The Two Clocks

One claim says the AI boom is 2008 again. Another says the revenue grows tenfold a year. What each one measures, what Anthropic’s prospectus now shows, and why the closer precedent may be neither.
Two claims about artificial intelligence travel together on social media this autumn, often in the same feed. The first is that the boom is a rerun of 2008: that behind the data centres there is a pile of debt, hidden in special vehicles and lent by funds nobody watches, waiting for one bad quarter to bring it down. The second is that the doubters are looking at the wrong number, because the business underneath is growing tenfold a year, and Anthropic is the usual example.
Both claims are about time. The first counts the bills: what has been borrowed and promised, and when it falls due. The second counts the income: how fast it grows, and for how long. A month ago, in Priced for Tomorrow, this blog set out the valuations of the AI companies against their revenues and concluded that the argument is about timing. Since then the numbers have become firmer. On 26 October Anthropic filed a public prospectus, the first full account of a frontier laboratory’s finances, and the big technology companies reported their quarters.[1] This essay uses them to test both claims.
It takes no side. The same disclosure as before belongs at the start, and matters more here: the essay was written with a tool made by Anthropic, whose figures are at the centre of it. They are treated like any other company’s: the audited numbers as audited, the company’s claims as claims.
A claim about sales
The phrase has an author. In February Dario Amodei, Anthropic’s chief executive, told the podcaster Dwarkesh Patel: “there’s this bizarre 10x per year growth in revenue that we’ve seen. So in 2023, it was zero to $100 million. In 2024, it was $100 million to $1 billion. In 2025, it was $1 billion to $9-10 billion.” Those are run-rates, a month’s revenue multiplied to a year. In the same conversation he explained why the number matters to a company that must buy computing years ahead: “if the growth rate is 5x a year instead of 10x a year, then you go bankrupt.”[2]
The prospectus now gives the revenue recognised in the accounts, which is smaller and audited. Anthropic recognised $41 million in 2023, $386 million in 2024 and $4,585 million in 2025: 9.4 times and then 11.9 times a year. In the first nine months of 2026 it recognised $33.2 billion, 13.1 times the same months of 2025, and in the third quarter alone $17.0 billion (Figure 1).[1] Looking backwards, the claim holds on any like-for-like measure, and for 2026 it understates.
Looking forwards is another matter, and the same filing is the source of the doubt. Growth on the previous quarter was 132 per cent in the first quarter of 2026 and 144 per cent in the second, then 47 per cent in the third; the quarter added $5.4 billion of revenue, against $6.8 billion the quarter before. Forty-seven per cent a quarter is still about 4.6 times a year, but it is not ten. The company gives two reasons of its own. About $600 million of revenue was lost in the third quarter to “price reductions introduced as part of new model versions and contractual discounts”, and an estimated 5 to 7 per cent of its revenue in the first nine months came from customers later found to be using its models against its rules, which it says “may reduce our overall revenue and growth rates in future periods”.[1]
The headline figure in Anthropic’s announcements has a definition for the first time. Its run-rate is the usage fees of the last 28 days multiplied by 13, plus monthly subscriptions multiplied by 12, and it counts sales made through the cloud companies’ marketplaces at their full price rather than Anthropic’s share. On that basis it was “approximately $81 billion as of October 31, 2026, up $10 billion from September 30”. The same document warns that “our Run-Rate Revenue growth rate is expected to decline over time”.[1] OpenAI reportedly counts the same kind of sales net, which is one reason its own figure of about $50 billion is not directly comparable.[3]
The revenue is also less concentrated than the leaks suggested. Two customers each provided 12 per cent of Anthropic’s revenue in 2025; in the first nine months of 2026 none provided more than 10 per cent, and the top five accounted for about 21 per cent of the third quarter, down from about 36 per cent in the first.[1]
Ten times a year of what?
On social media the claim often changes subject on the way: the revenue grows tenfold, so the “capacity” of the model must too. Other quantities are measured, and they grow at different speeds (Figure 2).
The computing used to train the largest models has grown about five times a year since 2020, with an uncertainty of four to six, according to Epoch AI, which tracks it; Anthropic does not disclose its own.[4] The amount of computing needed to reach a given level of performance has halved about every eight months, a gain of roughly three times a year from better methods.[5] Multiplied together, these give something over ten times a year in “effective compute”, but that measures what goes in, not what comes out. The philosopher Toby Ord, looking at the reinforcement-learning stage that has driven recent gains, found that “increasing the RL training compute by 100x as shown in the o1 chart only improved performance from about 33% to 66%”, which he called “breathtakingly inefficient scaling”.[6]
The output measure closest to ten times a year comes from METR, a research group that times how long a software task, measured by how long a skilled person takes, a model can complete half the time. Since 2023 that length has doubled about every 129 days, about seven times a year; measured from 2024 the doubling time is shorter, and a post on the forum LessWrong in February was titled “METR Time Horizons: Now 10x/Year”.[7] Anthropic’s models went from about an hour with Claude 3.7 Sonnet in February 2025 to about twelve hours with Claude Opus 4.6 a year later, with a wide margin of error. The measure has limits that METR itself states: it covers software and research tasks only, at a 50 per cent success rate; “measurements above 16 hrs are unreliable with our current task suite”; and the page has not been updated since 8 September.[7] In METR’s review of the laboratories’ internal models in the spring, on whether AI was speeding up AI research itself, “Anthropic explicitly argues that they had not seen a 2X increase in the pace of progress as of April 2026”.[8]
Two measures point in other directions. The price of a fixed level of benchmark performance has fallen about 47 per cent a quarter since 2023, about thirteen times a year, according to Epoch AI; but the cost of running the newest models at the frontier is rising, by between three and eighteen times a year by one estimate, because they are larger and think longer.[9][10] And when productivity is measured in real work, the effects are far smaller. In METR’s trial in early 2025, experienced developers took 19 per cent longer with AI tools; its follow-up in 2026 was too noisy to give a reliable answer, and a survey of 349 technical workers found a median self-reported gain of 1.4 to 2 times in “the value in their work”, with the caveat that such results “are not necessarily grounded in reality”.[11] The gap between benchmarks and jobs is a subject this blog examined in The Machine and the Hour.[a]
So the claim is true of sales, roughly true of one benchmark and of falling prices, and not true of the computing, the methods or the measured gains at work. That matters for the other claim, because the money is borrowed against the sales.
What has to be paid
Anthropic’s prospectus also lists what it has agreed to buy: “aggregate committed spend of approximately $900.9 billion, substantially all of which is related to compute and infrastructure”, of which $451.6 billion falls before the end of 2030 and $70.7 billion within twelve months (Figure 3). The press figure of about $450 billion “over the next decade” was the first slice; the rest runs to 2045.[1] At the October run-rate, the total is eleven years of revenue; the amount due within a year is about 87 per cent of it.
Two things make this less precarious than the headline. The company holds $132 billion in cash and short-term investments and generated $6.6 billion of cash from operations in the first nine months; and it does not own most of its computing but rents it, so the debt behind the data centres sits with its suppliers. Its operating result swung from a loss of $1.7 billion in the first quarter of 2026 to a profit of $974 million in the second and back to a loss of $567 million in the third, when the cost of computing rose from 48 to 58 per cent of revenue, partly because of $5.0 billion of short-term capacity “which currently costs significantly more”.[1] The net loss of $104 billion for the nine months is almost entirely an accounting charge: the convertible notes and similar instruments sold to investors rose in value with the company, and the rise is booked as a loss.[1]
The offering itself shows the circle that worried analysts in October. The offering is of 825 million shares at $70 to $78, and alongside them $26.0 billion of non-voting shares to its suppliers of chips and computing: AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC, with Google and Nvidia taking $10 billion each, Google’s part counting towards money it had already committed.[1] At the range, by this essay’s arithmetic on the shares outstanding after the offering, the company would be valued at roughly $0.92 to $1.02 trillion, close to the $965 billion of its private round in May and far below the $2 trillion talked about in September. Pricing is expected this week.[1][12]
The suppliers carry the larger bills. Alphabet spent $51.4 billion on property and equipment in the third quarter against $42.3 billion of cash from operations, its second quarter in a row of negative free cash flow; Meta’s free cash flow was negative $2.1 billion; Amazon’s over the past twelve months was negative $42.4 billion.[13][14][15] The gap is borrowed or raised. In the weeks after their results Meta sold $12.5 billion of mandatory convertible preferred shares, $15 billion of dollar bonds and €9 billion of euro bonds; Microsoft priced $25 billion of notes; Oracle, which had promised investors on 28 October to limit new capital to $20 billion through 2030, raised the whole $20 billion within a week, half in shares and half in bonds paying between 7 and 8.65 per cent.[14][16][17] The demand is still there: Google’s cloud revenue doubled in a year and Microsoft’s Azure grew 48 per cent. But Microsoft’s contracted future revenue, up 73 per cent in a year, was flat on the previous quarter at $677 billion, and Microsoft alone booked $10.0 billion of revenue from OpenAI in a single quarter, a reminder of how much of the laboratories’ spending is the suppliers’ income.[13][16]
Houses, leverage and overnight money
To test the comparison it helps to recall what 2008 was made of. American mortgage debt doubled from $5.3 trillion in 2001 to $10.5 trillion in 2007. In 2006 alone $600 billion of subprime loans were made, most of them packaged into securities; between 2003 and 2007 Wall Street created $4 trillion of mortgage-backed securities and nearly $700 billion of the more complex instruments built from them.[18] The subprime segment came to about $1.5 trillion, 10.9 per cent of American output.[19]
Three features turned losses on those loans into a world crisis. The first was leverage. The five large investment banks had, “by one measure, their leverage ratios ... as high as 40 to 1, meaning for every $40 in assets, there was only $1 in capital to cover losses. Less than a 3% drop in asset values could wipe out a firm.” The second was short-term funding: “much of their borrowing was short-term, in the overnight market”, so lenders could leave in a day. The third was a multiplier: insurance contracts and synthetic securities that bet on the same mortgages, so that one loss was owed several times; AIG alone had written protection on $79 billion of them.[18] The IMF later put the potential write-downs at nearly $4.1 trillion, about two-thirds borne by banks.[20] And the debt sat with households: some four million families lost their homes, $17 trillion of household wealth disappeared within 21 months, and the American economy lost 3.6 million jobs in 2008 alone.[18]
Debt that is hard to see
The case for the comparison rests on a few real features, and its most detailed version comes from the writer Ed Zitron, who argued in July that “the comparison between an AI data center and a CDO might seem a little ridiculous, but they’re actually remarkably similar”, and that “the AI bubble is, like the great financial crisis, a product of information asymmetry”.[21] The investor Michael Burry, famous for betting against mortgages before 2008, has written that “the shenanigans are apparent today for those that care to look”, though his usual comparison is with 2000.[22]
The first feature is debt kept off the books. Meta financed a data centre in Louisiana through a vehicle that sold $27.3 billion of bonds, described by the Bank of England as one of the “project-finance style off-balance sheet vehicles” now common in the industry; the IMF counted about $650 billion of debt in the circular deals between AI companies, of which “about $320 billion ... is taken on by investees”, including guarantees, contingent financing and “other off-balance-sheet debt”.[23][24] Howard Marks, who lends for a living, noted that such vehicles “were among the biggest contributors to Enron’s precariousness and eventual collapse”.[25]
The second is speed. Global AI-related debt issuance reached about $450 billion by early September, according to Morgan Stanley figures cited by the Bank of England, more than double the whole of 2025; the hyperscalers’ share of the world’s investment-grade bond issuance has risen to 10 per cent from under 1 per cent in 2024.[26][24] The Bank of England’s committee wrote in September that “the increasing indebtedness of AI firms combined with opacity and, at times, ‘circular arrangements’ ... can complicate the assessment of risks and could amplify losses if expectations disappoint”.[26]
The third is where the debt ends up. Much of it is held by private credit funds, which lend outside the banks and which this blog described in Owed to Ourselves; up to 30 per cent of the $3 trillion that data centres are expected to need could come from them, according to market intelligence cited by the European Central Bank.[19] Pension funds and insurers hold more than a quarter of the public corporate bonds in the United States and the euro area.[24] And the banks are not outside the circle: their loans to non-bank lenders rose 18 per cent in 2026 and to private equity and private credit firms 30 per cent, according to the IMF, and their credit lines to non-banks reached $2.6 trillion, according to the Federal Reserve.[24][27] The president of the Kansas City Fed, Jeff Schmid, asked in September: “are we moving to a too-big-to-fail AI ecosystem”.[28]
The fourth is the price of borrowing for the weakest names. Insurance against default by CoreWeave, which rents out Nvidia chips, cost the equivalent of more than 11 per cent a year in early November; its interest bill was $846 million in the third quarter, on revenue of $3.7 billion.[29] Those are the prices of a company the market thinks may not survive, and they are the part of the industry where the comparison bites hardest.
No households, no overnight money
The official bodies that watch the financial system describe the risks above, and none of them says that this is 2008. The IMF’s October report, which put an AI-driven correction first among rising risks, makes no comparison with the crisis, and finds that stress in private assets on its own “appear[s] absorbable by banks”.[24] The Bank of England wrote that “the predominantly cash and equity-financed nature of AI development to date limits the likelihood of severe systemic risk implications of a change to AI-driven asset prices in the near term”.[23] The one official text that runs the comparison explicitly, the ECB’s review in May, set all of American private credit, $1.4 trillion or 4.7 per cent of output, against subprime’s 10.9 per cent, and concluded: “Unlike subprime lending, leverage in private credit is low, and most funding is long term and not subject to run risk.” In its stress scenario, losses for euro-area banks did not exceed 1.3 per cent of their equity.[19]
The three features that made 2008 a crisis are mostly missing. Leverage is lower: Torsten Slok, chief economist of the private-credit firm Apollo, an interested party, put bank leverage today at “only around 10”, against more than 30 at Bear Stearns and Lehman, and the leverage of the listed private-credit funds at one to two.[30] Funding is longer: insurers lend from annuities of twenty or thirty years, not from overnight money that can leave in a day. No multiplier has been found: Zitron himself concedes that “the systemic risks aren’t driven by derivatives”.[21] And households owe none of it. They are exposed as owners of shares and pension funds, which could lose a great deal, but nobody loses a home because a data centre stands empty.
The scale is smaller too (Figure 4). The press tallies of AI data-centre debt, a little over $500 billion, come to about 2 per cent of the $24 trillion owed by American non-financial companies; the mortgage debt of 2007 was some twenty times larger.[31][21]
Two qualifications keep the defence honest. “Funded from cash” was true in 2024 and 2025; in 2026, with free cash flow negative at three of the four largest builders, it is cash, shares and rising debt. And the warnings are about direction more than size: the ECB noted that private credit “could quickly grow much larger” if it becomes the main lender to AI.[19] A system can be safe at today’s size and not at the size it is heading for. How tight couplings turn small failures into large ones is the subject of this blog’s Nothing to Spare.[b]
The closer precedent
If not 2008, then what? The precedent that Marks and others reach for is the telecom bust of 2001, the debt-funded half of the internet bubble.[25] Telecom companies laid “$90 billion worth of fiber-optic cable between 1997 and 2001”, and the glut drove “usage of the long-distance backbone below 3 percent in April 2001”. The communications-equipment index fell 86 per cent, “wiping out $793 billion in shareholder equity”, and the industry was left “with massive debt loads and few means of paying creditors”.[32] In the first nine months of 2002, “about 57% of the high yield default dollars were from telecom companies”, and telecom bondholders recovered “just 11.64%” of their money, the lowest in the authors’ records.[33] Equipment makers that had lent their customers the money to buy their equipment took losses alongside them.[25]
The bust ruined shareholders, bondholders and suppliers, and left behind fibre that carried the internet for the next two decades. It did not produce a banking crisis, and the recession of 2001 was mild. That is the shape that most of the evidence above points to if the AI build-out disappoints: heavy losses for those who lent to the weakest builders and for the shareholders of the strongest, without the chain from households to banks that made 2008. There is one difference that cuts the other way: fibre lasted decades and chips last a few years, so the collateral behind the newest loans loses value faster.
A harder month
The correction that began in October has widened. On 5 November the Nasdaq Composite fell 5.5 per cent, its worst day since April 2025, leaving it more than 10 per cent below its October record.[34] OpenAI, which had sought at least $30 billion at about $1.4 trillion, reportedly raised between $11 and $15 billion at $1.25 trillion.[3] On 11 November Nvidia’s board accepted the resignation of its founder and chief executive, Jensen Huang, “effective immediately”, and named its finance chief, Colette Kress, interim chief executive; the filing gives no reason.[35] Nvidia reports its quarter on 17 November, Anthropic is expected to price its shares the next day, and the ECB publishes its financial stability review on 18 November.
None of this settles either claim. A falling market says that investors are less sure of the revenue clock; it says nothing yet about the debt clock, which runs on contracts, coupons and maturities, and which has not missed a payment in the places that matter.
Two clocks
Put the two claims side by side and they turn out to be about the same thing seen from opposite ends. The tenfold growth is real in the one place it is usually claimed, the sales of the laboratories, and it is already slowing towards something like four or five times a year. The debt is real too, and growing faster than anything else in the credit markets, but it sits mostly with large companies, long-term investors and funds with little leverage, not with households and overnight lenders. Whether this ends well depends on which clock runs faster: whether the revenue keeps growing quickly enough, and reaches the companies that borrowed, before the commitments, the coupons and the worn-out chips fall due.
Several numbers will show it within a year. Anthropic’s quarterly growth, now reported in public, and whether its share of revenue spent on computing falls again. The free cash flow of the four largest builders, and how much of their spending is borrowed. The price of insurance against default by the weakest borrowers, CoreWeave and Oracle among them, and whether any data-centre vehicle misses a payment. And the size of private credit’s lending to AI, which is the one channel through which a telecom-style bust could become something worse. If those hold, the 2008 comparison will look like a category error. If they do not, the losses will be large and widely shared through pension and index funds, and the right comparison will be 2001: painful, and not the end of the banks.
This piece was written collaboratively with Claude Opus 5.5 (Anthropic): human specification and editorial direction; machine research, analysis and drafting. Anthropic, the maker of the tool used, is the company whose figures are at the centre of the essay; they come from its registration statement and amendments filed with the US Securities and Exchange Commission, read in the original, and its chief executive’s remarks are treated as any other source. Three research agents gathered the sources; quotations from the filings, from the IMF, Bank of England, ECB, Federal Reserve, FCIC, Dallas Fed, NYU Salomon Center, METR, Epoch AI, Oaktree and Zitron were read in the original. Burry’s, Schmid’s and Slok’s remarks, OpenAI’s funding, market moves and CoreWeave’s credit prices are as reported, and the CFR transcript of Slok’s remarks is machine-generated. The IPO had not been priced at the time of writing; the valuation range is this essay’s arithmetic on the shares the filing says will be outstanding, before options and unvested shares. The figures and ratios come from a script, scripts/two_clocks.py; yearly multiples from a single quarter are annualisations, not forecasts.
Cover: computed for this essay (scripts/two_clocks_cover.py).
Authored by: Luis Matos Ferreira — Physicist, Developer, Writer
Ideas this essay uses without stopping to explain them, and where this blog does. The letters in the text point here; the links go straight to the section.
- Priced for Tomorrow — the valuations of the AI companies against their revenues, and both sides of the bubble argument.
- Owed to Ourselves — how money and debt work, and when debt becomes too much.
- Nothing to Spare — whether the world has become more fragile.
- Rented Ground — how much of the world’s computing runs through a few American companies.
- Anthropic, PBC, registration statement on Form S-1, 26 October 2026, and amendments no. 1, 2 November, and no. 2, 10 November 2026 (quarterly results, run-rate definition, commitments, offering and concurrent private placement); Anthropic, “Anthropic files registration statement”, 26 October 2026.
- D. Amodei, interview with D. Patel, “We are near the end of the exponential”, Dwarkesh Podcast, 13 February 2026.
- Financial Times, 8 October 2026, on OpenAI’s annualised revenue, as reported by Axios; Bloomberg, 6 November, and CNBC, 12 November 2026, on its new funding round; as reported.
- Epoch AI, “Training compute of frontier AI models”, Trends, updated 5 February 2026.
- A. Ho, T. Besiroglu et al., “Algorithmic progress in language models”, arXiv:2403.05812, March 2024.
- T. Ord, “How Well Does RL Scale?”, 20 October 2025.
- METR, “Time Horizon 1.1”, 29 January 2026, its data file (corrected March 2026) and the time-horizons page, last updated 8 September 2026; J. Crox, “METR Time Horizons: Now 10x/Year”, LessWrong, 13 February 2026.
- METR, “Frontier Risk Report”, 19 May 2026.
- B. Emberson and D. Roodman, “The plunging price of thought”, Epoch AI, 22 September 2026.
- H. Gundlach, J. Lynch, M. Mertens and N. Thompson, “The Price of Progress”, arXiv:2511.23455, version 2, March 2026.
- METR, “update on the developer productivity study”, 24 February 2026, and “AI usage survey”, 11 May 2026.
- Decrypt and BNN Bloomberg, 16 November 2026, on the expected pricing date, as reported.
- Alphabet, third-quarter 2026 results, 28 October 2026.
- Meta Platforms, third-quarter 2026 results, 21 October 2026; Forms 8-K of 29 October, 5 November and 12 November 2026 (preferred stock and notes).
- Amazon.com, third-quarter 2026 results.
- Microsoft, first-quarter fiscal 2027 results and Form 10-Q for the quarter to 30 September 2026; Form 8-K of 4 November 2026 (notes).
- Oracle, investor day presentation, 28 October 2026, and Form 8-K, 4 November 2026.
- Financial Crisis Inquiry Commission, The Financial Crisis Inquiry Report, January 2011.
- European Central Bank, Financial Stability Review, May 2026, special feature on private credit, section 4, and chapter on financial markets.
- IMF Survey, “IMF sees total financial sector losses at $4.1 trillion”, 21 April 2009.
- E. Zitron, “The Subprime Data Center Crisis”, Where’s Your Ed At, 22 July 2026 (citing Bloomberg for the data-centre debt tally).
- M. Burry, Cassandra Unchained, August 2026, as reported by 24/7 Wall St.
- Bank of England, Financial Stability Report, December 2025.
- IMF, Global Financial Stability Report, chapter 1, “Beneath the Resilience: Rising Risks, Unresolved Vulnerabilities”, October 2026.
- H. Marks, “Is It a Bubble?”, Oaktree Capital memo, 9 December 2025.
- Bank of England, Financial Policy Committee summary and record, September 2026.
- Federal Reserve Board, Financial Stability Report, May 2026.
- J. Schmid, remarks of 25 September 2026, Reuters, as republished by Kitco.
- CoreWeave, third-quarter 2026 results; Deutsche Bank data on its credit default swaps, early November 2026, as reported by CNBC.
- T. Slok, “Could the AI debt problem create the next financial crisis?”, The Spillover, Council on Foreign Relations, 26 August 2026 (machine-generated transcript).
- Federal Reserve, Financial Accounts of the United States, Z.1, second quarter 2026, 11 September 2026.
- J. Thompson, “Is Telecom Disconnected or Just on Hold?”, Federal Reserve Bank of Dallas, Southwest Economy, January/February 2002.
- E. Altman and G. Bana, “Defaults and Returns on High Yield Bonds: Analysis Through September 30, 2002”, NYU Salomon Center, October 2002.
- Market reports, 5–13 November 2026 (Nasdaq Composite), as reported.
- NVIDIA, Form 8-K, 12 November 2026.
Comentários
Enviar um comentário