The BIS documents that combined capex of the top US hyperscalers (~$370B TTM) has surpassed operating cash flow for the first time this cycle, forcing Meta, Oracle, and Alphabet to tap bond markets at record scale. They frame this as the classic pattern where a tech boom outgrows its cash generation and credit steps in to bridge — strongly implying this is when systemic risk begins to accumulate.
By surfacing the BIS bulletin to the HN front page with a straightforward title, the submitter frames the financing shift as newsworthy for a developer audience — treating it as a signal that the AI infrastructure narrative has entered a more precarious phase worth examining.
The BIS highlights a rapid move toward off-balance-sheet vehicles where private lenders like Blackstone, Apollo, and KKR fund the buildings and GPUs while hyperscalers sign long-dated leases. This composition change — from equity holders to bondholders and private lenders holding the marginal AI dollar — is presented as a distinctive feature of this cycle, not just a scaling of existing debt.
The editorial argues that until now, the AI buildout was comfortably financed by ads and cloud cash flows, which made 'AI capex' look like a rounding error on a great business. The BIS bulletin marks the inflection where that framing breaks — the marginal dollar is now debt, and the risk profile of the entire AI infrastructure trade has changed for developers and investors alike.
The Bank for International Settlements — the central bank of central banks, not usually a source that developers read on a Tuesday — published Bulletin No. 120 on financing the AI boom. The headline finding is quiet and specific: for the first time in this cycle, the combined capex of the largest US hyperscalers has moved beyond what their operating cash flows can cover, and the gap is being filled with debt.
The numbers the BIS pulls together: Microsoft, Alphabet, Meta, Amazon, and Oracle spent roughly $370B on capex over the trailing twelve months, most of it on AI infrastructure. Their combined operating cash flow, while still enormous, has not scaled at the same rate. Meta issued $30B of investment-grade bonds in October. Oracle issued $18B in September, its largest raise ever. Alphabet returned to the bond market in April for the first time since 2020. Beyond public debt, the BIS flags a rapid shift toward private-credit-financed data center SPVs — off-balance-sheet vehicles where a private lender funds the building and GPUs and the hyperscaler signs a long-dated lease. Blackstone, Apollo, and KKR have all raised dedicated infra-credit funds targeting exactly this deal shape.
The bulletin frames this as a familiar pattern: a technology boom outgrows its own cash generation, credit steps in to bridge, and the composition of the marginal dollar shifts from equity holders to bondholders and private lenders. What the BIS is careful not to say — but strongly implies — is that this is the point in the cycle where systemic risk starts to accumulate.
Up to this point, the AI infrastructure buildout has been comfortably self-funded. Google, Microsoft, and Meta generated so much cash from ads and cloud that they could pour tens of billions into GPUs without touching the debt markets in any material way. That's the version of the story that made "AI capex" look like a rounding error on a great business. The BIS bulletin marks the moment that story stopped being true.
The mechanics matter. When capex is funded from cash flow, the useful-life assumption on a GPU is an accounting choice with cosmetic effects on reported earnings. When capex is funded from five-to-seven-year debt, the useful life becomes a covenant question. Meta and Microsoft both extended their server depreciation schedules from four to six years in 2023 and 2024 — a move that flatters earnings by spreading the cost over more quarters. If H100s and H200s turn out to be economically obsolete in three years rather than six (Blackwell is already sampling; Rubin lands in 2026), the debt doesn't get any shorter. The mismatch between how fast the silicon depreciates and how slowly the bonds mature is the specific thing the BIS is worried about.
There's a second concern the bulletin surfaces more delicately: circularity. Nvidia has invested in CoreWeave. CoreWeave has signed capacity deals with Microsoft. Microsoft has an equity stake in OpenAI. OpenAI has committed hundreds of billions in future compute to Oracle, Microsoft, and CoreWeave. Oracle borrows to build the data centers to serve that commitment. The BIS notes, in central-bank prose, that "revenue commitments among vertically related counterparties can obscure the underlying credit quality of the buildout." In plainer English: a lot of the demand backing these debt issuances is a promise from one AI company to pay another AI company using money that has not yet been raised.
Community reaction on Hacker News centered on two threads. First, that the private-credit angle is the least visible and most concerning — SPV-financed data centers sit off the balance sheet of the tenant, so the aggregate leverage in the sector is systematically understated by looking at 10-Ks. Second, that this isn't 2000. The counterparties are cash-generative giants, not pre-revenue startups, and even a sharp correction wouldn't produce dot-com-style zeros. Both things can be true. The 2008 comparison — where the risk wasn't in the assets themselves but in the funding structures wrapped around them — got more traction in the comments than the 2000 one, and probably deserves to.
For developers, the temptation is to file this under "macro noise, doesn't affect my sprint." That's mostly right for the next 12 months and mostly wrong past that. Every capacity-planning assumption you make that depends on GPU prices continuing to fall, or on your cloud provider absorbing another round of margin compression, is now leaning on the bond market rather than on Google's free cash flow.
A few concrete implications. If you're on reserved-instance pricing for H100/H200 capacity, your provider's incentive to hold that price steady weakens if their cost of capital rises — reserved pricing has already crept up 15-20% at the major clouds over the last two quarters, and the BIS data suggests that trend has a floor higher than most planners assumed. If you're building on managed inference APIs, the pricing floor is now partly a function of what OpenAI, Anthropic, and Google need to charge to service the compute commitments they've signed, not just marginal token cost. And if you're one of the shops still evaluating whether to buy your own GPUs versus rent — the calculus that made renting obviously correct at 2023 prices is closer than it used to be, especially for steady-state inference workloads with a two-to-three-year horizon.
The less obvious implication is on model choice. When compute funding gets tighter, the pressure to run smaller, quantized, distilled models — the boring engineering work rather than the frontier — gets stronger. Expect the next 18 months of provider announcements to lean harder on efficiency wins (MoE, speculative decoding, quantization-aware training) and less on parameter counts. That's a good thing for anyone shipping products; it's a mixed thing for anyone whose roadmap assumes the frontier keeps moving at the current pace.
The BIS doesn't make predictions and neither will we. But the bulletin has a specific tell in its final section: it flags concentration risk in the private credit funds that are underwriting the SPV wave, and notes that stress in that segment historically leads equity market repricing rather than following it. If you want a leading indicator for whether the AI capex cycle is turning, watch private-credit spreads on data center SPVs, not Nvidia's stock. The plumbing usually cracks before the fixtures do.
High growth scenario and medium growth scenario (Graph 2). I feel like an idiot asking - aren't we missing some, or at least one, scenario? Is "medium growth" for the next 4 years really the worst people can think of?
Usefulness aside, I see little evidence AI is making money (profit, not revenue) for any firm whose profit doesn't come from the AI itself or the infrastructure, including supply chain. I'd love to hear a counterexample. One such example would be of a hypothetical company that does transla
Speaking of financing: how is the Anthropic IPO going, what is the timeline? They filed over a month ago, no news since. (I would have expected some spectacular news headlines that would be designed to fuel public interest in the impending IPO, but can't detect anything of substance)
At least if the datacenters usage crashes, we'll have cheap power from all the infra that got built.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
BIS released a larger report in June that identified AI financing/sustainability as one of the biggest risks for the global economy:https://www.bis.org/publ/arpdf/ar2026e.htm