In the AI Boom, the Losses Only Surface When Someone Needs a Loan

The most revealing financial number of the week did not come from an earnings report. Jane Street, the intensely private New York trading firm, lost roughly $15 billion in July, its first losing month in about a decade, according to Reuters reporting¹. The loss traces to the firm's stake in Situational Awareness, the AI hedge fund run by 24-year-old former OpenAI researcher Leopold Aschenbrenner, which returned 439% in the first half of the year and then faced margin calls that forced a fire sale of its equity book to Citadel³. And here is the detail that matters most: the number became public only because Jane Street was preparing to issue $14.6 billion of bonds² and had to tell prospective lenders about it.

That disclosure channel, sideways and accidental, is the story of AI finance in 2026. The buildout is no longer funded mainly by equity that trades every day and reports every quarter. It runs through private credit funds, special-purpose vehicles, vendor guarantees, and lease commitments that sit off balance sheets. The losses in this system are real and, so far, absorbed. But you learn about them late, and usually because a financing transaction drags the number into the open. If you want to know who bears the risk when AI capital spending outruns AI revenue, follow the disclosure lags, because that is where the risk has pooled.

Consider the same week's other headline. Nvidia had been negotiating a financial backstop for OpenAI's planned 10-gigawatt data center campus in Ohio, effectively lending its investment-grade balance sheet to a customer that, per reporting on the deal⁵, remains unprofitable at an $852 billion valuation. After Nvidia's stock fell 5% when the $250 billion figure first leaked, the guarantee was cut to under $120 billion⁴, covering only the project's first phase, and Nvidia simultaneously launched platforms with six major financial institutions to mobilize over $500 billion of third-party capital for AI infrastructure. Read charitably, that is price discovery working: shareholders objected, the commitment shrank, and the risk moved to institutions paid to underwrite it. Read less charitably, the world's most valuable chipmaker was days from guaranteeing a quarter-trillion dollars of a customer's debt, and we know the size of the walk-back only through leaks.

The institutions absorbing what Nvidia declined to hold are the ones the Bank for International Settlements now worries about. In its March Quarterly Review, the BIS described how hyperscalers (the cloud giants: Amazon, Microsoft, Google, Meta, Oracle) increasingly finance data centers through special-purpose vehicles in which the tech company holds a minority stake and commits to long-term leases, while private credit funds and insurers hold the debt. The BIS calls these arrangements "shadow borrowing"⁶, obligations economically akin to debt that reside outside corporate balance sheets. The scale is not marginal. Moody's found the five largest hyperscalers carry $662 billion in signed-but-not-commenced data center leases⁷, equal to 113% of their adjusted debt, within a total of $969 billion in future lease commitments. Those obligations are invisible in standard leverage metrics today and will land on balance sheets over the coming years⁸ as facilities are delivered, whether or not AI revenue has arrived to service them.

And the first tremors have already run through the retail end of this chain. Private credit, meaning loans made by nonbank funds rather than banks or bond markets, has become the workhorse of AI-adjacent lending, and its semi-liquid retail vehicles are showing strain. The Federal Reserve's May Financial Stability Report documented that redemption requests at nontraded business development companies surged well past the standard 5%-of-assets quarterly cap⁹ in early 2026, driven partly by fears that AI itself will disrupt the software companies these funds lend to. Blue Owl permanently halted redemptions in its OBDC II fund¹¹ in February, converting investor exits into return-of-capital distributions on the fund's schedule, not theirs, while its tech-focused fund saw redemption requests hit roughly 15% of net asset value¹².

I want to be careful about what this evidence does and does not show, because several facts genuinely narrow my alarm. Jane Street absorbed its loss and still reports over $40 billion in trading revenue year to date¹; no bank was impaired, no margin went unpaid. Blue Owl raised cash by selling loans to pensions and insurers at prices near par, not in a distressed liquidation. The Fed judged the redemption wave "limited and manageable"¹⁰. And the revenue underneath the valuations is not imaginary: Anthropic disclosed run-rate revenue above $47 billion¹³ when it raised $65 billion at a $965 billion valuation in May, which works out to roughly 20 times revenue, actually the cheapest multiple¹⁴ among the big private AI names. Anyone declaring this a replay of 2008 has to explain why the ultimate obligors, cash-gushing hyperscalers with contracted leases, resemble subprime borrowers. They do not.

But notice what those reassuring facts have in common: every one of them describes a system that absorbs shocks by deferring them. The gated fund does not realize losses; it stops letting investors force the question. The uncommenced lease does not strain a credit metric; it waits. The secured, covenant-protected private loan does not trade down in public; it gets remarked quarterly by its own manager. These features are sold as stability, and in one sense they are. In another sense they mean the AI capital stack has systematically replaced daily price discovery with episodic, transaction-forced disclosure. Stress does not clear continuously. It accumulates and then surfaces in lumps, the way Jane Street's July surfaced in an August bond prospectus.

Which is why the capability question matters more than the finance crowd admits. Mark Zuckerberg told Meta staff in July that agentic AI development "hasn't really accelerated in the way that we expected"¹⁵, even as Meta spends up to $145 billion on AI infrastructure this year. That is one company's four-month product miss, not proof of an industry plateau, and Anthropic's revenue curve argues the other way. Still, the entire deferred-disclosure edifice, the leases, the SPVs, the 20x multiples, is underwritten by an assumption of steep, sustained revenue growth. If growth merely slows to excellent, the marks in the private stack are wrong, and the people holding them are, in order: retail investors in semi-liquid credit funds who have already discovered they cannot exit, the insurers and pensions buying loan portfolios from gated vehicles, and finally hyperscaler shareholders, onto whose balance sheets $662 billion of commitments will migrate regardless.

So return to that bond prospectus. Jane Street's loss was survivable, even trivial against its year. What was not trivial is that a $15 billion hole in a systemically active trading firm was invisible until a refinancing required honesty. The next test arrives with Anthropic's expected autumn IPO, when the largest private AI valuation finally meets continuous public pricing. I expect the offering to succeed and the multiple to compress, and I expect the same pattern everywhere else: not a crash but a series of forced disclosures, each revealing that the risk was where the reporting wasn't. The system is bending in exactly the places we cannot watch, and we keep finding out by accident.