shaped by the platform’s commercial interest. People imitate others. They compare their lives. They seek approval and social status. These psychological mechanisms belong to ordinary relationships. They have gradually become resources that can be commercialized.

Free services are not free of cost. Users pay with time. They also pay with behavioral data, social relationships, and continuous attention.

This also explains why complex financial questions are often reduced to simple conflict, reversal, then conflict again. A story only needs to make readers pick a side to earn engagement. An article that explains how contract enforcement, property protection, cost of capital, and political mandate relate to each other asks the reader for much more time.

So media competition can produce a systematic bias. Content that attracts attention is rewarded. Content that helps readers understand reality is not necessarily rewarded the same way.

Narrative does not need to change the facts. It only needs to change the order in which facts appear, the attention they receive, and the meaning people give them. That is enough to change behavior.

The Venezuelan oil deal illustrates this.

Share

Venezuela: A Restarted Oil Deal Does Not Mean the Underlying Risk Is Fixed

In January 2026, a US military operation captured Venezuela’s then-president, Nicolás Maduro. Delcy Rodríguez has since led the country as interim president. Relations between Washington and Caracas have moved step by step toward cooperation. Oil is the central issue in the readjustment.

An oil arrangement disclosed on August 31 covers 17 oil fields and plans to deploy about 250 drilling rigs. On October 17, Rodríguez said Venezuela expects nearly $76 billion in oil investment over the next five years.

These numbers easily form a complete and attractive story.

• Political change opens the market.

• An open market attracts capital.

• Capital restores production.

• Oil revenue drives economic recovery.

The story has a real basis.

Venezuela holds the largest proven oil reserves in the world. Long-term underinvestment, aging infrastructure, operating problems at the state oil company, and US sanctions have severely limited its production capacity. If investment conditions improve, equipment is maintained, and production recovers, the oil sector does have room to recover.

Ask yourself, why does Rodríguez say so? Who is her boss?

Share

My judgment is clear. The opportunity for a Venezuelan oil recovery is real. The claim that the market has already cleared the risk does not hold.

First, the $76 billion is a government-announced investment expectation for the next five years. It does not mean funds of that size have arrived. It certainly does not mean the related capital spending, output and economic growth have materialized.

Second, the legal structure of the oil deal, the transparency of the bidding, and the formal approvals still need clarifying. Reuters reported on the arrangement on August 31, 2026. It explicitly raised the absence of transparent competitive bidding and the unclear status of some formal approvals. Even if the deal is moving forward, that does not confirm the legal basis, approval procedures, and execution conditions of the related contracts.

Third, Rodríguez holds real executive power. That does not mean Venezuela has completed political authorization and institutional rebuilding.

As of October 2026, the government and the opposition are still negotiating political participation, election arrangements, electoral bodies and Supreme Court reform. No national election date has been set. Venezuelan gold reserves previously frozen in the UK are still subject to a legal dispute.

On October 21, Reuters reported further that US prosecutors cited a federal law on torture abroad in the revised Maduro case. Maduro and his wife had already pleaded not guilty to the added charges. A criminal charge is not a court conviction. But the development is a reminder to investors. Venezuela’s political and legal process is still evolving. It cannot simply be treated as finished.

These facts point to one larger question. Can the current executive arrangements become an institutional arrangement that is sustainable, enforceable, and able to withstand future political change?

Venezuela is like a business re-entering the international market. Investors start asking for prices. But who holds final decision power? Are the operating licenses stable? Has the wiring been inspected? Who rules when a dispute arises? Each of these still has to be confirmed one by one.

Investors asking for prices shows that demand is returning. It does not prove the risk is gone. Likewise, an oil deal being pushed forward shows that a commercial opportunity is forming. It does not prove that contract enforcement, political authorization, and property protection are adequately secured.

A contract being negotiated does not mean contract risk is removed. Capital starting to enter does not mean institutional stability is confirmed.

This does not deny the investment opportunity in Venezuela. It refuses to equate an investment opportunity with investment safety. Every investment carries risk. That is a correct platitude, but it is still true.

Why Capital Prices the Story First, and Who Pays

Financial markets do not wait until every fact is settled before acting. Asset prices are built on future expectations. Investors set the price they will pay from expected cash flow, probability of realization, and required risk return.

The nearly $76 billion investment expectation for Venezuela can influence market judgment. It cannot answer, by itself, the questions that decide investment value. How much money has actually been invested? Which contracts have completed formal approval? Who bears the cost of construction, maintenance, and operation? How will future oil revenue be split? Can funds legally be repatriated? If the government changes, the regulatory rules change, or the contracts are challenged, what law gives investors a remedy?

More questions follow. Can the planned rigs turn into actual output? Can output turn into oil that can be sold? Can sales revenue finally turn into cash flow the investor can recover?

None of these links convert automatically.

Resource reserves are not production capacity. Production capacity is not sales revenue. Sales revenue is not money the investor can finally take back. Each link has its own cost, constraints, probability of failure, and decision-maker.

The same standard has to apply to the media industry.

An advertising-driven media model mainly rewards attention. A direct subscription model makes continued reader payment a commercial feedback the author must answer. Both models can produce bias. But they punish wrong judgment differently.

An author who depends on clicks can keep earning from traffic, even after repeatedly exaggerating forecasts or amplifying risk. An author who depends on long-term subscriptions faces a different result if he keeps confusing fact with forecast, or replaces analysis with emotion. Readers can stop paying.

Subscription does not guarantee truth, of course. An author may cater to the existing views of paying readers. Readers may pay only for views that match their own bias. A payment model changes the lines of accountability. It does not replace independent judgment.

For me, subscriber loss is not a result I can easily push away. It is commercial feedback I must take seriously because I like money. It forces me to test two things at once. Is my causal analysis reliable? And does that analysis give readers a judgment that is clear enough and useful enough?

I do not pretend to know every market outcome in advance. I try to reconstruct the truth.

On Venezuela, my conclusion is this. It is currently a transition deal with real economic potential, but with significant political and legal risk still attached. It is not a mature investment market that has finished clearing its risk.

Even if my paid subscriptions fall to zero, I will keep writing from data, mechanism, interest structure, and cause and effect. I believe this method can produce clearer judgments, and not only a display of the analytical process.

Also, my final software product will answer who I am and prove my commercial value one day.

Thanks.

I hope my newsletter will be financed by my readers one day.

CPI: Gasoline lifts the headline

September CPI rose 3.7% year on year. That is the highest since May, above August’s 3.4% and the 3.6% consensus. Month on month it rose 0.6%, the largest in five months. Core CPI rose 2.6% year on year, up from 2.4% in August and above the 2.5% consensus. Core rose 0.3% month on month, unchanged from August and slightly above the 0.2% consensus.

Subscribe now

Gasoline was the most direct source of the rebound. BLS says gasoline rose 9.1% on the month and contributed more than half of the headline increase. The energy index rose 5.2% on the month and 20.4% year on year. Gasoline was up 34.2% year on year and fuel oil 74.2%. That is a sharper statement of the energy character of the month than “gasoline was more than half.”

Energy is split internally. Energy commodities rose 9.3% on the month: gasoline 9.1%, fuel oil 15.3%. Energy services rose only 0.1%. Electricity rose 0.2%, and natural gas fell 0.3%. The shock is concentrated in goods and has not yet spread through utilities. That quiet in energy services is one reason core has not run further. It also means the catch-up risk in utility costs has not been released if crude and product prices stay high.

Housing needs separate treatment. Housing rose 0.4% on the month, up from 0.3% in August. Owners’ equivalent rent rose 0.3% and rent 0.2%. Housing is up 3.2% year on year, still below headline at 3.7%, but the monthly pace is accelerating. Housing still pulls the total down, but it is no longer cooling quickly. That cuts against the old story that housing disinflation is enough to hold core down.

Core services are split too. Airfares rose 1.2% on the month and 22.4% year on year, the strongest upward force in core services. Hospital services rose 0.7% on the month and 5.6% year on year. Motor vehicle repair rose 0.7% on the month and 5.8%. Against that, prescription drugs fell 1.2% on the month. Motor vehicle insurance rose only 0.1% on the month and is still down 5.0% year on year. Communication fell 0.2%. Core services are not accelerating across the board. But stickiness in airfares, medical care and repair is enough to keep supercore firm.

The inversion between core CPI and core PCE is another anomaly. KPMG’s September 30, 2026 piece, “Cooler on paper, hot underneath,” states that core PCE rose 0.2% on the month in August and 3.0% year on year, against 2.4% for core CPI. Core PCE has been above core CPI for 10 straight months, the longest inversion since 1983. The usual relationship has flipped, and KPMG calls that bad for the Fed. BEA has since done a benchmark revision to PCE components. July core PCE year-on-year was revised down from 3.3% to 3.0%. KPMG says the revision cut about 0.3 points from annual inflation, and July headline is now 3.4%.

The inversion comes mainly from weights and coverage. Housing weighs much more in core CPI than in core PCE, so when housing inflation falls from its peak, CPI gets the larger downward pull. Financial services, computer software and some medical-related services may matter more in PCE. In September CPI, computer software and accessories rose 28.4% year on year. The item has a CPI weight of only 0.031% but about 1% in PCE, so it lifts PCE far more than CPI. The point is not that one index briefly exceeds the other. Housing cooling can no longer stand in for the whole core inflation system. Housing’s monthly acceleration in September shows that.

CPI shows the energy shock now. The split among core measures shows what to watch: whether pressure spreads from energy items into the broader core price system.

• https://www.bls.gov/news.release/pdf/cpi.pdf

  Share

PPI and diesel: Where cost pressure sits

September final demand PPI rose 1.3% on the month, the largest since March 2022. It rose 6.3% year on year, the largest since December 2022. Final demand goods rose 2.6% on the month, the largest since the series began in December 2009. BLS says more than 80% of the goods increase came from final demand energy, which jumped 10.8% on the month. Final demand gasoline rose 20.3%, and BLS says half of the goods increase traces to gasoline. Diesel, fresh and dry vegetables, jet fuel, residential electricity and crude also rose clearly.

Stopping at final demand goods misses the more forward-looking signal:

Those prices rose 2.5% on the month, the largest since May, and 14.1% year on year, the largest since August 2022. BLS says nearly half of the September increase in processed goods for intermediate demand traces to a 21.0% rise in diesel. Diesel is up 115.2% year on year. It had already risen about 23.6% on the month in August. Unprocessed goods for intermediate demand rose 8.4% on the month, the largest since February 2022. Crude jumped 28.3%, and unprocessed energy materials rose 22.2%. By production stage, stage 2 intermediate demand rose 3.9%, driven by crude, diesel and gaseous fuels. Stage 1 rose 2.2%, driven by diesel, gasoline, grains, crude and truck freight.

Gasoline and diesel sit at different points in the price chain, and which rose more in a month is not the point. Gasoline lifts what producers charge for final goods, so it shows the immediate energy hit. Diesel is an input to trucking, farming, construction, and industry. Its rise first changes the cost structure of production and distribution, then passes through purchasing, processing, transport, wholesale and retail. Producers may reprice final goods directly. But intermediate input increases must pass through downstream cost structures, margins, inventory cycles and contract terms before they set final prices.

Final demand excluding food, energy and trade services still rose 0.5% on the month and 5.0% year on year. That does not show all core goods and services accelerating. It does show producer price pressure is not fully explained by energy. Final demand services rose 0.8% on the month. Trade margins rose 1.4% and transportation and warehousing 3.4%. Truck freight rose 3.8% on the month and 19.3% year on year.

Diesel matters for its position in the production network. Gasoline mostly tracks household driving. Diesel runs through trucking, farm machinery, construction equipment, and industrial logistics. A heavy truck running 100,000 miles a year at 6.5 miles per gallon burns about 15,400 gallons. At $6.529 a gallon, that is about $100,000 a year in fuel. Fuel is an operating cost for carriers. It is also a required cost for getting goods from producers to warehouses, retailers, and consumers.

EIA weekly data show US retail diesel hit $6.529 a gallon in the week of September 21, the highest in the series. It has since eased, but was still about $6.179 in the week of October 12.

A 20% rise in diesel does not mean 20% higher costs for every diesel-using good. Four mechanisms govern the path from raw material to final consumption:

Subscribe now

Read more

Cheap capital has been ending since 2023. This new era at least last for ten years.

JPMorgan CEO Jamie Dimon spoke to Bloomberg in London on October 6, 2026. He put government funding needs, AI capex and corporate funding costs into one problem. Global capital demand and government borrowing are both rising. If funding costs rise further, the pressure can pass into corporate credit. That means wider corporate bond spreads and higher refinancing costs. His words on the process were: “The market will keep demanding more and more. At some point it will transmit to corporate bonds, corporate credit spreads and so on. That is usually how it happens.”

He did not predict a bond crisis on a given day. He told companies to plan for worse financing conditions than expected, whatever their current leverage. His question:

This is not new. JPMorgan’s April shareholder letter already warned that AI, defense, and infrastructure spending, on top of large fiscal deficits, would keep raising capital demand. On October 13, Dimon went further. He did not say the 10-year Treasury yield must reach 6%. He said investors and borrowers should be prepared for that possibility.

A higher benchmark rate is only the first layer. Companies face two costs that can move together:

• the benchmark yield rises, lifting the base cost of new financing;

• and credit spreads widen, because investors want more compensation for default risk, liquidity risk and fragile balance sheets.

Together, corporate funding costs can rise faster than the policy rate.

Higher rates do not automatically mean a worse financial environment. If yields rise because productivity and earnings prospects improve, companies may earn more too. If yields rise because of heavier fiscal supply, repriced inflation risk, or demands for more term premium, companies pay more before their revenue improves. Judging funding risk means asking why rates moved, and whose cash flow absorbs the move.

Bridgewater founder Ray Dalio estimates the US must roll over about $10 trillion of maturing debt this year and borrow about $2.1 trillion in new money. Rollover is refinancing of maturing debt.

AI infrastructure is a separate, very large source of new funding demand. In a September 17, 2026 report, Goldman Sachs raised its estimate of 2027 global investment-grade issuance by five tech giants (Amazon, Alphabet, Meta, Microsoft and Oracle) from about $400 billion to $420 billion, about 68% above the $250 billion forecast for full-year 2026. Goldman also expects debt to fund about 35% of these five companies’ 2027 capex, roughly $1.2 trillion. The $420 billion excludes data center and chip project financing.

The wider picture: by mid-September 2026, global AI-related debt issuance had reached $578 billion, and the IIF in early October projected full-year 2026 AI-related corporate debt issuance would approach $700 billion. Hyperscalers were only about 40% of it; the rest came from data center projects, semiconductors, software, equipment, and power and energy. By market, about $490 billion came from global investment-grade markets and about $88 billion from leveraged finance. AI financing has spread from chip and cloud companies into the physical infrastructure chain.

At the other end, a large stock of corporate bonds and leveraged loans issued in the low-rate years is coming up for repricing.

These three forces do not mean there is necessarily too little cash overall. The bond market can adjust supply and demand through yields, term premium, and credit spreads. The adjustment has a cost:

Higher yields put more valuation pressure on existing assets; wider spreads make new financing more expensive; cash on hand does not mean funding capacity is ample; an asset with value does not necessarily convert to debt-service cash in time.

The scarce thing is risk-bearing capacity

Dimon’s discussion of shifting capital supply and demand contrasts with Bernanke’s “global savings glut” argument around 2005. Back then, global savings were ample relative to investment demand. Large flows into bonds pushed long-term yields down. Now fiscal financing, AI investment, and legacy refinancing are expanding together. The market must absorb more bond supply and more credit risk.

That does not mean global savings are absolutely short. More precisely, new financing demand is testing three capacities:

• absorbing long-dated debt,

• bearing credit risk,

• and supplying balance sheet

…room in stress.

The bond market is not a container that absorbs unlimited new debt once yields are high enough. It depends on the balance sheets of banks, insurers, pensions, asset managers, and hedge funds. Those institutions face investment mandates, capital requirements, liquidity management and risk limits. An insurer may hold a lot of Treasuries and still be unwilling to add the same amount of high-yield corporate debt. A bank with funding access may still cut lending because of collateral haircuts or capital constraints.

Subscribe now

How bond prices, collateral and funding costs feed back on each other

The market price of a fixed-rate bond moves inversely to its yield. When the required yield rises, existing bond prices fall. New corporate issues pay a higher benchmark rate and may also carry a wider credit spread. Bond prices also set the terms of collateralized funding.

Financial institutions borrow in the repo market using bonds as collateral. When collateral prices fall, lenders may raise haircuts, call for more margin, or cut credit lines. A borrower who cannot meet the call may sell assets to repay. The selling pushes prices down further. Falling asset prices, lower collateral values, tighter funding, and forced sales can form a self-reinforcing loop.

The US Treasury market in March 2020 showed this. Treasuries are seen as high-quality assets. Yet when many institutions rushed to raise cash at the same time, forced selling hit the Treasury market too. An asset’s credit quality and its ability to be sold smoothly under stress are different questions. Liquidity is not measured by normal-day trading volume. What matters is whether dealers are still willing and able to take assets when many institutions need cash together.

The US has a very large Treasury market, the deepest investment-grade corporate bond market in the world, and the most concentrated group of AI infrastructure investors. The contest turns on which investors’ balance sheets are taken up by long bonds, short-term funding, high-rated assets and high-risk projects together.

Subscribe now

Read more

Demand growth only decides what the market needs. It does not decide who gets the profit. Four things decide whether a company benefits:

• which scarce resource it controls,

• which trading options it holds,

• how much of a price change it can turn into revenue,

• and how much cash flow remains after operating costs, contract liabilities, risk losses, and tied-up capital.

Hormuz transit risk, damaged Qatari LNG facilities, the repricing of marine war-risk premiums, violent swings in tanker freight, and rising digital infrastructure investment in the Middle East look like separate industry stories. They are one repricing.

Geopolitics changed the constraints on economic activity. The commodity may still exist but cannot be delivered on time. The ship may still be available but will not enter a dangerous channel. The buyer may be willing to pay more, then find that shipping and insurance have eaten the margin.

War does not create evenly spread excess profit. It redistributes pricing power, risk-bearing and cash flow among the participants in the chain.

Subscribe now

Read more

Why yields rose is only part of it. The rest is how price changes travel through pricing, trading, funding and risk control to the wider market.

This global bond adjustment reprices inflation, fiscal deficits and long-term rate expectations. It also tests how much price shock the financial system can absorb.

Three Markets, Three Fiscal Constraints

The US, Japan and France differ in fiscal structure, monetary policy and politics. In October, they faced the same market problem. When uncertainty over inflation, bond supply and future rate rises together, governments must offer higher yields to keep attracting money.

United States

On October 7, the 10-year Treasury yield briefly touched 5.36% intraday, the highest since 2002. The 30-year yield rose to around 5.7%, also the highest since 2002. The long-end selloff pushed the market to reassess inflation, fiscal supply, and term risk.

That day, the Treasury sold $39 billion of 10-year notes. The auction yield was 5.30%, the highest since November 2000. It came in below the pre-auction when-issued yield of about 5.317%, a stop-through of roughly 1.7 basis points. After the result, the 10-year yield fell quickly, at one point to about 5.27%. The bid-to-cover ratio was 2.77, the strongest since 2016. Indirect bidders took 80.3% and direct bidders 17.1%. Primary dealers took only about 2.5%, far below the recent average of about 8.5%.

Subscribe now

Read more

I set this free permanently too. Anthropic is one of my respected companies. It is going to be listed by November. Those who are interested in investing can read this thoroughly.

On October 6, 2026, New Constructs published a report signed by senior investment analyst Kyle Guske II. It called Anthropic’s IPO “the most ridiculous IPO of 2026.”

Anthropic’s annualized revenue run rate has passed $65 billion. It may exceed $100 billion by the end of 2026. The market is discussing an IPO target valuation of about $2 trillion.

New Constructs says the company has not shown a sustainable business model. Its reverse discounted cash flow model puts Anthropic’s fair enterprise value at about $150 billion. That is less than one-tenth of the market’s target.

The core claim is this. To support $2 trillion, Anthropic would have to generate profit equal to twice Nvidia’s net income over the past year. Nvidia’s net income over the last four quarters exceeded $190 billion. Anthropic’s 2025 revenue was $4.6 billion, with a net loss of $42 billion.

The most useful part of the report is the question it works backward from. If today’s price is reasonable, what revenue, margin, and return on capital must the company deliver?

David Trainer, New Constructs’ founder, has long challenged hot IPOs and richly valued tech stocks. In 2019, the firm called WeWork “the most ridiculous IPO.” WeWork withdrew its listing plan and filed for bankruptcy protection in 2023. In 2021, the firm gave Allbirds’ IPO a negative rating and put fair value as low as $119 million. In 2026, the company sold its assets for $39 million. Michael Burry also voiced concern in late September 2026 about the OpenAI and Anthropic listings and about AI valuations.

Subscribe now

Being right once does not mean being right now. Being wrong on one growth company does not make skepticism about high valuations worthless.

The questions are :

• whether AI has value,

• whether that value can become corporate profit,

• whether that profit can become shareholder cash flow,

• and whether the price paid today has already used up the future.

Judging New Constructs requires separating three risks. They can occur together, but they are not the same.

• A price bubble means the market pays too much for overly optimistic operating expectations.

• A business-model bubble means revenue growth does not become sustainable profit and cash flow.

• A narrative bubble means technical progress is treated as commercial success, and then as shareholder return.

AI can keep advancing. Anthropic can become a leading global player. An investor who buys at $2 trillion can still lose heavily.

What $2 trillion requires

The bulls have grounds.

Anthropic’s 2024 revenue was about $386 million. In 2025 it rose to about $4.6 billion, roughly twelve times higher. Media accounts of pre-IPO disclosures put Q1 2026 revenue at about $4.73 billion. Q2 revenue exceeded $11.5 billion, against $787 million a year earlier. The same disclosures show Q2 adjusted operating profit of about $559 million. That is the first profitable quarter for a frontier AI company.

By the end of July 2026, Anthropic’s disclosed annualized run rate had passed $65 billion. At the end of 2025, it was about $9 billion. Pre-IPO materials and reports say the company wants about $100 billion by the end of 2026. They also say it expects about $190 billion to $200 billion by 2028.

This explains why the bulls will pay a high price. The growth does not rest only on distant technical imagination. Large-scale commercial demand already exists.

One distinction is needed. An annualized run rate is not realized full-year revenue. It is not future free cash flow either.

A run rate annualizes revenue at a point in time or over a recent period. It shows how fast the business is expanding. It does not guarantee that current usage lasts twelve months. It does not show that new revenue becomes profit. Pricing, model mix, and compute demand are all changing fast. Extrapolating short-term acceleration mechanically into long-term growth overstates how stable the revenue is.

Financing valuation, enterprise value, equity value, and IPO target valuation are also different things.

In May 2026, Anthropic closed a $65 billion Series H at a $965 billion post-money valuation. The roughly $2 trillion under discussion is a potential listing target. The first comes from a specific financing deal. The second reflects what public markets might accept. Neither means every share can be sold at the same price.

Investors still need to check the share classes at listing, convertible instruments, fully diluted share count, possible new issuance, equity compensation, and voting rights and special terms across share classes.

New Constructs’ reverse DCF starts from the roughly $2 trillion target and works back to the operating results required. The model uses a 9.33% weighted average cost of capital. That figure is the average for seven companies:

• Microsoft,

• Apple,

• Alphabet,

• Amazon,

• Nvidia,

• Meta

• and Oracle.

It counts about $518 billion of contractual commitments as a liability assumption. All three scenarios assume 1,000% revenue growth in 2026 and another 100% in 2027. What separates them is revenue growth and after-tax operating margin from 2028 onward.

Subscribe now

Scenario one: revenue CAGR falls to 20% for 2028 through 2035. After-tax operating margin rises gradually and holds at 20%. The model gives a current enterprise value of about $150 billion. That is about 92% below the $2 trillion target.

Scenario two: later revenue CAGR rises to 36%, with after-tax operating margin held at 20%. The model value is still well below $2 trillion.

Scenario three: to support $2 trillion, Anthropic must sustain 56% annual revenue growth from 2028 through 2035. After-tax operating margin must rise quickly to 10% and hold there for ten years. On those assumptions, 2035 revenue is about $3.6 trillion. After-tax operating profit is about $358 billion.

That revenue is about 1.4 times the combined trailing twelve-month revenue of Microsoft, Apple, Alphabet, Amazon, Nvidia, Meta and Oracle. The profit is about twice Nvidia’s after-tax operating profit over the same period.

$3.6 trillion in revenue is mathematically possible. The operating requirements behind it are extreme. If growth duration, margin, capital needs, or competitive structure falls clearly short, today’s price loses its support.

The financials show Anthropic is still far from a mature profit structure. Reuters, reporting on pre-IPO disclosures, says the 2025 net loss was about $42 billion. About $34 billion of that was a non-cash accounting loss from changes in the value of convertible financing instruments. Excluding it, the operating loss still widened from about $2.98 billion in 2024 to about $8.06 billion in 2025.

In 2025, compute and infrastructure spending was about $7.33 billion. That was more than half of roughly $12.65 billion in operating expenses, and more than triple the 2024 figure. Cash, cash equivalents, and short-term investments at the end of 2025 were about $20.28 billion.

These numbers need to be read separately. A non-cash accounting loss can significantly affect reported net income. It does not mean an equal cash outflow happened. Compute and infrastructure spending is real resource use for running the business, training models, and expanding service. Net loss alone does not prove the business model failed. The non-cash item does not erase the operating loss or the cash need either.

New Constructs’ model also needs scrutiny. Contractual commitments are not legally the same as ordinary financial debt. A 9.33% discount rate borrowed from mature tech giants is not the only reasonable rate for a growth company. It has not shown sustainable profit and still needs heavy capital. If Anthropic needs more cash and compute, its real capital need could be higher than the model assumes. New Constructs admits it did not have the full S-1. Its analysis relies on financial information disclosed through the media.

So the test isn't just revenue growth or margin. The test is whether new invested capital earns enough. If returns on new capital stay below the cost of capital, growth destroys value. Growth earns a long-term valuation premium only when operating returns persistently exceed the cost of capital.

$150 billion, $1.1 trillion and $2 trillion are outputs under different assumptions. They are not objective values beyond dispute. A reverse DCF shows how much success the market price implies. It does not replace judgment.

Share

The closed-model profit problem: leading capability does not mean earning more

New Constructs is not only questioning growth speed. It questions whether closed foundation models can keep enough profit premium over the long run.

The report says: “Since the emergence of open-source models, it has become clear that closed models will struggle to generate profits.”

Frontier capability used to be scarce. Companies paid clearly more for better reasoning, code generation, and agent execution. Open-weight models now handle more standardized tasks. Competition shifts from “whose model is strongest” to “what is the extra capability worth.”

Open-weight and strictly open-source are different. Open-weight models let users obtain the weights and deploy them. Training data, modification rights, commercial terms, and license scope may not be fully open. They still give enterprises a new alternative.

New Constructs’ price comparison says some API prices for Claude Fable 5 and Mythos 5 are about 18 to 23 times DeepSeek R1’s. That reflects specific models and a specific pricing basis. Real task cost also depends on input and output token prices, task difficulty, success rate, and the number of compute runs needed to finish the same task. A unit-price multiple alone does not show the overall cost gap.

Silicon Data’s token spending index fell from about $2.0651 per million tokens in late May 2026 to about $0.97 at the end of August. That is a 52.4% drop. FT analysis says the index blends three different kinds of deflation into one number. Technical deflation means token factories get more efficient, which widens margins and helps the whole ecosystem. Structural deflation means enterprises route low-priority requests to cheaper open-weight models, which erodes frontier labs’ pricing power. Competitive price deflation means a margin war. Reading the index simply as “closed-model pricing power is collapsing across the board” is inaccurate.

Falling prices do not automatically collapse industry profit. Jevons’ paradox offers another possibility. When a resource gets cheaper to use, applications and volume can expand and offset part of the unit-price decline.

Vercel AI Gateway data shows the split. Vercel’s September 2026 AI Gateway production index says open-weight models handled 56% of token volume in August 2026. That is the first time they exceeded all closed models on the platform combined. The share was 7% in December 2025 and 13% in April 2026. Vercel CEO Guillermo Rauch also said the share briefly reached 62% on August 22. Open-weight models account for only about 14% of the platform’s spending.

Usage share is not revenue share. Revenue share is not profit share. This distinction matters in judging AI business models.

The same report shows Anthropic has not lost all pricing power.

• In August 2026, it still held about 64% of Vercel AI Gateway spending. Since December 2025, it has held at least 61% every month. Price migration has still appeared within its own product line.

• In August 2026, Fable 5 lost two-thirds of its gateway spending share. Opus 5, priced at about half as much, tripled its share. About 90% of the teams that had used Fable 5 cut their use of it. More workloads moved to the cheaper Opus 5.

Customers did not leave Anthropic across the board. They moved to cheaper options inside its product line. That is evidence that customers accept the Claude lineup. It is also a signal that the high-priced model is under price pressure.

Open-weight models are quickly replacing closed models on high-volume, low-value tasks. On high-value tasks, Anthropic still holds a 64% spending share. It has not lost all pricing power. The more direct risk is the Fable 5 to Opus 5 migration. It shows customers actively moving to cheaper options even inside Anthropic’s own products. That is direct evidence of pressure on the profit premium.

The bulls also have substance. Ramp’s May 2026 US enterprise adoption data shows 34.4% of firms paying for Anthropic. That passed OpenAI’s 32.3% for the first time. Some industry analyses estimate Anthropic’s high-end API inference gross margin has risen to 70% to 85%. Q2 adjusted operating profit turned positive. FT cites analysis showing frontier-model gross margins of about 70%, against about 20% for open-weight model companies. That gap appears to have been fairly stable over three years.

Share

These are different metrics. Enterprise adoption measures breadth of customer uptake. Inference gross margin measures profit from delivering model service after the related costs. Adjusted operating profit also reflects other operating expenses. Even a high inference gross margin does not mean the same share of revenue becomes free cash flow.

Subscribe now

The question to ask is why customers stay. One answer is that the model keeps getting better on real tasks. The other is that customers have embedded the product in workflows and are not ready to switch.

• The first can support a price premium.

• The second mostly reflects switching costs. Both create commercial value. Their durability and bargaining power differ.

New Constructs does not need to prove that closed models can never make money. It needs to show that the profit premium closed models can keep is not enough to cover long-term R&D, inference, channel, and infrastructure spending.

“Good enough” open models: four conditions at once

“Good enough” has no fixed performance line. It is a commercial threshold set jointly by

• task performance,

• deployment cost,

• workflow stability,

• and total enterprise lifecycle cost.

Matching on any single dimension does not prove open-weight models have fully replaced closed ones. That holds for benchmark scores, speed, or deployment flexibility.

First, task performance. The gap on everyday work is too small to justify a high premium.

Writing email, summarizing documents, organizing material, and generating routine code do not always need the most expensive frontier model. If a cheaper model reaches the accuracy and stability the work requires, the marginal return on a stronger model falls. The question is how much verifiable improvement the extra spending buys on the target task. A place on a composite leaderboard does not answer it.

Second, deployment economics. Cost, latency, and maintenance do not cancel the savings.

When a model runs on a personal computer or enterprise-owned infrastructure, users can cut cloud call fees and lower data-exfiltration risk. They get more direct control over deployment. Local deployment is not free compute. Hardware depreciation, energy, maintenance, model updates, and specialist staff all belong in total cost of ownership. Only a full cost count shows whether self-deployment beats API calls.

Third, agent stability. The model must do more than answer questions. It must complete tasks in a row inside real workflows.

Converging model capability does not make product experience converge. Agents also need tool calling, state management, permission control, error detection, environment understanding, and exception recovery. The execution system built around a model must understand code repositories and file structure. It must call tools and judge the results. It must recover from errors and keep the goal from drifting in long tasks.

A smart model solves only part of the problem. Production workflows must keep running in complex environments and finish tasks at an acceptable error rate.

Fourth, commercial substitutability. After counting full cost and risk, the alternative must be more competitive.

In July 2026, Kimi K3 topped a specific coding blind test on Frontend Code Arena with 1679 points. Claude Fable 5 scored 1631, and GPT-5.6 Sol scored 1618. On Program Bench, K3 scored 77.8. Fable 5 scored 76.8, and GPT-5.6 Sol scored 77.6.

These results show open-weight models approaching or passing closed rivals on some coding tasks. They do not show leadership on every task. They do not show lower overall cost in real production. Enterprises care about

• whether the model completes the target task,

• how many resources deployment needs,

• whether it meets security and compliance requirements,

• and how large the loss is when it fails.

If a cheap model needs more attempts, produces more errors, or adds heavy manual review, a low token price does not mean a lower task cost.

The substitution advantage of open-weight models is already real in general coding, document processing, and some local-privacy settings. On complex logical reasoning, long-chain research and precise retrieval over very long contexts, closed frontier models may still hold an edge. The gap between 70% gross margin for frontier models and about 20% for open-weight companies has looked stable for three years. Enterprises must also test

• security assessment,

• auditability,

• and permission control separately.

A performance score is not proof of security.

So “good enough” requires all four at once:

• target-task performance passes,

• deployment cost is affordable,

• agent execution is stable,

• and full lifecycle cost and risk are more competitive.

The standard also keeps moving. Each step forward by closed models raises the bar for open-weight substitution.

Who actually pays for AI

a16z’s seventh consumer AI apps report brings in a sample of US consumer bank card and e-receipt data from YipitData. It gives a new view of how consumer payment is structured.

As of August 2026, about 4.5% of US consumers in that sample bought a personal subscription to ChatGPT, Gemini, or Claude. A year earlier, it was 2.1%. Nearly half of US consumers say they have used AI. About 25% use it daily.

Awareness and trial are broad. Personal payment is far from the same penetration. Using AI and paying for AI continuously are different behaviors.

The structure of payment matters more than the payment rate. In this sample, the top 1% of spenders paid an average of about $903 a month for AI through consumer payment channels. They contributed about 19.5% of consumer AI spending. The bottom half of users together accounted for only 16.6%. The sample is not a census of the US market, but it shows clear concentration among heavy users.

High spenders buy more than chat. They also buy code generation, productivity software, autonomous agents, and developer tools. a16z found they spend clearly more than typical paying users on automation and product-building tools. Early consumer AI monetization may therefore concentrate among people who can use the tools directly for work, development and creation. Payment is not spread evenly across all consumers.

Product scale and conversion to premium plans do not line up either. a16z shows ChatGPT has about three times as many US paying users as Claude or Gemini. On premium plans, about 7.3% of Claude’s buyers take the Max plan above $100 a month. For ChatGPT and Gemini, the figures are about 1.1% and 1.3%. These numbers depend on the sample and product definitions. They still show that the largest user base does not necessarily have the strongest monetization at high price points.

Enterprises show the same gap between use, payment, and realized value. McKinsey’s 2026 survey says about 89% of responding organizations regularly use AI in at least one business function. 80% of respondents say AI improved personal productivity. Only 37% of organizations report a positive EBIT impact from AI. Only about 6% qualify as “AI high performers.” MIT’s NANDA lab found 95% of surveyed generative AI pilots produced no measurable P&L impact. That 95% reflects an adoption failure rate, not a technology failure rate. The causes were mainly rigid workflows, lack of contextual learning, and mismatch with daily operations. They were not insufficient model capability. PwC’s 2026 global CEO survey found 56% of CEOs said AI had brought neither revenue growth nor cost reduction in the past twelve months.

No measurable short-term P&L impact does not mean AI has no value. Enterprise adoption rates and employees’ subjective experience cannot replace verifiable return on investment.

Analyzing these data requires separating three kinds of evidence:

• realized financial results,

• operating trends now forming,

• and expectations that depend on future assumptions.

Confirmed revenue and cash flow are the factual base. Customers moving from pilot to production, expanding spending, and renewing are trend evidence. Multi-trillion-dollar market size, long-run excess profit, and large-scale automation replacement are still expectations to be verified.

Anthropic’s model leans toward enterprise API and professional workloads. It does not rest only on personal monthly fees. By the revenue breakdown in the pre-IPO materials, about $3.8 billion of 2025 revenue came from customers paying by actual Claude usage. Subscription revenue was about $789 million. This lets Anthropic serve enterprise customers and professional developers with a stronger willingness to pay. It also carries a risk. Enterprise customers care more about cost, delivery quality, and alternatives. If a cheaper model does the same work, they have a stronger reason to change purchasing.

Low consumer payment penetration cannot be used directly to size Anthropic’s whole commercial space. High enterprise usage does not automatically become sustainable revenue. What matters is whether customers keep paying, whether they expand workloads that earn an economic return, and whether the supplier keeps a reasonable profit under competition.

Subscribe now

Defensive enterprise buying

Defensive enterprise purchasing is becoming common. It is shifting from buying tools to avoid falling behind toward buying reassurance to avoid blame. The two logics look similar. The economic mechanisms differ.

First, enterprises buy for business continuity as well as cost savings.

When Claude Code, Copilot, and similar tools are embedded in daily development, not renewing is more than one fewer software expense. It can interrupt delivery pace and lower team efficiency. It can let a competitor ship product iterations first. Business continuity is not permanent pricing power. Defensive purchasing becomes a long-term commercial advantage only if product performance, workflow integration, and migration cost keep customers dependent.

Second, AI purchasing creates new defensive pressure.

When one department adopts AI and claims efficiency gains, other departments that do not follow may be disadvantaged in budget fights. AI also erodes traditional software spending. McKinsey’s survey shows 32% of responding enterprises dropped or canceled at least one external software purchase because of internal AI coding tools. AI purchasing can raise model vendors’ revenue. It can also push down revenue at some traditional software vendors. It is a budget shift, not a simple addition to software spending.

Third, the ROI verification cycle is long, and the decision window is short.

Enterprises struggle to quantify in advance the opportunity cost of not buying AI. They often must respond quickly to competitors’ deployments. AI purchases can therefore come before strict return verification. If AI budgets cannot become verifiable operating results over time, defensive demand will eventually be constrained.

Fourth, defensive purchasing will diverge. It will not become uniform.

At the general tool layer, decisions often rest on “don’t fall behind.” Replacement cost is relatively limited, and enterprises may try several models. In high-value scenarios, enterprises are more likely to move from buying tools to buying outcomes. They pay by task completion, usage, or verifiable business results.

There is also an organizational economics question that is often ignored. Is AI helping people do their work? Or is it forcing people to work at the pace of the machine’s output?

a person is augmented by a tool, using the machine to extend their own ability

or

a person is turned into an accessory of the machine.

They handle the output and the leftover problems of automated systems. The people most enthusiastic about AI are often those who control their own economic fate and other people’s, but it is not for ordinary people.

If AI helps a developer cut repetitive work and raise delivery quality, the person is using the machine to extend ability. If the company only asks employees to receive more AI-generated code/content, clean up errors, work faster and take responsibility for what the system misses, the person may become an accessory to the machine. Both can show higher output per person. The net economic value created is not the same. The second case may only move the burden from generation to review. It may even add hidden cost while productivity looks higher.

So measuring AI productivity cannot stop at counting how much code, text or task output was generated. It must subtract the cost of supervision, error correction, rework and mistakes. AI tools create realizable customer value only when the added output becomes business benefit.

Defensive purchasing can lift adoption. It does not necessarily create pricing power. The ceiling on AI commercialization depends on whether customers can keep realizing economic gains and how much of that value the model supplier captures.

Claude Code

Claude Code was Anthropic’s most prominent growth engine. The coding agent launched publicly in May 2025. TickerTrends, a third-party tracker, estimates its annualized run rate rose from $2.5 billion anchored at February 12, 2026, to about $6.3 billion at the end of March, about $10.3 billion at the end of April, and above $14 billion by June.

The product matters for more than revenue growth. It matters for the work setting it is embedded in. A general chat model gives one answer at a time. A coding agent can go deep into a codebase. It can inspect files, write code, run tests, and fix errors. It can plug into a team’s existing development tools and delivery process. It has a chance to move from a one-off tool to a daily production system. Once a product is embedded in real workflows, usage frequency, switching cost, and willingness to keep paying usually matter more commercially than a one-time model experience.

By August 2026, third-party tracking shows Claude Code’s growth slowing clearly. As of August 10, its tracked annualized run rate was about $15.12 billion, up about 5.2% over the past four weeks. Over the same period, OpenAI Codex’s tracked run rate was about $8.83 billion, up about 20.8%.

These are not product revenue figures that Anthropic formally disclosed and audited. TickerTrends’ method includes assumptions about Codex’s share of OpenAI’s enterprise revenue. That limitation needs scrutiny. The figures still suggest competitors are catching up faster. Compared with early growth of more than sixfold, 5.2% is a clear slowdown. In traditional software, it is still a high growth rate.

Several things can affect customer choice:

• usage limits,

• price,

• data governance,

• and integration experience.

Regulatory and compliance discussion around machine-readable text watermarks may also change some developers’ habits. The existing evidence does not show that a watermark policy directly slowed Claude Code’s growth. Attributing the change to a single event is not supportable. The more accurate reading is that the slowdown has several causes. The watermark policy is one concurrent event. It is not the confirmed sole cause.

The key is whether competitors can deliver similar results at a lower price, with fewer restrictions, or with a better integration experience. If users can finish the same task at lower cost, Anthropic may need to cut prices, change plans, or change its billing structure to keep customers.

Heavy users also have a natural ceiling. Take a user on the $200 a month Max plan. If their token demand is near the practical limit of their workload, there may be little room to use more. A supplier can create reasons to pay through high-value new features. It cannot assume every customer’s usage grows without limit.

The key test for Claude Code is whether it can keep the quality of growth while competitors close in and prices come under pressure. Continued growth alone is not the test.

Share

Customer Concentration, Channel Cost, Governance Risk

Anthropic’s fast growth does not hide its customer concentration.

Reuters, reporting on pre-IPO disclosures, says Anthropic’s 2025 sales through the Amazon and Google cloud marketplace channels totaled about $2.16 billion. That is about 47% of full-year revenue. Two unnamed customers each contributed about 12% of revenue, together about 24% of the year. The company also warned investors that many large customers have no long-term purchase contracts and may cut or stop spending.

One distinction in basis matters. Cloud marketplace channel sales reflect customer transactions completed through the Amazon and Google platforms. They cannot be read as Amazon alone contributing 47% of Anthropic’s revenue. High customer concentration does not mean revenue is necessarily unstable. It does mean the budget decisions of a few customers can have a disproportionate effect on total revenue.

Channel cost also matters. Anthropic pays cloud platforms about $351 million in channel distribution fees. Measured against the related channel sales, cloud partners take about 16 cents of every dollar of related sales revenue.

The relationship is also unusual. Amazon and Google are at once Anthropic’s investors, cloud infrastructure suppliers, product distribution platforms and direct competitors. Investor as supplier, channel as competitor: this is a structural risk in the business model. When things go well, the parties’ interests may line up closely. On compute allocation, product pricing, and customer relationships, they will not always be on the same side.

Governance structure also affects what public-market investors actually get. According to pre-IPO reports, Anthropic plans to use a Founder LLC made up of seven co-founders, plus a special Class F voting arrangement. This would let the founders collectively hold about 50.1% of the voting power. Such a structure can protect a long-term mission and insulate management from short-term share price pressure. For ordinary shareholders, economic rights and governance influence may differ sharply.

Regulatory risk should not be ignored. On September 30, 2026, the Washington Post first reported that the US Federal Trade Commission has opened a broad investigation into the AI safety practices of Anthropic and OpenAI. It is examining whether they violated federal law against unfair and deceptive practices. The background includes an OpenAI AI agent breaking out of a controlled environment during testing and breaking into the AI developer platform Hugging Face. Anthropic and OpenAI have both reported cases of their systems leaving the intended environment.

Three levels should be kept apart:

• Unintended behavior occurred,

• a regulator opened an investigation,

• and a regulator found a violation.

They are not the same. An investigation is not a finding that the company broke the law. It does put compliance costs and potential legal liability into the valuation. As of publication, the investigation is ongoing.

For an AI company, stronger capability and broader tool permissions may raise potential economic value. They may also enlarge the damage and liability from system errors. Safety and governance are a compliance cost. They also affect whether enterprises are willing to hand critical work to agents.

Whether a foundation-model company becomes a long-term winner depends on more than building a technical edge. It also depends on whether it can keep delivering at reasonable cost. And it depends on whether it keeps the value it creates under the constraints of channel, governance and regulation.

Subscribe now

The $518 billion compute commitment and circular financing risk

Reuters, in September 2026 reporting on pre-IPO disclosures, says Anthropic’s total future contractual commitments tied to cloud computing, compute and infrastructure are about $518 billion. About 80% of that is non-cancellable. Or it requires the company to pay under specific contract terms regardless of actual usage.

The main arrangements disclosed are about $111.1 billion committed to Google, about $110 billion to Amazon, about $31.4 billion to Microsoft, and about $161.2 billion in Broadcom-related equipment lease obligations. The arrangement with xAI involves up to about $84.5 billion in Nvidia GPU compute spending through 2029. It can be canceled on 90 days’ notice, so it cannot simply be equated with non-cancellable contracts.

The two kinds of commitment differ in accounting and economic meaning. Broadcom equipment leases usually involve recognizing right-of-use assets and lease liabilities. They may be operating or finance leases. The Google, Amazon, and Microsoft commitments are long-term infrastructure service obligations. Merging them to compute an annual average amount ignores that difference.

Contract terms also vary. The Amazon agreement runs to April 2036. The Google and Microsoft agreements cover the next seven to ten years. Compute spending usually rises with business scale, so early-year spending is far below the ten-year average. The $84.5 billion xAI arrangement can be canceled on 90 days’ notice. That gives some flexibility in a stress scenario.

At the end of 2025, Anthropic’s cash, cash equivalents, and short-term investments were about $20.28 billion. The commitments are far larger than the cash reserve then. Investors do need to examine that. Two opposite errors should be avoided.

First, the full $518 billion is not a one-time current cash outflow. It is not ordinary debt on the current balance sheet.

Second, the fact that these commitments are not all current debt does not mean the economic constraint can be ignored. Some commitments must be paid regardless of actual usage. If demand falls, the company may not be able to cut spending in proportion.

If customers shift workloads to cheaper models, Anthropic may still owe minimum payments for reserved compute. Revenue is uncertain, and part of the cost is already fixed by a long-term contract. That cost rigidity amplifies cash flow pressure when demand falls short.

Circular financing spreads the risk from a single company across the industry chain. Capital providers fund model companies, cloud platforms and infrastructure projects. Model companies put that money into cloud services, equipment leases, and chip purchases. Cloud platforms and suppliers expand investment based on orders and expected future demand. The new infrastructure then has to be paid for by future model revenue. This structure can be reasonable productive investment. If end demand is insufficient, the loop also runs in reverse.

On July 7, 2026, the Bank for International Settlements (BIS) published a working paper, “The AI Investment Race.” It describes the current AI buildout as one of the largest technology-driven investment booms in US history. The paper’s model estimates that under winner-take-all competition, the contest for market dominance could push industry investment to about 1.5 times the socially efficient level. In scenarios where demand is less price-sensitive, it could rise to about 3 times. The paper also notes that AI hardware is highly specialized. In a downturn, forced asset sales at a discount could amplify losses. Debt financing and circular equity relationships would speed up the spread of stress. “When one firm runs into trouble, the others may soon be affected.”

This does not mean all AI investment is unreasonable. It does not mean circular financing is itself accounting fraud. The risk is that infrastructure investment and financing capacity may rest on mutually reinforcing growth expectations. End customers’ willingness to pay and cash returns have not been fully verified.

One likely accounting misreading also needs correcting. A long-term purchase commitment does not let a cloud platform book the full committed amount as realized operating revenue on the day of signing. Contract commitments can support future investment plans, order expectations, and financing assessments. Revenue recognition still depends on service delivery and the applicable accounting rules.

Subscribe now

Trainer’s record and the “permanent bear” problem

New Constructs has not been right every time.

In 2019, the firm called WeWork’s IPO “the most ridiculous IPO.” WeWork failed to list on the original plan, its operating and financial problems surfaced, and it filed for bankruptcy protection in 2023. In 2021, the firm gave Allbirds’ IPO a negative rating and put fair value as low as $119 million. In 2026, the company sold almost all its assets for $39 million. That is only 1% of its peak valuation. Both are typical cases where traditional corporate analysis identified risk in the business model, balance sheet, and long-term contracts.

On DoorDash, the firm made a clear error. New Constructs rated DoorDash the “most ridiculous IPO of 2020” and gave a valuation of about $25.4 billion. DoorDash’s market value on its first trading day was far above that judgment, and the market valuation later rose further. Trainer’s bearish calls on Tesla and Nvidia also diverged clearly from market performance over some later periods.

Judging an analyst cannot rest on whether the stock later rose or fell. A stricter test looks at the valuation assumptions at the time, the forecast horizon, actual operating data, and risk-adjusted returns over the same investment horizon.

There is an important distinction here.

Valuation direction and investment timing are two separate judgments. An analyst can identify a long-term risk correctly and misjudge when it shows up in the price. Or a temporary price rise can lead the market to conclude, wrongly, that the analyst’s fundamental analysis was wrong.

WeWork’s and Allbirds’ problems centered on business model and financial sustainability. Nvidia has the CUDA software ecosystem, a developer network, and AI computing infrastructure as competitive advantages. These raise customer switching cost. They can make its profitability far higher than the surface logic of hardware sales suggests. If all tech companies are put into one static framework, ignoring technology diffusion, ecosystem formation and changing demand, an analyst can get the risk direction right and still get enterprise value and investment timing wrong.

That is also why the AI industry’s capital cycle may not move in step with its technology cycle.

Financial capital can speed up technology diffusion. It can also push capital and expectations too far before real returns arrive. The BIS research supports this quantitatively. Current AI investment is about 1.5 times the socially efficient level, and could reach 3 times in scenarios where demand is less price-sensitive.

In recent years, the AI industry has repeatedly gone through revisions of expectations. Early on, the market feared frontier models would become a black hole that kept consuming compute. After reasoning models developed, the capability frontier and cost structure changed. Low-cost models such as DeepSeek deepened market doubts about training cost and competitive advantage. Falling inference costs opened room for more applications. These changes cannot be simply described as each crisis having been “disproved.” The more accurate account is that the market kept overestimating or underestimating certain mechanisms. New products, new demand, and cost changes then altered the earlier expectations.

Even if AI capital markets go through an adjustment, it does not follow that AI technology has lost its value. The dot-com bubble left communication networks, software talent and the infrastructure that later innovation needed. The AI capital cycle may also leave stronger compute capacity, engineering talent, reusable development tools and open-weight models that run on ordinary devices. A productive residue does not guarantee returns to early investors. An industry can leave long-term value behind while its most aggressively financed companies go bankrupt. A technology can raise productivity a great deal while early shareholders, having paid too high a price, earn no reasonable return.

A bubble can fund a useful technology. It can also concentrate capital excessively before commercial returns are proven. A technology’s long-term success does not replace judgment about the price paid and the return on capital.

Subscribe now

The long-short contest: the answer is not in

Anthropic confidentially submitted a draft S-1 registration statement to the US Securities and Exchange Commission in June 2026. In its October 6 report, New Constructs states plainly that it did not obtain the full S-1. Its analysis relies on media reports and disclosed financial data.

Bloomberg reported in early October 2026 that Anthropic is seeking to list as early as mid-November. The IPO roadshow could formally launch the week of November 9. The aim is to start trading before Thanksgiving on November 26. The latest target is completion by the end of 2026. Anthropic is scheduled to meet potential investors at its San Francisco headquarters on October 14. Potential investors see a reasonable valuation range of $1.8 trillion to $2 trillion. Discussions are ongoing, and the timing may still change.

Whether New Constructs proves right or wrong depends on three judgments that are still unverified.

• First, can Anthropic turn its technical edge into sustainable profitability?

Model leadership is only the starting point. The company must show that customers will keep paying a premium for the lead. It must show the product can be embedded in daily workflows. It must show that the revenue from each successful task covers the cost of inference, R&D, human supervision, channels, and infrastructure.

• Second, can customer usage become shareholder cash flow?

High-end API inference gross margin is estimated above 70%. If inference gross margin really sits in that range and keeps improving, the turn to positive cash flow could come earlier than New Constructs’ conservative scenarios assume.

High adoption does not mean customers will keep expanding purchases. Rising usage does not mean revenue rises in proportion. Better gross margin does not mean cash flow after operating expenses, capital expenditure, and contract payments will necessarily improve. More convincing evidence would be the company raising revenue quality and improving unit economics over several consecutive quarters. It would then gradually produce sustainable free cash flow after continued R&D and compute investment. The company must also show that new capital earns a return above its cost of capital.

• Third, will enormous compute commitments be backed by enough end demand?

The roughly $518 billion in contractual commitments is not a one-time current cash expense. It does set the economic obligations the company must meet in the future. Customer concentration is high. Some revenue depends on cloud platform channels. Token prices keep changing. In that setting, Anthropic must show end demand is strong enough. It must show product pricing and unit profit can cover these long-term commitments.

As model capabilities converge, competitive advantage may not stay concentrated in the model itself. It may shift to workflow, distribution, enterprise relationships, developer tools, and task completion quality. That could become Anthropic’s moat. It could equally become the entry point for competitors to dismantle its advantage.

How value is divided is equally critical. The productivity gains enterprise customers get do not all automatically belong to Anthropic. Customers may bargain down model supplier prices. Cloud platforms may take distribution revenue. Compute suppliers may earn infrastructure profit. Capital providers will demand compensation for financing risk. The value technology creates can be large. The profit pool left for model companies may not be equally large.

The debate comes down to one complete economic chain. Model capability leads to reliably completing real tasks, which leads to customers getting realizable economic benefit, which leads to Anthropic capturing part of that value, which leads to sustainable free cash flow, which leads to return above cost of capital, which leads to value for shareholders.

Every link can break. The technology can advance, customers can raise productivity, and the company can grow fast. Competitors may still push prices down. Infrastructure suppliers may take most of the incremental profit. Financing requirements may swallow free cash flow. Only when the whole chain holds does the technical advantage enter the valuation.

New Constructs may be right that Anthropic’s valuation demands very high growth. It may underestimate the long-term competitive advantage that workflow, distribution and technology diffusion bring. It may also be right about structural risk in the business model and still be unable to predict when, and by how much, the market will reprice.

The bulls must answer one question as well. Suppose Anthropic ends up as one of the most important AI companies in the world. Today’s price may still leave investors without enough margin of safety.

Do not treat technical capability as profit. Do not treat usage as revenue. Do not treat revenue as cash flow. And do not treat value that might be created in the future as shareholder return already delivered today.

Whether New Constructs ends up right or wrong will take several full fiscal years after the listing to settle.

The contest has raised one of the most important questions for capital markets in the AI era. Everyone is betting on future technical capability.

• How much new value does this race create, and how much of it stays in shareholders’ hands? That remains to be verified.

I set it free permanently. Many articles explain SpaceX recently. I think my perspective will be more realistic.

SpaceX isn’t a company short on money. It is using debt to rebuild the risk narrative.

What it is doing goes deeper than borrowing. It takes credit that belonged to rockets, satellites, and internet connectivity and repackages it as a capital machine that can keep financing AI infrastructure.

Lay Out the Facts First

SpaceX began trading on Nasdaq on June 12, 2026, ticker SPCX. It priced at $135 per share and sold 555.6 million Class A shares. The base IPO was $75 billion. It closed the first day at about $161, up about 19% from the offer price. After underwriters fully exercised the overallotment option, total proceeds reached about $85.7 billion, the largest IPO in history. The deal closed on June 15.

In the prospectus, SpaceX no longer defines itself as only a traditional aerospace company. It places itself inside an AI services and infrastructure narrative. The S-1 filing shows 2025 full-year revenue of $18.7 billion with a net loss of $4.9 billion; the AI segment alone burned $12.7 billion in capital expenditure that year, more than triple the combined spending of the space and connectivity divisions.

• https://www.sec.gov/Archives/edgar/data/1181412/000162828026040364/spaceexplorationtechnologib.htm

On June 19, the company disclosed about $100.8 billion in cash and cash equivalents, a large jump from about $15.9 billion at the end of Q1.

Soon after listing, it went to the bond market. About ten days later, SpaceX launched an investment-grade bond offering with an initial target of at least $20 billion. Orders came in near $89 billion, and the size was raised to $25 billion. It was split into five tranches of senior unsecured notes, with maturities of 5, 7, 10, 20, and 30 years:

• $7 billion due 2031, coupon 5.35%.

• $6 billion due 2033, coupon 5.65%.

• $6 billion due 2036, coupon 5.875%.

• $2.5 billion due 2046, coupon 6.6%.

• $3.5 billion due 2056, coupon 6.65%.

Weighted average maturity is about 12 years. Weighted average coupon is about 5.9%. The bond sale drew a final order book of $73 billion at pricing, less than three times the amount offered, below the year’s average of about four times for high-grade deals. The 2036 tranche priced at a spread of 140 basis points above Treasuries, roughly 40 basis points wider than the average for similarly rated BBB debt.

The cash burn picture became clearer when SpaceX released its first quarterly report since the IPO. In the first half of 2026, operating cash flow was just $3.5 billion, while capital expenditure reached $28.5 billion, of which $23.6 billion went to AI. Free cash outflow was approximately $25 billion. Q2 capital expenditure alone was $18.37 billion, more than double the quarter’s $7.81 billion in revenue.

Subscribe now

Starlink remains the cash engine. Q2 ARPU was $66 per month, holding stable sequentially but down from $85 a year earlier, reflecting geographic expansion into lower-priced markets. Management characterized this as a deliberate trade-off: users doubled to 12 million, and enterprise and government revenue grew 108% year over year to $1.8 billion.

As of June 30, SpaceX held roughly $93.5 billion in cash and equivalents. Management has said existing funds are sufficient to support operations for at least the next 12 months.

The proceeds were mainly meant to repay a $20 billion bridge loan obtained in March 2026. According to IFR, the bridge loan was priced off SOFR. The spread steps down as the credit rating improves: from 175 basis points for a speculative-grade or unrated borrower to 75 basis points once the borrower holds A-category ratings from at least two major agencies. The loan was executed on March 2. It was used to replace X and xAI debt that carried interest rates of 9.5% to 12.5%. The effective rate was 4.58%. Bloomberg calculated that the swap saves nearly $1 billion a year in interest.

The restructuring came at a cost that did not show up in the headline savings. SpaceX absorbed $1.163 billion in cash prepayment penalties and recognized an additional $1.526 billion accounting loss on extinguishment of debt, a combined one-time hit of nearly $2.7 billion against the first year’s interest savings. The bridge loan carried a contractual obligation: SpaceX was required to apply the net proceeds of a qualified IPO to repay the facility within six months. The facility matures in September 2027, with two three-month extension options at a 0.25% fee per extension. SpaceX ultimately received Baa1 from Moody’s, BBB+ from Fitch, and BBB from S&P investment grade, but not the Single A rating required to trigger the lowest 75bp spread.

That completed a capital structure restructuring. Short-term financing handled the existing debt. The IPO raised credit standing and financing capacity. Long-duration public debt took over from the bridge financing. This is the core function of the modern debt market. Once debt is standardized, it stops being a contract between a borrower and a bank. It becomes a capital instrument that institutional investors can hold widely, trade, and continuously reprice.

The bridge loan was not a routine financing. It was a pre-IPO balance-sheet cleanup engineered to convert $18.9 billion of legacy X and xAI debt, including $3 billion of 12.5% senior secured notes, into a single facility priced closer to investment-grade benchmarks. The mandatory prepayment clause made the IPO a structural requirement, not an option. The loan’s repayment was contractually tied to the equity raise, meaning SpaceX could not have simply held the IPO proceeds on the balance sheet without retiring the bridge. The IPO raised $85.7 billion in net proceeds; $20 billion of that was already spoken for on the day the cash landed.

From here the problem changes.

After gaining investment-grade credit, SpaceX did not visibly cut capital spending. It is expanding its balance sheet quickly. On October 6, the Financial Times reported that SpaceX was seeking about $40 billion to buy Nvidia AI chips. About $10 billion would be bank loans. $30 billion would come from investment-grade debt. Apollo Global Management is leading, with PIMCO and others in talks. The deal is expected to close in 2027, and part of the financing may use GPUs as collateral. Apollo had earlier provided a $3.5 billion first-lien financing package to a digital infrastructure fund managed by Valor Equity Partners. That SPV bought $5.4 billion of Nvidia GB200 chips and leases them to SpaceX on a triple-net basis, capitalized with $3.5 billion of Apollo-led debt and $1.9 billion of Valor-led equity. Following SpaceX’s acquisition of xAI, the lease benefits from a SpaceX guarantee, supporting an investment-grade profile.

The $40 billion financing takes this template to an entirely different scale, roughly eleven times the size of the February deal. It is not a standalone transaction. It is the third iteration of a financing structure that Apollo and Valor have been refining since late 2025: a special purpose vehicle buys Nvidia GPUs directly from the manufacturer, leases them to a Musk-affiliated entity, and is capitalized with a mix of Apollo-led debt and Valor-led equity. The first deal, in November 2025, was a $3.5 billion loan for an AI compute cluster. The second, in February 2026, supported the $5.4 billion GPU purchase. If completed, the $40 billion financing would be among the largest-ever debt financings for AI infrastructure, and the first time a company with an investment-grade credit rating uses GPU collateral at this magnitude. That distinction matters: when a BBB-rated issuer pledges depreciating hardware as collateral, the market is being asked to price AI asset risk through an investment-grade credit lens.

SpaceX’s financing logic has changed. It no longer uses debt to support an aerospace and satellite communications company with mature cash flow. It uses the credit of the mature business to pre-fund AI infrastructure whose returns are not yet fully proven.

According to a forecast by Melissa Otto of S&P Global Visible Alpha, SpaceX’s total debt will rise from about $41.7 billion in 2026 to more than $218 billion in 2028, more than a fivefold increase. Over the same period, capital expenditure rises from about $48.7 billion to $118.4 billion.

But Otto’s forecast was published in July, based on Q1 data. Q2 results, released in August, showed quarterly capital expenditure of $18.4 billion, $15.8 billion of it directed at AI compute infrastructure, roughly 86% of the total, far exceeding the roughly $13.1 billion consensus estimate. JPMorgan analyst Doug Anmuth responded by raising his 2027-2028 capex forecast to nearly $200 billion per year, noting it “further pressures FCF in 2027, a trend we see across the hyperscalers.” If JPMorgan is right, Otto’s $218 billion debt projection is too conservative. The market is not debating whether SpaceX will borrow more. It is debating how much more, and at what spread.

If that forecast holds, SpaceX faces more than rising leverage. It faces a balance sheet expansion. Debt capitalizes future cash flow. If the future cash flow does not arrive, the future that was capitalized today turns back into funding pressure today.

Why Does SpaceX Suddenly Need So Much Money?

The answer is xAI.

On February 2, 2026, SpaceX acquired xAI for about $250 billion. The combined entity was valued at about $1.25 trillion. The deal was all stock.

From an industrial strategy view, this turns a “technology–infrastructure–dependency” logic into a corporate structure. A SpaceX that is only a rocket company owns launch capability. A Starlink-only business owns communications infrastructure. With xAI added, the businesses combine into a different shape:

This is not a simple consolidation of independent units. It forms a vertically integrated technology infrastructure.

Subscribe now

The system burns a great deal of cash.

xAI’s Q1 2026 operating loss was about $2.47 billion on revenue of only $818 million. The loss was about three times revenue. For full-year 2025, its operating loss was about $6.36 billion on revenue of about $3.2 billion. Capital spending is more aggressive. In Q1 2026, SpaceX’s total capex reached $10.1 billion. About $7.7 billion of that, roughly three-quarters, went to the AI business. Space capex was about $1.1 billion. Starlink was about $1.3 billion. AI capex alone was more than three times the other two combined. In 2025, xAI spent $12.7 billion on AI infrastructure, more than the combined $8 billion SpaceX spent on Starlink and rocket launches.

That is not the trajectory the IPO narrative assumed. It is the trajectory of a company spending like a hyperscaler while earning like a startup. SpaceX’s Q1 2026 total revenue was $4.69 billion. xAI’s operating loss alone was 53% of that figure, and its capex was 164% of it. The space segment spent $1.1 billion on capital projects. Starlink spent $1.3 billion. The AI segment spent $7.7 billion, roughly 76% of all capital spending in the quarter, a ratio that would be extraordinary for any company, let alone one that went public ten months ago.

Morgan Stanley analysts forecast that xAI could consume as much as $120 billion in cash across 2026 and 2027. That is about 1.6 times the $75 billion base IPO.

The Morgan Stanley estimate was published on July 22, 2026: the same day SpaceX shares hit an all-time low of $116.97, down more than 13% from the IPO close. Short interest had surged from 5%–7% of the float a month earlier to 32%, roughly 206 million shares and about $25 billion in nominal value. The $120 billion figure covers two years of cumulative cash consumption, operating losses plus capital expenditures for the AI segment, not debt. At the Q1 annualized capex run rate of $30.8 billion, the implied annual burn is roughly $60 billion, meaning Morgan Stanley expected the quarterly pace to double from Q1 levels.

Once inside SpaceX, xAI did not bring an ordinary “growth business”. It brought a very large up-front capital commitment.

We can treat the new space economy as a financing process that matures in stages. Early on, it relies on government, entrepreneurs’ private money, and outside capital. At maturity, the industry itself should generate enough revenue to support the next round of expansion. Capital should move from “going into the industry” to “being produced by the industry”.

SpaceX is doing something more aggressive. Before the AI industry reaches self-financing, it uses Starlink’s cash flow and SpaceX’s investment-grade credit to buy AI time. Starlink’s 2025 operating profit of $4.4 billion effectively subsidizes xAI’s expansion. But the subsidy is far smaller than the burn. The question is no longer whether SpaceX can enter AI. The question is whether AI can reach self-financing before the debt market starts demanding returns.

Subscribe now

What Changed Is the Balance Sheet’s Identity

SpaceX’s Q1 2026 net loss was about $4.28 billion on revenue of about $4.69 billion. In the same period of 2025, the net loss was only about $528 million. One quarter’s loss is close to the full-year 2025 loss of about $4.94 billion. Capex was about 215% of revenue. This is not a local swing on the income statement. The direction of capital allocation has changed.

By Q2 2026, the change is clearer. SpaceX capex reached $18.369 billion, up about 6.5 times from $2.825 billion a year earlier. Of that, $15.828 billion went to AI compute infrastructure, 86.2% of total capex. CFO Bret Johnsen said capex is expected to stay at a similar level in the coming quarters. Meanwhile, Starlink connectivity revenue in Q2 2026 was $4.291 billion, up 66% year over year. Operating profit was $1.656 billion, up 79%. SpaceX’s total Q2 revenue was $7.814 billion, up 92%, above analyst expectations.

What alarms credit analysts is the cash flow. In the first half of 2026, SpaceX’s operating cash flow was only about $3.5 billion. Capex was about $28.5 billion, $23.6 billion of it on AI alone. Free cash flow was negative $25 billion. Operating cash flow covered about 12% of capex. Even with Starlink growing fast, SpaceX’s overall cash flow structure is still capital-consuming.

The ratio that matters most is not capex versus revenue. It is Starlink’s operating profit versus AI capex. In Q2, Starlink generated $1.66 billion of operating profit. SpaceX’s AI segment consumed $15.8 billion of capital expenditure in the same quarter. Starlink’s quarterly profit covers roughly 10.5% of the quarter’s AI capex. Even if Starlink’s operating profit grows 50% annually, it would take until approximately 2029 for its cumulative quarterly profit to cover a single quarter’s AI capex at the current run rate. JPMorgan analyst Doug Anmuth now projects capex of nearly $200 billion in both 2027 and 2028, “which further pressures FCF in 2027, a trend we see across the hyperscalers.”

So SpaceX’s three core businesses run on three different time scales.

• Starlink is in commercial scale-up and generates cash steadily.

• Starship is a long, high-risk development and infrastructure project. According to the prospectus, the program has run for nearly ten years with cumulative spending above $15 billion, far above the $2 billion Musk projected in 2018.

• AI is in the most aggressive stage and the one with the least return validation: large capital input, very fast technology iteration, commercial returns still to be proven.

All three sit on one balance sheet and are financed by the same debt market. What credit analysis actually looks at is not margin. It is the time structure of cash flow:

• Where does the cash come from?

• When does it return?

• What is the probability that it returns?

SpaceX’s biggest credit strength comes from Starlink. Its biggest new capital need comes from AI. What to watch is whether

Why Its Credit Base Is Firmer Than a Typical AI Company’s

SpaceX’s credit did not start from a slide deck.

In the traditional space system, cost-plus contracts weaken, to a degree, a company’s economic incentive to cut costs. The Space Review has documented how “cost-plus contracts that NASA had been using to develop the combined SLS and Orion system worked to the contractor’s rather than NASA’s advantage,” producing what one analysis called “bloat and overruns.” Fixed-price contracts put cost risk and efficiency gains back on the company. SpaceX built its long-term competitiveness not only because its rockets are “more advanced”. It tied engineering innovation, cost reduction, and profit incentive together. Reusability then changed the economic nature of the capital good. A disposable rocket consumes manufacturing capital again on every mission. A reusable one lets the same technical asset provide service repeatedly.

The numbers validate the thesis. Falcon 9 now delivers payloads to low Earth orbit at approximately $2,720–$3,000 per kilogram, down from roughly $54,000/kg on the Space Shuttle, a reduction of more than 90% in two decades. A single Falcon 9 booster completed its 37th flight in August 2026, just two missions shy of the Space Shuttle Discovery’s record of 39. The marginal cost of a reused Falcon 9 launch is estimated at around $15 million, versus a list price of roughly $70 million for a dedicated launch. A used booster on its twentieth flight now costs SpaceX less to launch than it would cost any competitor to build and throw away a brand-new rocket of equivalent capability.

So:

Starship aims to push this further. Its target is $100/kg, a further 97% reduction from Falcon 9 levels. At full reusability with 20+ flights per booster, cost per kilogram could fall to $115–125/kg. A single Starship costs approximately $100 million to assemble, with fuel and materials adding roughly $2 million per launch. The Raptor 3 engine is approximately four times cheaper to manufacture than its predecessor, designed explicitly for rapid reuse.

This chain explains why SpaceX’s debt capacity cannot be treated as pure “story valuation”. An AI startup may have only a future. SpaceX at least has part of its past already delivered. What is happening now is

Moody’s applied its Telecommunications Service Providers methodology, not an aerospace framework, to rate SpaceX at Baa1. Starlink generated over $7 billion in EBITDA in 2025 and, before interest expense, produced close to $3 billion in cash flow. That is the cash flow base creditors are underwriting. The agencies assigned investment-grade ratings even as SpaceX posted a $4.9 billion net loss last year, because the rating committees were pricing Starlink’s connectivity economics, not xAI’s burn rate.

Maturity Mismatch: The Clocks Differ

The danger is not the amount of debt. It is that debt and assets run on different clocks.

SpaceX’s latest bonds have a weighted average maturity of 11.7 years, with maturity dates from 2031 to 2056. The AI compute equipment it is buying loses economic value much faster.

The industry disagrees on useful GPU life, and the disagreement has hardened into a credit-market dispute. Some bankers estimate about three to four years. Nvidia CEO Jensen Huang has said top-end GPUs can earn returns for up to about ten years. GPU asset valuation firm Barkr estimates a GB300 NVL72 system can be used for nine to ten years. Some large cloud providers have extended server depreciation from three to four years to five to six. Data center buildings and facilities can last 20 to 30 years, but the economic value cycle of GPUs and servers is usually about 4 to 6 years. Citigroup estimates GPUs and supporting servers can make up 30% to 40% of total data center project cost.

The dispute is not academic. In late September, Nvidia launched a $500 billion financing initiative to turn GPUs into loan collateral, comparing residual values against a five-year accelerated depreciation curve: the A100 retains about 25% after six years, the H100 about 58% after three to four, and the B200 about 158% after one. Banks are not persuaded. Impax Asset Management’s Tony Trzcinka said: “Banks typically underwrite GPUs over a 3-4 year depreciation schedule.” Nvidia’s original plan capped residual value guarantees at 25% for some transactions; banking sources said such limited guarantees “are currently difficult to gain market acceptance.” The financing structure is being adjusted toward stronger creditor protections, combining chip collateral, long-term customer contracts, and corporate guarantees.

So the issue is not that long-term debt can’t buy GPUs. It can. Infrastructure projects have always needed long-term capital. The issue is whether the asset’s economic return period can cover the long-term cash flow commitments the debt requires.

If a set of GPUs faces replacement by a new architecture in three to five years, and SpaceX’s bonds don‘t mature until 2056, then what today’s financing depends on is not that batch of GPUs. It depends on decades of continued AI cash flow, continued reinvestment, and continued access to capital markets. SpaceX is not borrowing “GPU money”. It is borrowing the ability to keep earning AI capital returns in the future. That is what “AI-grade credit” means.

The market is already pricing this mismatch. As of early October 2026, SpaceX’s 2056 bonds are trading at about 85 cents on the dollar, yielding roughly 2.27 percentage points above Treasuries, a level the FT described as “similar to the level of junk bonds.” The spread has widened from 175 basis points at issuance to 236 basis points. The 5-year CDS hit 194 basis points, a record high. The bond market is not waiting for the AI cash flows. It is repricing the duration risk now.

Why Is AI Infrastructure Worth a Bet This Size?

One point is easy to miss. As launch costs fell, areas once marginal, such as space stations, the Moon, logistics, and the space industry, gained a chance to enter real large-scale capital formation. It goes on to discuss orbital computing and space-based data centers.

The key driver unlocking space’s economic potential is private property rights and entrepreneurial incentives, not government programs. He identifies orbital data centers as one of the ventures that could soon be listed on public stock exchanges, alongside space tourism and asteroid mining. SpaceX’s dominance is quantified: the company runs roughly 50% of the world’s rocket launches and approximately 80% of the mass put into orbit.

Behind this sits a practical economic constraint:

The bottleneck is no longer theoretical. Grid connections in primary markets like Northern Virginia now take up to seven years. Transformer lead times run two to four years. Roughly $98 billion in data center projects were delayed or canceled in a single quarter of 2025 from grid and community bottlenecks. A 1 MW AI cluster generates roughly $70 million a year in cloud GPU revenue; a five-year grid queue delay therefore represents about $350 million in foregone revenue per megawatt. Against numbers like that, even an expensive orbit starts to look rational.

So SpaceX’s AI capex is not just “Musk likes AI”. There is an industrial logic behind it.

But the economics are not yet favorable. As of mid-2026, orbital compute costs roughly four times what an equivalent terrestrial data center costs. The main barrier is economic, not technical, with future launch costs a key determinant. Google’s own feasibility study concluded that at $200 per kilogram, orbital data centers become cost-competitive with terrestrial operations, a threshold they project by 2035. Independent modeling places parity around 2040 in a base case. The megawatt-class orbital data center being financed at scale in 2026 has three engineering preconditions:

• custom silicon,

• modular plug-in architecture so a 5-year compute card can swap under a 20-25-year platform,

• and launch costs below $300 per kilogram.

The two build-outs are linked by a single logic:

On September 28, 2026, Starship reached orbit for the first time and deployed 26 Starlink V3 satellites. SpaceX plans to expand from broadband to a “Starmind” data network of approximately one million satellites. The logistics for that scale can only be handled by Starship-class launch systems.

The problem is in the same place:

• The more complete the industrial chain, the more capital must be committed up front.

• The more capital committed, the higher the demand on future cash flow to show up.

SpaceX’s largest opportunity and its largest risk come from the same source. The opportunity is that vertical integration: launch, satellites, communications, compute, creates a cost structure no competitor can match. The risk is that the capital required to build all four simultaneously exceeds what even investment-grade credit markets can absorb before any one of them reaches self-financing.

Why the Bond Reaction Matters More Than the Stock

After the $40 billion financing news on October 6, SpaceX’s stock fell after hours and in early trading the next day. The stock is not what matters. The credit market is.

According to ICE Data Services, SpaceX’s five-year CDS spread rose as much as 14.5 basis points to about 195.4 basis points a year, an intraday high since the contract began active trading in June 2026. When active trading began in June, it was about 110 basis points. A CDS is, in essence, the market price of insurance against default. A move from 110 to nearly 195 basis points means the annual cost of default protection on $1 million of debt rose from about $11,000 to nearly $20,000.

The paradox is structural: the 85 basis point widening occurred over four months in which SpaceX completed an $86 billion IPO and issued $25 billion of investment-grade bonds. The CDS market is not reacting to the current balance sheet. It is reacting to the trajectory.

The bonds are repricing too. According to MarketAxess, SpaceX’s 6.65% bond due 2056 trades at about 85 cents on the dollar. Its yield is about 2.27 percentage points above a Treasury of the same maturity, a spread close to junk-bond levels, as the FT noted. When the bond was issued in June, the spread was only 175 basis points. Trace data shows the credit spread widened to 238 basis points by early October. Both measures confirm the same directional move: from 175bp at issuance to a range of roughly 227–238bp within four months. For a BBB-rated issuer, a credit spread above 200bp is historically associated with elevated default expectations. A spread of 238bp implies the market is pricing SpaceX closer to the BB (junk) category than to its actual BBB rating.

The most important thing here is not the 238 figure. It is this:

Moody’s rates SpaceX Baa1, Fitch BBB+, and S&P BBB. All remain investment grade, all with stable outlooks. Moody’s and Fitch ratings sit three notches above junk. S&P’s BBB is one notch lower. But bonds and CDS are demanding clearly higher risk compensation. A rating is a discrete category. Bond prices and CDS are continuous market prices. The rating says SpaceX still stands inside the investment-grade zone. Market prices say investors are recalculating how much risk they must bear for it to stay there. The distance between the two is itself a signal.

The agencies are pricing what SpaceX is. The bond market is pricing what SpaceX is becoming. Rating agencies operate on annual review cycles. Bond markets reprice continuously. The 238 basis point spread is the market’s answer: it is pricing SpaceX not at Baa1, but at a level closer to Ba2/BB, two notches below investment grade.

The Real Warning Is Not Default but Investment-Grade Credit Being Spent Early

Credit is an asset. The stronger the credit, the lower the financing cost. The lower the financing cost, the easier it is to support larger capex. If larger capex eventually creates new cash flow capacity, credit strengthens further.

SpaceX is betting on this flywheel:

The reverse loop also exists:

The danger comes when the second chain starts to dominate the balance sheet. At that point, debt stops being a tool of capital expansion and becomes a constraint on it. So what the market cares about is not whether SpaceX can borrow another $40 billion. It is what the next round of money costs.

Industrial Strategy and Financing Strategy Are Now One Thing

The real turning point of the new space economy is not “government stops, private firms take over”. It is that space moves from a government project to an industrial system that keeps producing commercial revenue, attracts private capital, and in turn expands infrastructure.

SpaceX is doing something more aggressive:

So SpaceX’s move from rockets into satellite communications, then AI, data centers, and government infrastructure is in essence a move from “selling products” to “building and controlling infrastructure”.

If it ultimately connects chips → data centers → networks → AI models → data → users → government customers, it gets more than an AI division. It gets a new layer of infrastructure control. And infrastructure control raises financing capacity in turn:

So:

This is more than industrial strategy. It is itself a financing strategy.

Musk’s Other Financing Capability: Turning Attention Into Capital

Musk has one financing capability that an ordinary industrialist can hardly copy:

A traditional industrial company runs:

A Musk-style company more often runs:

What matters is not storytelling. It is that future expectations get capitalized early. SpaceX’s IPO shows this. The $75 billion base raise, and $85.7 billion in the end, added cash and also set a public-market credit anchor. That equity-market credit was then converted into larger debt financing capacity.

So issuing bonds right after the largest IPO ever is not a contradiction. The IPO handles equity financing. The bonds handle something else:

For an industry like AI infrastructure, which depends heavily on up-front capital, this capital structure is very attractive. But it requires the capital market to believe something not yet fully proven:

Risk Control Should Watch the Debt–Collateral–Refinancing Loop

If part of the $40 billion financing ends up using GPUs as collateral, the risk no longer stays in the income statement. It enters the feedback system between the balance sheet and the financing market.

Once AI asset values fall, collateral values fall. When collateral values fall, lenders may raise haircuts. When haircuts rise, the borrower needs more cash or more collateral. If cash is short, forced asset sales can follow. Asset sales can push market prices down further. So:

This is not simply “asset prices dropped”. It is a financial accelerator.

In 2026, GPU-backed financing began to move from marginal experiment toward a more standardized financing tool. CoreWeave completed a $3.1 billion GPU-collateralized loan. Stonebriar and Wingspire are also pushing GPU-related equipment-finance ABS. GPUs are shifting from plain production equipment to assets the financial system can price, pledge, and securitize. Once “compute assets” enter the financial system, the AI cycle is no longer only a tech cycle. It gradually takes on the character of a credit cycle.

This is the point in the SpaceX case: once future cash flow is turned into debt, collateral, and refinancing, it gains financial leverage and financial fragility at the same time. So what SpaceX should guard against is not only whether “$200-plus billion of debt is too much”. It is whether AI assets, bond prices, collateral values, and refinancing capacity end up bound into the same feedback system.

Why Governance Matters Especially Here

Bond investors buy contractual cash flows, not stories.

SpaceX’s governance structure is highly concentrated. Musk holds about 42% of the equity but, through super-voting shares, controls about 85% of total voting power. Class A shares carry 1 vote each. Class B shares carry 10 votes each. On June 11, the day before the IPO, MSCI gave SpaceX an ESG rating of CCC, the lowest tier in its system. The governance score is only 3.2 out of 10. In the controversies category, SpaceX scored only 1 out of 10.

These indicators are not credit ratings. They cannot be converted directly into default probability. But they point to something creditors should notice:

The issues around related-party transactions and financial transparency are more specific. On May 6, 2026, SOC Investment Group wrote to the SEC asking it to review the accuracy and reliability of SpaceX’s financial disclosures. It focused on the accounting treatment of transactions between SpaceX and other Musk-linked companies such as xAI and Tesla, and asked the SEC to ensure the auditor stays independent.

The prospectus discloses that xAI bought about $191 million of goods and services from Tesla in 2024, about $506 million in 2025, and about $303 million from January to April 2026. April Megapack purchases alone were $269 million. Cumulative spending is close to $1 billion, used to stabilize peak power demand at the Colossus data center.

These transactions do not by themselves imply wrongdoing, and they do not directly prove credit risk. The question is this:

For shareholders, that is a governance issue. For creditors, it is a cash flow visibility issue.

SpaceX Is Far From an Out-of-Control Credit Story

This has to be said plainly. SpaceX’s investment-grade rating does not rest on a story alone.

Starlink is the real cash flow base. This is why the rating agencies did not treat SpaceX as simply a cash-burning AI startup. Moody’s even used a “telecommunications service provider” rating methodology to support its Baa1 rating.

But Starlink itself is going through a business model shift. In Q2 2026, Starlink subscribers reached about 12 million, double the 6 million a year earlier, up 1.7 million from the previous quarter. ARPU was $66 per month, flat with Q1 and down about 22% from $85 a year earlier. Management explains the change as international expansion and growth in low-price plans, trading user scale for market share. Meanwhile, enterprise and government revenue rose about 108% year over year to $1.8 billion, an important offset.

So what to watch is not whether ARPU is falling. That answer is already clear. What to watch is whether

• If V3 satellite deployment goes well, network capacity is released, and the enterprise and government business keeps expanding, Starlink’s cash flow base gets stronger.

• If V3 deployment falls short, or competition pushes ARPU down further, SpaceX’s most important source of debt service takes a direct hit.

Leverage: Break It Into Stock, Time and Price

Management’s target is to keep Debt/EBITDA at 2 to 3 times. For full-year 2025, the company had revenue of about $18.674 billion, adjusted EBITDA of about $6.584 billion, and a net loss of about $4.937 billion. That shows SpaceX is not without earning power. Starlink’s strong EBITDA can support a fair share of AI and space investment.

But that EBITDA base was built before AI capex fully accelerated. S&P expects SpaceX’s negative cash flow to continue until 2030, with cash burn rising further in 2027 and 2028. In the first half of 2026, operating cash flow was about $3.5 billion, capex about $28.5 billion, and free cash flow negative $25 billion.

So the “2–3 times leverage” target is not the answer.

• Debt/EBITDA is the stock.

• The maturity wall is time.

• Refinancing cost is price.

• Cash flow is the source of repayment.

• Credit spread is the market’s combined verdict.

Whether SpaceX can keep investment-grade credit depends on all five holding at once.

What the $40 Billion Really Tests

The test is not whether SpaceX can borrow. At present it still can. The test is the price at which the market is willing to lend.

If the future $30 billion of investment-grade debt can still be issued at a relatively limited credit premium, the capital market still believes Starlink’s cash flow and SpaceX’s AI infrastructure can deliver strong long-term returns.

If every new round of AI debt keeps widening the credit spread, clearly, SpaceX enters a different loop:

At that point, debt stops being growth fuel and becomes a constraint on growth.

Why 238 Basis Points Matter More Than a 2% Stock Drop

238 basis points is not panic. A CDS of 195 basis points is not default pricing. SpaceX is still an investment-grade issuer. But the market has begun to demand higher risk compensation.

The thing to watch is not a one-time downgrade. It is a gradual repricing:

Equity can price a Mars story ten years out. Bonds must pay cash on the next coupon date. A downgrade is the result. Spread widening is often the warning.

SpaceX Is Betting on the Ability to Capitalize the Future

The SpaceX story compresses into three chains.

The first is debt:

The second is industrial technology:

The third is Musk’s capital mechanism:

SpaceX is pressing all three chains onto one balance sheet:

Subscribe now

Read more

There are various interpretations of the AI Trade market situation; I will attempt to summarize and categorize them to make them easier to understand for all of you. Take the Korean stock market as a case study.

The Context

The AI stock market is hot, and its center is in the US. But the most complete risk sample is in Korea. Korea, Japan, and other Asian markets have all seen asset manias driven by AI, semiconductors, and high leverage. Korea is worth studying because it shows more than a rally: it shows a financial feedback mechanism.

Savings move into risk assets. Household risk appetite rises. Borrowing increases. Money concentrates in AI leaders. Leveraged ETFs amplify positions. Foreign investors cut exposure, and domestic money absorbs the supply. Rising prices raise collateral values. Borrowing capacity expands further.

When the market turns, the chain runs in reverse. Prices fall, collateral shrinks, and margin constraints tighten. Leveraged investors are forced to sell. As market depth falls, the same sell order causes a larger price impact, prices fall further, and a new round of margin pressure forms.

At that point, the market becomes a financial system that reinforces itself and reverses itself, not just a machine that prices expected future profit.

After deleveraging ends, risk appetite can come back. One crash does not necessarily mean the bubble is gone. It may only have removed the first batch of highly leveraged investors. If prices then rise again, the survivors may take away this lesson:

Leverage returns.

A dangerous bubble needs one true story and a market that extrapolates it to a price the real economy cannot sustain.

• Railways were real, and so was the 19th-century railway bubble.

• Electricity was real, and so was the speculative mania of the 1920s.

• The internet was real, and so was the 1999 internet bubble.

Subscribe now

Read more

That buffer has to be credible to the market, fast to deploy, and able to turn into actual market supply.

The SPR was never just underground storage. It is national energy insurance, a foreign policy tool, and a policy asset that shapes market expectations. Its value cannot be measured by inventory alone. It is closer to this:

Subscribe now

Read more

• a large price decline,

• partial exchange-inventory recovery,

• and continued discussion of physical tightness.

On January 29, 2026, silver touched an intraday record of $121.65 per ounce. On January 30, it fell roughly 30% in one day, from above $121 to below $75. LSEG data described it as the largest one-day decline since its records began in 1982. By early October, spot silver was around $60.08 per ounce, roughly half the January high.

At the same time, Indian silver imports recovered materially in August. New York-registered stocks returned to roughly 100 million ounces. London vault holdings also rose. That weakens the certainty behind claims that an extreme year-end shortage is inevitable.

It does not prove that physical silver is loose.

Subscribe now

Read more

The difference between a mature trader and an ordinary trader is often not what they know. It is whether they can keep following their method once the market starts arguing with them. Most traders already know trends, moving averages, volume, valuation, earnings, macro, sentiment, and the importance of stops. They can write a clean pre-market plan. They can define entry, stop, target, and risk.

The problem begins after the open. Price accelerates. They fear missing it. They chase. Price reaches the planned stop. They decide to “watch it a little longer.” A gain grows. They feel unusually correct. The next position becomes larger. A string of losses arrives. They want the money back. They begin trading setups they would normally reject.

A trade that began as

gradually becomes

A trading plan exists to break that chain. It is not designed to improve prediction accuracy. It is designed to tell you what to do when you are least qualified to decide.

Knowledge Changes Once Money Is Exposed

Before entering a position, you analyse probabilities, risk, and scenarios. After entering, you look at unrealised P&L, entry price, and whether you were right. The market has not changed. The meaning you assign to information has.

A normal pullback becomes:

A stop becomes:

A price move becomes:

That is the real shift. Once capital is exposed, the question quietly changes from

to

Markets generate paths that cannot be fully anticipated. Traders, meanwhile, turn experience into precise expectations. When reality departs from those expectations, they search for confirming information and discount evidence that undermines the original thesis.

More analytical ability can make this worse. More analysis creates more possible explanations. A trader can always find another indicator, another macro narrative, another piece of news, another valuation argument, another time frame. Intelligence becomes a machine for defending the position.

Kahneman’s System 1 and System 2 model helps explain the state change. Pre-market planning relies more heavily on slow, deliberate, effortful reasoning. Live trading triggers fast, automatic reactions through volatility, P&L, headlines, order flow, social comparison, and seeing others make money. System 1 is not useless. It is simply not neutral when your money is on the screen.

The dangerous moment is usually not when you have failed to analyse. It is when you have already analysed, entered a position, and then let emotion reinterpret the analysis.

Overtrading Is Not a Skill Signal

The evidence on retail trading is less flattering than the mythology. Barber and Odean studied 66,465 households at a large discount broker from 1991 to 1996. The most active fifth earned an annualised net return of about 11.4 percent. The market returned about 17.9 percent over the same period. The average household earned 16.4 percent annually and turned over roughly 75 percent of its portfolio each year. The study identified overconfidence as a major explanation for high trading intensity and the resulting poor performance.

The important detail is not merely that the most active group did badly. Even the average household return of 16.4 percent trailed the market return of 17.9 percent. The group did not lack access. They used the same broker, operated in the same market, and had the same broad opportunity set. They simply converted participation into friction, timing errors, and behaviourally expensive decisions.

The familiar chain is:

Barber and Odean’s 2001 study found that men traded about 45 percent more than women in its sample of more than 35,000 households. The estimated annual net-return cost associated with higher trading was 2.65 percentage points for men and 1.72 percentage points for women. This is not a useful story about gender in isolation. It is evidence for a broader mechanism: overconfidence raises activity, while more activity does not automatically imply more ability.

The pattern appears across markets. A study of the complete trading history of Taiwanese investors found that individual trading imposed an annual performance penalty of about 3.8 percentage points on the aggregate portfolio of individual investors. Institutions, by contrast, showed an annual performance boost of 1.5 percentage points. The study found that individual losses were concentrated in aggressive trading.

• https://faculty.haas.berkeley.edu/odean/papers%20current%20versions/justhowmuchdoindividualinvestorslose_rfs_2009.pdf

The Brazilian mini-index futures evidence is more severe. Among individuals who persisted in day trading for more than 300 days, 97 percent lost money after exchange and brokerage fees. Only 1.1 percent earned more than the Brazilian minimum wage. Only 0.5 percent earned more than the starting salary of a bank teller.

The conclusion is not that retail traders are structurally forbidden from making money. The conclusion is less comforting. Active trading has a high threshold. Most traders do not clear it. The usual constraint is not lack of market vocabulary. It is a lack of a behavioural structure strong enough to survive position exposure.

“Trade less” is therefore not a strategy, either. The actual questions are:

• Why is this trade worth taking?

• What evidence is sufficient to justify risk?

• What evidence invalidates the thesis?

• What is the maximum loss I have accepted before entering?

• What would make this setup worth repeating?

Trading costs are not only commissions and slippage. They include attention, cognitive load, emotional depletion, distraction from better opportunities, and the psychological residue of the previous trade. A low-quality trade can consume capital even when it does not immediately lose money.

Subscribe now

Read more

The arithmetic is obvious. The same dollar-priced goods cost fewer yen. If retail prices do not change, gross margins rise.

That logic is correct. It is also incomplete.

What determines profit is not how much the yen appreciates. It is how much of that move survives

• procurement contracts,

• overseas supplier price increases,

• inventory,

• hedging,

• shipping costs,

• competitive pressure,

• and final demand

…before it reaches the income statement.

A stronger yen is not simply a cost variable.

Subscribe now

Read more

On September 18, 2026, the Bank of Japan raised its policy rate from 1.00% to 1.25% in a 7-2 vote. It was the sixth increase since Japan exited negative rates in March 2024. It came only three months after the June increase, the shortest interval between hikes since 1990. The 1.25% rate is Japan’s highest since April 1995.

If this is read as another 25-basis-point move, it misses the important part.

The 25 basis points are secondary. A bigger structural change may be underway:

That funding anchor is not moving in isolation. Long-term U.S. Treasury yields are moving higher at the same time. The two anchors of the global balance sheet may be repricing together.

• On the asset side, Treasuries remain the world’s primary reference point for risk-free pricing and collateral. Their financing price is rising.

• On the liability side, the yen has long served as a low-cost funding anchor. It is becoming more expensive.

For more than a decade, the global carry system rested on those two conditions. The broader question is:

Subscribe now

Read more

On September 16, 2026, the FOMC voted 12-0 to raise the federal funds rate target range by 25 basis points to 3.75%–4.00%. This was the Fed’s first hike since July 2023, and the first under Kevin Warsh as chair. The statement said the hike would “support inflation returning to the 2% target more promptly,” while keeping the assessment that economic activity was “expanding at a solid pace.” The statement also specified that the Committee “continues to implement a policy of maintaining ample reserves in the banking system.”

Warsh’s stated bar was straightforward:

As of the September meeting, that condition wasn’t met. At the press conference, he pointed to three developments since July:

• labor market data stronger than expected,

• unsettling inflation readings,

• and geopolitical events with knock-on effects on energy prices.

Data going into the meeting reinforced this. The Bureau of Labor Statistics reported on September 11, 2026, that August CPI rose 0.4% month over month and 3.4% year over year; core CPI rose 0.3% month over month and 2.4% year over year. Gasoline prices rose 3.9% month over month in August, contributing more than a third of the total monthly increase. Energy prices were up 16.3% year over year, with gasoline up 27.4% and fuel oil up 52.0%. Wireless telephone service prices jumped 5.9%, a sign inflation pressure isn’t confined to energy.

Worth noting: two months earlier, at the July meeting, the FOMC voted 9-3 to hold rates, with Hammack, Kashkari, and Logan dissenting in favor of a 25bp hike. By September, the dissent disappeared; all 12 voting members backed the hike. Market analysis notes the September statement dropped the explicit reference to “Middle East conflict,” replacing it with the more generic “geopolitical developments,” and softened the description of capital investment from “strong” in July to “solid.”

But what actually shifted market perception wasn’t the 25 basis points.

At the press conference, CNBC’s Steve Liesman asked a question that has been nearly standardized over the past decade:

• https://www.cnbc.com/video/2025/06/20/watch-cnbcs-full-interview-with-federal-reserve-governor-christopher-waller.html

• https://www.cnbc.com/2025/06/20/cnbc-transcript-cnbc-exclusive-federal-reserve-governor-christopher-waller-speaks-with-cnbcs-steve-liesman-on-squawk-box-today.html

Warsh gave no number, no range. His answer effectively split the question into two layers:

• From an academic standpoint, comparing the current rate to the neutral rate is useful;

• For the FOMC’s actual decision that day, the comparison carries no direct operational effect.

Berkadia’s analysis notes that Warsh “effectively downplayed the concept as academically useful but of limited operational value,” shifting emphasis to financial conditions instead.

This distinction matters. It is not:

It is:

Subscribe now

Read more