The AI Bubble

In November 2025, at the Wall Street Journal’s Tech Live event, OpenAI’s chief financial officer Sarah Friar was asked how the company intended to finance its chip and data center commitments. Her answer was specific: OpenAI was looking for an ecosystem of banks, private equity, and a federal “backstop” or “guarantee” that could lower financing costs and increase the amount of debt the company could take on. The interviewer pressed: a federal backstop for chip investment? Friar confirmed.

The retraction arrived within twenty-four hours, from three directions. Friar herself, on LinkedIn: “I used the word ‘backstop’ and it muddied the point.” Sam Altman, on X: “We do not have or want government guarantees for OpenAI datacenters.” David Sacks, Trump’s AI and crypto policy czar: “There will be no federal bailout for AI.” [1]

Three people denying something in twenty-four hours is unnecessary unless the thing they are denying is the plan. She has since been excluded from key financial meetings, her absence described as “notable and awkward.” [2] The person who said the quiet part out loud was not corrected. She was sidelined.

To understand why a federal backstop is the plan — and to understand what Friar was looking at when she said it — you have to look at the financial architecture she was asked to finance. It has a topology that will be familiar to anyone who remembers what happened to telecommunications.

The gap #

This is not an argument that AI does not work. AI is useful. In specific domains it is genuinely productive. Some agentic tools are delivering real value — automating well-defined workflows, accelerating code review, compressing research cycles. The software wrapping LLM calls is evolving, and is already proving it’s use.

But the capital expenditure being committed is not priced for a technology that is useful. It is priced for a technology that is transformative — for something approaching artificial general intelligence, arriving on a timeline measured in months rather than decades. The problem is that the underlying architecture cannot deliver that. Large language models predict the next token; they do not model the world, plan across steps, or reason about consequences. The people who built them have begun saying so publicly. Ilya Sutskever, the OpenAI co-founder whose work established the scaling hypothesis the industry runs on, said in late 2025 that pre-training scaling is “essentially tapped out” and that another 100x of compute “won’t get a qualitative change in capability.” [3] Dario Amodei at Anthropic predicted powerful AI “as early as 2026” in October 2024. [4] By February 2026 he was saying “I don’t believe we’re basically at AGI” and acknowledging that if his revenue forecast was off by a year, “there’s no force on earth, there’s no hedge on earth that could stop me from going bankrupt.” [5] The commitments that underpin the industry were made against the 2024 timeline, not the 2028 one. The estimates have moved. But the money has only increased.

What the field requires, by their own account, is research — not half a trillion dollars in concrete and copper. The buildout is an answer to a question the technology has not yet resolved, built at a scale that forecloses the possibility of changing course when the research points somewhere else.

J.P. Morgan, modeling what the buildout would need to earn to clear a ten percent return on current capital expenditure, arrived at roughly six hundred fifty billion dollars per year in AI-sector revenue — the equivalent of thirty-five dollars per month, in perpetuity, from every iPhone user on earth. The current run-rate is about twenty-five billion. The gap is twenty-six-fold. [6] Goldman Sachs’ chief economist concluded that AI had contributed “basically zero” to U.S. economic growth in 2025 and observed that “FOMO, not ROI, is driving hyperscaler capex.” [7] The San Francisco Federal Reserve’s February 2026 consensus: “While GenAI and related applications are useful, they are not the innovation that spurs broad-based reorganization of the economy.” [8]

The current architecture will not close this gap. It will not close it because the capabilities the spend assumes — autonomous reasoning, reliable multi-step planning, self-correction without human oversight — are not properties of next-token prediction at any scale. The people who built the systems have said so.

The flagship products themselves are evidence of the deficit. ChatGPT, Claude, and Gemini are not LLM calls — they are complex engineering systems that use the model’s output to orchestrate conventional software: search engines, code interpreters, calculators, file systems, external APIs. The LLM produces text; the application routes that text through tools that do the actual work. The industry is compensating for the limitations of the architecture with the same software engineering the architecture was supposed to replace. If the models could plan, verify, and reason about consequences, the scaffolding would not be necessary. Its existence is the admission.

Inflated demand #

The strongest evidence the bulls have is that token consumption is growing fast. ServiceNow and Uber exhausted their entire annual AI token budgets by mid-year and renegotiated their contracts. [9] Anthropic spent the spring of 2026 throttling third-party programmatic access to Claude because agent traffic was outpacing what its subscription model could absorb. Complaints of Claude Pro users running out of tokens after one or two queries are now prevalent on social media. The lab whose business model requires tokens to flow is rationing token flow. [10]

But a meaningful share of what is being reported as demand is friction, waste, and compensation for the very limitations the architecture cannot resolve. Users write longer prompts to constrain outputs that drift. They make multiple attempts at the same task because the model cannot reliably plan across steps. They build elaborate scaffolding — tracing, retrieval systems, verification loops, chain-of-thought prompting, recursion engines — to approximate capabilities the architecture does not natively possess. Each workaround consumes tokens. Each token is counted as demand. A portion of the consumption curve is not demand for the product. It is demand for the product to be something it is not yet. The spend rises precisely because agentic AI is not the AGI the investment assumes.

The operational waste compounds the inflation. A developer was recently charged a thirty-thousand-dollar overnight invoice from a single runaway agent loop on AWS Bedrock. [11] I have watched senior engineers accidentally spend thousands of dollars in tokens due to a process that didn’t exit. Code-generation tools produce suggestions that get rejected, regenerated, rejected again, each cycle billing tokens for output that is never shipped. The models themselves are getting more expensive per request: Anthropic’s latest flagship ships a new tokenizer that encodes the same text into up to thirty-five percent more tokens, inflating the bill without changing the output; separately, higher default reasoning effort levels produce longer responses the user did not ask for. [12]

The buyers have not learned to manage and the sellers have not learned to price, the two failures meeting in the middle and being reported, in the aggregate, as demand. The buildout is being sized against consumption figures that include their own inefficiency — and the revenue projections required to justify it assume this inflated consumption will grow, not contract, as teams mature and architectures stabilize.

The ouroboros #

Combined capital expenditure by the five companies building the infrastructure — Amazon, Alphabet, Meta, Microsoft, and Oracle — reached roughly $725 billion in 2026, approximately three-quarters directed at AI-specific infrastructure. [13] The figure in 2022 was $162 billion. Capital intensity ratios — Meta at 54%, Oracle at 57%, Microsoft and Alphabet in the mid-to-high forties — would have been unthinkable for trillion-dollar-revenue firms in any prior decade. [14]

The problem is not the scale alone. It is where the money goes. It does not reach the broader economy. It circulates among the same handful of companies, and the financial architecture connecting them has an interesting topology:

• Microsoft has committed approximately thirteen billion dollars to OpenAI. OpenAI has committed two hundred and fifty billion to Azure cloud spending. Microsoft records the Azure consumption as revenue, reports AI business at a thirty-seven-billion-dollar annual run rate — a figure traced directly to OpenAI’s spend flowing back through Azure — and uses the growth to justify $192 billion in 2026 capex. [

[15](#ref-15)] - Alphabet has invested up to $40 billion in Anthropic and contracted a multi-gigawatt TPU deal worth tens of billions in return. [
[16](#ref-16)] - Amazon has invested up to $33 billion in the same company, with over a hundred billion in commitments flowing back over ten years. [
[17](#ref-17)]

The investment generates losses, which generate infrastructure spending, which generates revenue, which absorbs the losses. The economist Paul Krugman and the technologist Azeem Azhar have each described the arrangement as a “financial ouroboros.” [

[18]] In the five years following the Telecommunications Act of 1996, telecom companies built an identical loop. Qwest sold indefeasible-right-of-use fiber capacity to Global Crossing while Global Crossing sold capacity back — bilateral swaps that inflated reported revenue on both sides. In a single quarter in 2001, reciprocal transactions accounted for thirty-two percent of Global Crossing’s reported cash revenue; without them, adjusted EBITDA swung from positive four-hundred-seventy-two million to negative forty-three million. [19] The SEC later found Qwest had fraudulently recognized over $3.8 billion in revenue through arrangements of this kind. Employees had a name for the practice. They called it heroin, because “each quarter’s gap between real revenue and the projection required a larger dose to fill.” [20]

When Microsoft invests thirteen billion in OpenAI and books the returning Azure spend as AI revenue growth, the topology is the same one Qwest built: each participant reports the other’s spending as income, and the aggregate is presented to investors as organic demand. The question is whether the underlying economics are different enough to justify the arrangement. The gap — twenty-six-fold — suggests they may not be.

The intermediaries #

Where money circulates at this scale, intermediaries appear to position themselves inside the loop. In telecom it was companies like ITC DeltaCom and McLeodUSA — debt-financed carriers that leased capacity from the majors and sold services back into the same ecosystem, each borrowing against revenue that depended on the majors continuing to spend. The arrangement worked until it didn’t; twenty-three telecom companies filed for bankruptcy between 2000 and 2002. [21]

The AI equivalent is CoreWeave. It has raised approximately $28 billion in combined equity and debt in twelve months. It derives roughly two-thirds of its revenue from Microsoft. Its largest remaining contracts are with OpenAI and Meta — the same companies whose hyperscaler backers are, at one remove, the source of the capital CoreWeave borrowed to buy the hardware it rents to them. In March 2026, CoreWeave closed an $8.5 billion term loan rated A3 by Moody’s, the first investment-grade-rated financing in history secured by GPU infrastructure. [22] The rating is a judgment about Microsoft’s and Meta’s creditworthiness applied to debt whose underlying collateral is rapidly depreciating semiconductor hardware — replaced roughly every four to six years due to obsolescence.

Because of that rating, pension funds, insurers, and money-market vehicles can now hold this paper by mandate. The risk has been intermediated; it has not been reduced. The same was true of the telecom debt that pension funds held in 2001. When the loop collapsed, California’s public employees’ retirement system lost $565 million on WorldCom alone; New York’s state retirement system, $300 million; Texas, $277 million.

The lock #

The reason none of them can stop is that the investment, the revenue, and the justification for the next investment are the same transaction. If Microsoft reduces its OpenAI commitment, it loses one of Azure’s largest customers, the AI revenue line that justifies $192 billion in capex, and the earnings growth that holds its stock price — all at once. The same logic binds Alphabet and Amazon to Anthropic: the equity position and the cloud contract are the same bet, and unwinding one unwinds both.

WorldCom’s CEO told investors that internet traffic was doubling every hundred days. In reality it was doubling roughly once a year. [23] But the lie was not just promotional: WorldCom needed the projection to justify the debt that financed the acquisitions that produced the revenue that covered the interest on the debt. Each quarter’s gap between real revenue and the projection required a larger dose to fill. The arrangement could not tolerate the truth about its own demand curve.

The AI buildout has the same property. Beneath the hyperscaler commitments, the debt compounds the lock. CoreWeave’s A3 rating, Meta’s SPV financing, Oracle’s bond offering — each is underwritten by the assumption that hyperscaler commitments continue. If any one pulls back, the ratings reprice, potentially to junk in a single cycle, and the institutional investors holding the paper absorb the difference. No single company can exit without triggering consequences for every other company in the chain.

Of the five, in my opinion, Meta is the one to watch for the first pullback. Its capex-to-revenue ratio is the highest among the trillion-dollar-revenue hyperscalers; Barclays models a roughly ninety-percent collapse in its free cash flow. Meta is the only one of the five hyperscalers without a cloud computing business to monetize the infrastructure externally. JPMorgan downgraded the stock in April 2026, citing a “challenging path” to capitalizing on AI spending. When asked on the Q1 earnings call about evidence of return on investment from $145 billion in annual AI capex, Zuckerberg called it “a very technical question.” The trigger, when it comes, will likely be a quarter in which ad revenue growth decelerates but Meta keeps telling investors it will spend just as much or more on AI infrastructure. The fallout would not be contained to Meta’s stock price. Its SPV financing, its bond issuance, and the pension and insurance portfolios holding both would reprice simultaneously.

The regulatory friction that might have slowed the spend has been removed. The Trump administration launched Stargate from the White House with Altman and Ellison on stage, issued executive orders streamlining data center permitting, and in December 2025 preempted state AI laws entirely. Between April and June of 2025, twenty data center proposals worth $98 billion were blocked or delayed by local opposition. The industry’s response was not to slow down. It was to secure federal override of local objections.

The backstop #

This is where Friar’s remark becomes legible. She was not freeforming. She was describing the mechanism that can close the gap between what the buildout costs and what it earns. The people building these systems know the architecture is not AGI — they have said so publicly, repeatedly, on the record. They are selling it at a price only AGI could justify, building software scaffolding to maintain the appearance of progress, and financing the difference with instruments designed so that when the gap becomes undeniable, public money covers it. The backstop is not a contingency for a risk they hope to avoid. It is the mechanism that makes the entire structure viable.

The Federal Reserve’s Spring 2026 Financial Stability Report elevated artificial intelligence from fifth to third among perceived threats to American financial stability, behind only geopolitical risk. [26] The institutions whose job it is to identify systemic exposure have begun identifying it.

The precedent Friar was reaching for is not hypothetical. In 2008, $700 billion to stabilize banks while millions of homeowners lost their houses. In 2020, the Fed bought corporate bonds off the open market for the first time, extending the assumption of public rescue from banks to the corporate credit market. In 2023, the FDIC waived its own statutory deposit limit overnight for Silicon Valley Bank. Each intervention was described as exceptional. Each established that institutions whose failure would propagate will be caught — and that the people on the other side of the transaction will not.

The financial system now prices AI infrastructure debt on this assumption. The plan, to the extent that anyone in the room would use the word, is a bailout. No one has voted on it. No legislation has been passed. The assumption holds because the precedent holds. The government will step in, because the alternative — letting the repricing cascade through pension funds, insurance portfolios, and money-market vehicles holding investment-grade AI debt — won’t happen, and surely not with this administration.

Who pays #

The telecom collapse erased more than two trillion dollars in market capitalization. Bond investors recovered twenty cents on the dollar. The FCC chairman testified that the industry owed one trillion dollars, much of which would never be repaid. [21] Tens of thousands of workers lost their jobs and their retirement savings simultaneously.

But there was one consolation. Fiber-optic cable, once laid, has a shelf life measured in decades. When demand eventually caught up, the dark fiber lit up and became the backbone of the modern internet. The investors who overpaid were not made whole, but the physical asset retained value.

Semiconductors will not. GPUs depreciate on a cycle of roughly six years, driven by obsolescence; each new generation renders the prior one uneconomical to operate. The data centers being built today will house hardware that is outdated before the demand the buildout assumes has had time to materialize. When the correction comes — and it is a question of when — the assets at the center of it will not be waiting patiently underground for the world to catch up. They will be waste.

The losses will reach ordinary people through channels that did not exist in 2001. They are already reaching them.

Under the Pension Protection Act, auto-enrollment routes employees into target-date funds that track the S&P 500 by default. The ten largest companies in the index — the same hyperscalers and AI-infrastructure companies driving the buildout — now account for roughly 40% of its weight while contributing 32% of earnings. About forty cents of every default 401(k) dollar flows into these companies. Households are price-insensitive buyers of bubble exposure by design. [27]

The index weights themselves are being inflated by accounting that produces no cash. In Q1 2026, more than half of Amazon’s quarterly profit came from marking up the value of its Anthropic stake — not from selling products or cloud services but from updating the estimated value of an investment. Alphabet reported $28.7 billion of its $62.6 billion quarterly profit from the same source. [28] Each revaluation inflates earnings, lifts the company’s index weight, and reweights the 401(k) that buys more of the inflated stock. No cash changes hands. Neither Anthropic nor OpenAI has gone public. Anthropic is now in discussions to raise at least $30 billion at a valuation exceeding $900 billion. [29] The growth was captured privately. The exposure is distributed publicly.

State and local governments are subsidizing the buildout directly. Virginia’s data center sales tax exemption cost $1.6 billion in fiscal 2025 — the state projected $1.54 million when the program was created. Texas will lose $3.2 billion over the next biennium; it projected $180 million. The subsidies produce roughly one permanent job per million dollars of public cost. [30] In neighborhoods near the facilities, electricity prices have risen as much as 267% over five years. [31] Two-thirds of data centers under development since 2022 are in water-stressed areas. [32] The communities absorbing these costs were not consulted.

It has always been a move available to those who build complex systems to use the complexity itself as leverage — to secure investment, financing, and time from people who do not understand what is being built. The railroad barons did it with federal land grants. The defense contractors did it with cost-plus procurement. The telecom executives did it with demand projections they knew were false.

They are building it anyway, because the financial structure ensures they do not need it to work. It needs only to continue. The companies and people at its center are enriched at every step — equity in the labs they fund, stock in the companies they run, fees on the debt they structure, carry on the funds they manage. When Friar said “backstop,” she was not describing a contingency. She was describing the business model. The buildout is underwritten by an implicit guarantee that none of the people who made the bet will be the ones who pay for it.

The same people the buildout was sold as replacing will be the ones who pay for it. They are already paying — in inflated electricity bills, in foregone state revenue, in retirement accounts overweighted to a trade they never chose. The correction, when it arrives, will not introduce a new cost. It will reveal the full scale of the one that has been accumulating. That is the lesson of every bubble that came before this one. What makes this one different is not the pattern. It is the scale of the bet, the speed of the depreciation, and the depth of the channels through which the cost will be distributed.

Employees at Qwest had a name for the practice of filling each quarter’s gap between real revenue and the projection. They called it heroin, because each quarter required a larger dose.

The dose is $544 billion this year.

Epilogue #

My previous post documented what happens to people and teams when output is decoupled from understanding. That was the micro picture: what the pattern looks like from inside the room where the work is supposed to happen. The response was larger than I expected. A considerable number of readers wrote in to corroborate the pattern from inside their own organizations. Because of this, I am drafting guidelines on recommended organizational AI use, downstream of those conversations, which will appear separately and soon. If you’d like to contribute please email.

References #

1. OpenAI CFO Sarah Friar says company isn’t seeking government backstop (CNBC, November 6, 2025) and OpenAI CFO walks back remarks about federal loan guarantees (The Register, November 6, 2025).

2. OpenAI CFO Excluded From Investor Meetings Amid IPO and Spending Clash With Altman (The Deep Dive, 2026), summarizing The Information’s reporting.

3. [Ilya Sutskever — We’re moving from the age of scaling to the age of research](https://www.dwarkesh.com/p/ilya-sutskever-2) (Dwarkesh Patel, November 25, 2025).

4. [Machines of Loving Grace](https://www.darioamodei.com/essay/machines-of-loving-grace) (Dario Amodei, October 2024).

5. [Dario Amodei in conversation with Dwarkesh Patel, February 2026](https://www.dwarkesh.com/p/dario-amodei-2).

6. J.P. Morgan calls out AI spend, says $650B in annual revenue required to deliver a mere 10% return on AI buildout (Tom’s Hardware, 2026).

7. AI boosted US economy by “basically zero” in 2025, says Goldman Sachs chief economist (Tom’s Hardware, 2026) and Goldman finds no meaningful relationship between AI and productivity (Fortune, March 3, 2026).

8. AI Moment: Possibilities, Productivity, Policy (Federal Reserve Bank of San Francisco Economic Letter, February 2026).

9. Anthropic tightens Claude limits as OpenAI courts agent users (Axios, May 14, 2026). ServiceNow and Uber burned through their entire annual AI token budgets before mid-year.

10. Anthropic cuts third-party usage (Axios, April 6, 2026) and AI demand is inflated, and only Anthropic is being realistic (CNBC, April 17, 2026).

11. Bedrock and a hard place: Claude adventure leaves AWS user staring down $30K invoice (The Register, May 14, 2026).

12. Claude Opus 4.7 Pricing (Anthropic, 2026). Opus 4.7 ships a new tokenizer that produces up to 35% more tokens from the same text at an unchanged headline rate; the model also defaults to a higher reasoning effort level, producing longer responses per request.

13. Big Tech’s AI Spending to Reach $725 Billion in 2026 (Statista, 2026), Hyperscalers Hit $700 Billion in 2026 (Yahoo Finance, 2026), and The Magnificent Capex (Ferguson Wellman, May 2026).

14. The Magnificent Capex (Ferguson Wellman, May 2026). 2026 capex as % of revenue: Meta ~54%, Oracle ~57%, Microsoft ~47%, Alphabet ~46%.

15. Microsoft calls for $190 billion in 2026 capital spending (CNBC, April 29, 2026) and What Microsoft’s 10-Q Says About OpenAI (Om Malik, May 1, 2026).

16. Google’s $40B Anthropic Bet (Tech Insider, April 25, 2026).

17. Amazon commits up to $25B more to Anthropic; Anthropic to spend $100B+ over 10 years on AWS (Amazon, April 20, 2026).

18. Should we worry about AI’s circular deals? (Noah Smith, 2025), discussing the “financial ouroboros” framing from Paul Krugman and Azeem Azhar; see also Paul Krugman, Talking AI With Martin Wolf.

19. Capacity Swaps by Global Crossing and Qwest: Sham Transactions Designed to Boost Revenues? (U.S. House Energy and Commerce Subcommittee hearing, 2002).

20. SEC v. Qwest Communications International Inc. — Litigation Release 18936 (Securities and Exchange Commission).

21. The Great Telecom Implosion (Paul Starr, Princeton, 2002) and Telecoms crash (overview).

22. CoreWeave Takes As Much Financial Engineering As It Does Datacenter Design (The Next Platform, April 9, 2026); CoreWeave closes $8.5B DDTL 4.0, rated A3 by Moody’s (CoreWeave 8-K, March 31, 2026).

23. The Spreadsheet that Fueled the Telecom Boom — and Bust (Anthony J. Pennings, PhD) and Did WorldCom Puff Up the Internet Too? (Light Reading).

24. A Behind-the-Scenes Look at How Meta Raised $30 Billion for Its AI Data Center (TipRanks, 2026) and Blue Owl and Meta close record $30bn financing for AI data centre expansion in Louisiana (Private Equity Insights, 2026).

25. Meta is spending up to $145 billion this year on AI. When asked about signs of ROI, Zuckerberg said ‘that’s a very technical question’ (Fortune, April 29, 2026).

26. Federal Reserve Spring 2026 survey highlights geopolitical risks, AI concerns as top threats to financial stability (2026), and Speech by Vice Chair for Supervision Bowman on artificial intelligence in the financial system (Federal Reserve Board, May 1, 2026).

27. Your 401(k) Is Propping Up the AI Bubble (ProMarket / Stigler Center, May 5, 2026).

28. Half of Google’s and Amazon’s “blowout AI profits” came from a stake in Anthropic — not from their actual business (Fortune, April 30, 2026).

29. Anthropic in talks for $30B at $900B valuation (Bloomberg, May 12, 2026).

30. Cloudy With a Loss of Spending Control: How Data Centers Are Endangering State Budgets (Good Jobs First, April 2025) and Many states don’t report losses from data center tax breaks (Stateline, April 15, 2026).

31. Data Center Power Demands Are Contributing to Higher Energy Bills (Environmental and Energy Study Institute, 2025).

32. The AI Boom Is Draining Water From the Areas That Need It Most (Bloomberg, 2025).

source & further reading

nooneshappy.com — original article