The BIS used its 2026 annual report to formally flag AI infrastructure investment as a systemic risk, citing hyperscaler capex tracking above $400B in 2026 versus $230B in 2024. It points to circular financing (Nvidia→CoreWeave→OpenAI→Microsoft→Nvidia), extended GPU depreciation schedules (4 to 6 years), and concentrated customer risk among money-losing AI labs as structural parallels to the late-1990s telecom buildout.
The editorial highlights that hyperscalers quietly extended assumed GPU useful life from 4 to 6 years, which flatters near-term earnings but defers the recognition of impairment risk. If AI demand softens or chip generations turn over faster than the new schedules assume, the writedowns will hit later and harder than current financials suggest.
The synthesis emphasizes the circular flow — Nvidia invests in CoreWeave, CoreWeave buys Nvidia chips, OpenAI contracts CoreWeave for compute, Microsoft funds OpenAI, Microsoft buys Nvidia chips — as the structural pattern most reminiscent of 1999 telecom vendor financing. The BIS stops short of calling it fraud, but the language is pointed enough to signal regulatory unease about whether reported revenue reflects genuine end-demand.
The BIS flags that OpenAI, Anthropic, and xAI absorb a disproportionate share of GPU offtake while losing tens of billions per year. That means the entire $400B+ capex cycle ultimately rests on a handful of cash-burning customers whose ability to keep buying compute depends on continued private funding rounds rather than profitable operations.
The Bank for International Settlements — the Basel-based institution that serves as the central bank for central banks — used its 2026 annual report to do something it rarely does: name a specific sector as a systemic risk. The BIS warned that the scale and financing structure of AI infrastructure investment now resembles the late-1990s telecom buildout closely enough that regulators should treat a sharp correction as a plausible scenario, not a tail risk.
The specifics, per the Telegraph's reporting and the BIS document itself: hyperscaler capital expenditure is tracking above $400 billion for 2026, up from roughly $230 billion in 2024. Microsoft, Google, Meta, and Amazon together now spend more on data centers and chips than the entire US federal R&D budget. A growing share of that spending — the BIS doesn't put a precise number on it but flags the trend — is financed via private credit, special-purpose vehicles, and vendor financing arrangements where chip suppliers (read: Nvidia) take equity or revenue-share positions in the buyers of their own product.
The report singles out three structural concerns. First, depreciation schedules: hyperscalers have quietly extended the assumed useful life of GPUs from 4 to 6 years, which flatters near-term earnings but pushes the writedown risk further out. Second, customer concentration: a handful of AI labs (OpenAI, Anthropic, xAI) account for a disproportionate share of GPU offtake, and most of those labs are themselves losing money at a rate measured in tens of billions per year. Third, the circular financing pattern — Nvidia invests in CoreWeave, CoreWeave buys Nvidia chips, OpenAI signs a compute contract with CoreWeave, Microsoft invests in OpenAI, Microsoft buys Nvidia chips — which the BIS describes in language careful enough to avoid the word 'fraud' but pointed enough that anyone who lived through 1999 will recognize it.
The dot-com analogy is overused, so it's worth being precise about which part the BIS is invoking. They are not saying the technology is fake. They are saying the financing is recursive, the depreciation assumptions are optimistic, and the revenue required to justify current capex doesn't yet exist at the customer level. That is exactly the WorldCom/Global Crossing problem — the fiber was real, it just got laid faster than demand could absorb it, and the debt that funded it didn't care.
The numbers underneath are the part that should make a senior engineer pause. OpenAI's annualized revenue is somewhere around $12-15B depending on whose leak you trust. Their compute commitments, between Microsoft, Oracle, and the announced Stargate buildout, are in the hundreds of billions over the next five years. The arithmetic only works if either (a) ARR grows roughly 10x in 24 months, (b) inference costs collapse by another order of magnitude, or (c) someone keeps writing checks. Historically, when an industry needs all three of those simultaneously, the third one runs out first.
Community reaction on HN (102 points, 340+ comments at time of writing) split predictably. The bull case: this is different because the underlying compute *is* being used, model capability *is* improving, and the productivity gains *will* eventually justify the spend. The bear case: every bubble in history has a true believer cohort whose technical analysis is correct and whose financial timing is catastrophic. The most-upvoted comment, from a user claiming to work in datacenter finance, was blunter: 'The PPA contracts being signed for 2027-2030 power delivery assume utilization rates that have literally never been observed in any compute market. We are pricing perfect futures.'
What's genuinely new in the BIS framing — and what separates this from the usual 'bubble' takes — is the focus on cross-border financial plumbing. A meaningful slice of AI infrastructure debt is held by European and Asian institutional investors reaching for yield, often via structured products that bundle data center leases, GPU residual values, and PPA contracts. If GPU resale values crater (the H100 secondary market has already softened) or if a hyperscaler quietly slows its capex commit, those structures unwind in ways that don't stay contained to Silicon Valley balance sheets.
For practitioners, the BIS report is not actionable in the 'rewrite your code' sense. It is actionable in the 'understand the ground you're standing on' sense. If your employer's runway is denominated in Anthropic credits, Azure OpenAI commits, or 'AI transformation' headcount approved by a CFO who has read exactly one McKinsey deck, you have macro exposure whether you wanted it or not.
Three concrete implications. First, contract terms matter more than they did 18 months ago. The reserved capacity deals that hyperscalers and AI labs were pushing through 2024-2025 — multi-year compute commits at locked prices — were structured to give the seller revenue visibility. In a downturn, the buyer (you) is the one stuck paying for GPUs you no longer need at prices that no longer reflect spot. If you signed one, read the termination clauses. If you're about to sign one, negotiate them.
Second, model portability is no longer just an ideological preference. Teams that built on a single frontier model with no abstraction layer are betting that the provider stays solvent and stays priced where it is today. Teams that invested in a thin routing layer (LiteLLM, OpenRouter, or rolled their own) bought themselves optionality that costs maybe two weeks of engineering and may turn out to be the cheapest insurance they ever wrote. The open-weight tier — Llama, Qwen, Mistral, DeepSeek — exists partly because the people running those projects believe what the BIS just said out loud.
Third, if you are a contractor or consultant whose pipeline is 'help company X implement AI', notice that procurement cycles for discretionary AI spend are already lengthening. The first sign of a capex correction is not bankruptcy; it's a budget that was committed in Q1 quietly getting deferred to Q3. That's happening now, in pockets, regardless of what the BIS does or doesn't warn about.
The BIS does not move markets. It moves the Overton window for what regulators are allowed to say. Expect bank stress tests in the EU and UK to start incorporating AI-capex exposure scenarios within 12 months, and expect at least one major US hyperscaler to quietly revise its GPU depreciation schedule back down before year-end as auditors get nervous. The technology will keep working. The question — the only question that ever matters in these cycles — is who is holding the paper when the music slows down. Right now the answer is: a lot more people than realize it.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.