Bought Long, Sold Short

On 9 March 2023, customers of Silicon Valley Bank withdrew 42 billion dollars in a single day. Around 100 billion more was queued for the next morning. That is what the US central bank's review says (Federal Reserve). The bank did not open again.

Silicon Valley Bank was not broke in the usual sense. It had put its customers' money into long-dated government bonds, safe paper that had merely lost value when interest rates rose. The problem was the term. Customers could withdraw at any time. The bonds ran for years.

Douglas Diamond and Philip Dybvig described this structure in 1983, and in 2022 they won the Nobel Prize for it, together with Ben Bernanke (Nobel Committee). A bank borrows short and lends long. That is useful, as long as not everyone wants their money at once. If they do, no amount of assets is enough, because they sit in the wrong term. That is why there is deposit insurance, liquidity rules and a central bank that steps in when needed.

The AI industry has a similar structure, only inverted. On the spending side it commits for years and decades. On the revenue side it sells to customers who can leave almost any time. For this piece I read Anthropic's IPO filing, submitted on 26 October 2026, along with the quarterly reports of Meta and Nvidia and the paper on financing the buildout that Stijn Van Nieuwerburgh presented to the Brookings Institution in September. This text is not about whether AI is useful. It is about who carries the time.

Sold short

Anthropic is the first large AI lab to publish audited figures. The filing is surprisingly frank about the customer side. "Many of our customer contracts allow customers to reduce usage or terminate with limited or no notice. Customers are not obligated to renew," it says on page 42 (Anthropic, Form S-1). 80 per cent of revenue in the first half of 2026 was consumption-based, only 20 per cent subscriptions (p. 138). Ask less, pay less.

How mobile these customers are shows in Anthropic's own rise. In September 2025, 16 per cent of US businesses on Ramp paid for Anthropic, according to the Ramp AI Index. A year later it was 43.8 per cent, more than for OpenAI (Ramp). The share of firms switching provider hit a record 8 per cent a month in September, as a three-month average. Ramp writes: "It could be an early sign that AI models are becoming commodities" (Ramp, 6 Oct 2026). The same mobility that made Anthropic big can make it smaller again.

It is not only the contract that is short, it is the product too. Models released in the six months to the end of September brought in more than 80 per cent of revenue in the third quarter (p. 143). After half a year a model is no longer a source of revenue but a predecessor.

Bought long

On the other side are obligations whose terms the filing lists one by one (pp. 170 f.). At Google, Anthropic has committed to minimum spending of 114.5 billion dollars, running until July 2033. At Amazon to at least 110 billion dollars, running until April 2036. At Microsoft to 44.9 billion, until May 2033. On top come five-year leases for TPU chips via Broadcom worth 161.2 billion dollars, computing power at SpaceXAI until 2029 and at Nscale and Lambda until 2033 and 2034.

These commitments, Anthropic writes, "must be satisfied regardless of our revenue or utilization" (p. 141). They apply whether or not a customer cancels tomorrow.

Further up the chain, the terms get longer still. Since August 2026, Nvidia has guaranteed up to 105 billion dollars for land, power and buildings of a data center in Ohio to be used by an affiliate of OpenAI, under leases running 20 years (Nvidia, 10-Q). And Meta's Hyperion data center in Louisiana is financed with bonds that do not mature until 2049.

Read all the terms on one axis and a simple picture emerges. What the industry sells ends after weeks or months. What it has bought ends after years or decades. In between lies the time in which someone carries the risk.

Covered

How big the gap is can be read from a number every IPO filing contains: remaining performance obligations, revenue secured by signed contracts but not yet booked. At Anthropic it was 10.8 billion dollars on 30 June 2026, about three quarters of it due within twelve months (F-26).

On the same date, fixed obligations for computing power and infrastructure came to about 589 billion dollars, according to the notes (F-35, F-58). By 26 October about 266 billion more had been added (F-65). Put as a percentage: 1.8 per cent of the June commitments were covered by customer contracts.

The calculation is harsh, and it needs two qualifications. First, the figure is typically low in consumption-based businesses: if you bill by usage, you have little signed future. At Oracle, which rents data centers to AI labs, it is the other way round. Its books show a backlog of 664 billion dollars, only about 13 per cent due in the next twelve months (Oracle, 10-Q). Oracle sells long, to customers who themselves sell short.

Second, Anthropic is growing so fast that any snapshot ages quickly. In the second quarter of 2026, revenue rose 1,368 per cent year on year, compute costs 300 per cent (p. 10). Existing customers spent about nine times as much in September as a year earlier (p. 135). The second quarter brought the first operating profit, 974 million dollars (p. 157). And at the end of June, 131.7 billion dollars sat in cash and short-term investments (p. 161). The commitments are a bet that this growth continues. They are not a bet that nobody cancels.

The chain

The gap between selling short and buying long does not disappear. It moves, and at every step it gets a little longer.

Hyperion shows it most clearly. Meta holds 20 per cent of the special-purpose vehicle that builds the data center, the investment firm Blue Owl 80. About 90 per cent of the roughly 30 billion dollars is financed with bonds, 27 billion, paying 6.58 per cent and due in 2049. Meta rents the data center, but not for 23 years. According to Meta's annual report the lease is fixed for an initial four years, with options of up to 20 years, backed by a residual value guarantee of about 28 billion dollars (Meta, 10-K 2025).

Van Nieuwerburgh explains in his paper why Meta builds it this way. It is not about cheaper money; the bonds cost at least one percentage point more interest than Meta's own. It is about "preservation of an asset-light corporate profile" (Brookings, pp. 8 to 10). Neither the future rents nor the residual value guarantee appear on Meta's balance sheet, "even though one of the two commitments will ultimately materialize for sure". Meta itself puts its maximum loss exposure from such vehicles at 68.4 billion dollars (Meta, 10-Q).

What this means for creditors, the paper sums up in one sentence: "Debt service is fixed, while project revenues depend on utilization, lease renewals, and uncertain residual values" (p. 12). The interest comes every year. Whether the tenants renew is decided every four years.

Not a bank run

It would be convenient to bring back Silicon Valley Bank at this point. It would also be wrong, at least in part.

Customers of an AI lab are not depositors. They have no money there that they could withdraw. When they cancel, nothing flows out, nothing flows in any more. A run in which everyone stands at the door on the same morning is hard to imagine for a software provider. What can happen is quieter: growth flattens, prices fall, the commitments remain.

Economic history knows this course better than the bank run. Around 2000, American telecom companies invested an average of 1.1 per cent of GDP a year in fibre, 1.4 per cent at the peak, according to Van Nieuwerburgh's comparison table (Table 1). The cables stayed in the ground for decades; prices for data transmission fell faster than expected. For the AI buildout he puts the figure at 3.6 per cent, more than for the railroads in the 19th century. And unlike fibre, graphics chips are outdated after a few years, in his words "within three to six years" (p. 24).

Silicon Valley Bank had assets in the wrong term. The AI industry has obligations in the wrong term. The difference is that a bank does not have to keep its customers to service its debts. An AI lab does.

Who carries the time

Every term a customer does not sign, someone else carries. At Anthropic it is first the owners, who have put in 141 billion dollars, and the investors now being invited onto the stock exchange. At Hyperion it is bondholders, at OpenAI's data center in Ohio it is Nvidia, and for the loans on the chips themselves it is increasingly private credit. According to Morgan Stanley, about 800 billion dollars of the financing through 2028 is to come from such loans (cited by Van Nieuwerburgh, p. 7).

Customers carry the least. They have the shortest term, the most choice and, as long as providers compete for them, the best prices. For a company buying AI that is a comfortable position. It stays comfortable only as long as nobody has to close the gap.

There are two ways to close it. Either revenue grows into the commitments, fast enough for the sums to work. Or providers lengthen the terms on the customer side: longer contracts, minimum volumes, prices that make switching more expensive. Anthropic hints at both. Margins "may compress", it says on page 133, and it may be required "to reduce our prices" (p. 60). The opposite move is not described in the filing.

A bank whose customers can leave at any time needs deposit insurance. An industry whose customers can leave at any time needs growth. So far it has it. The question is not whether the time between selling short and buying long is a risk. The question is whose risk it is when growth takes a break.

Sources and method

Anthropic: IPO filing (Form S-1), submitted to the SEC on 26 October 2026. Termination p. 42, revenue mix p. 138, model age p. 143, commitments by counterparty and term pp. 170 f., "regardless of our revenue or utilization" p. 141, remaining performance obligations F-26, growth of revenue and compute costs pp. 10 and 136, existing customers (net revenue retention, excluding Pro and Max plans) p. 135, operating profit p. 157, cash and short-term investments p. 161, capital raised p. 10, margins and prices pp. 60 and 133. The fixed obligations of about 589 billion dollars on 30 June 2026 and the roughly 266 billion from contracts signed between July and 26 October are my own sums from notes F-35, F-58 and F-65. Anthropic has not yet stated the total in the filing. The coverage ratio is my own calculation, 10.8 divided by 589. Terms on the axis are counted from October 2026.

Ramp: AI Index, revised series, as of 30 October 2026, payments of more than 85,000 US businesses. Switching rate from the post of 6 October 2026.

Meta: Hyperion lease and residual value guarantee from the 2025 annual report (10-K), maximum loss exposure from the 10-Q for 30 September 2026. Bond volume, coupon, maturity, leverage and spread per Van Nieuwerburgh. Nvidia: guarantee from the 10-Q for 26 July 2026. Oracle: backlog from the 10-Q for 31 August 2026.

Stijn Van Nieuwerburgh: "Financing the AI Buildout", Brookings Papers on Economic Activity, conference draft of 4 September 2026. A revised version with newer data appeared on 2 October 2026 as NBER Working Paper 35865. Silicon Valley Bank: Review of the Federal Reserve's Supervision and Regulation of Silicon Valley Bank, April 2023. Diamond and Dybvig: "Bank Runs, Deposit Insurance, and Liquidity", Journal of Political Economy, 1983.

Text and research: Jonas Höttler. As of October 2026. The cover image and chapter pictures are concept images, AI-generated with Bria.