The AI data centre debt bubble: believe the sellers, not the sermons
Credit downgrades, defensive risk disclosures and a rush to offload data-centre stakes are telling a very different story from the keynote stage. Businesses signing AI contracts should read the actions, not the projections.
Masayoshi Son says AI will need $5 trillion of investment every single year by 2040, and calls bubble talk absurd. The people who actually built the data centres, meanwhile, are hiring bankers to sell them. When talk and behaviour diverge this sharply, back the behaviour. That divergence is the best evidence available on whether the AI data centre debt bubble is real, and it's far more useful to a business signing AI contracts than any conference-stage forecast.
Forecasts are cheap. Credit rating actions, audited risk disclosures and equity sales are expensive signals: someone pays a price to make them. So start there.
The most concentrated bet in modern credit
In July, S&P Global cut Oracle's credit rating to BBB-, one notch above junk, calling OpenAI a "central credit risk": roughly half of Oracle's $638bn in remaining performance obligations trace back to one customer. Oracle's own annual report says the same thing in lawyer, warning that some customers "may be highly leveraged" and that cloud revenue is concentrated in a handful of very large accounts. The bet is enormous. OpenAI's own announcement describes up to 4.5 gigawatts of capacity in a partnership exceeding $300bn over five years, and funding it produced negative free cash flow of $23.7bn in fiscal 2026, with capex up 162%.
Now look at the shape of the liabilities rather than the headline number, because that's where the fragility actually lives. Data-centre leases typically run fifteen to twenty years. The bonds funding them mature in five to ten. And the revenue meant to service both depends on an unprofitable private company's ability to keep raising fresh capital, year after year, in whatever market conditions happen to prevail. Borrow medium, commit long, collect short: a textbook duration mismatch, and duration mismatches are how solvent-looking institutions get into trouble fast. No forecast from a stage addresses that structure.
Is the AI data centre debt bubble following the fibre playbook?
We've run this experiment before. Between 1996 and 2001, telecoms carriers borrowed vast sums to lay fibre against demand forecasts, famously, that internet traffic doubled every hundred days, which turned out to be marketing. Demand kept growing; it just grew slower than the debt. The result was the telecoms crash, and the recoveries on network assets were brutal. Global Crossing had run up more than $12bn of debt building its network; a controlling 61.5% stake in the restructured company changed hands for $250m, roughly two cents on the dollar of the debt, for control.
The detail that should worry today's lenders is buried in the collateral. Fibre was, in hindsight, decent security: glass in the ground lasts decades, and the distressed buyers eventually did well as demand caught up, much of today's AI traffic runs over fibre lit long after its original owners went under. GPUs are not comparable assets. Operators book them on five- or six-year depreciation schedules, but the period in which anyone pays top rates to rent them looks closer to two or three, because each new chip generation resets the market price of the old one. If recoveries on long-life fibre came in at cents on the dollar, recoveries on rapidly obsolescing silicon start from somewhere worse.
That matters because GPU-backed lending has become an asset class in its own right. CoreWeave has raised billions in debt secured against its Nvidia fleet, and it's far from the only borrower pledging chips as collateral. The wrong-way risk is baked in: the moment a borrower can't service a GPU-backed loan is precisely the moment demand for renting GPUs has softened, which is also the moment the collateral is worth least. Fibre lenders learned this in 2001. The depreciation curve says the GPU version of the lesson arrives faster.
Why are the builders selling now?
Watch the sellers. US data-centre builders are reported to be racing to offload majority stakes worth tens of billions of dollars, largely to private-equity funds under structural pressure to file these assets under stable infrastructure, a categorisation Moody's has pushed back on, warning of overbuilding, obsolescence and refinancing risk.
The more interesting signal is quieter. Larry Ellison has pledged 346 million Oracle shares, roughly 30% of his stake, as loan collateral. Pledging is subtler than selling: you raise cash without a disposal hitting the tape, you keep the upside and the headlines, and if the shares fall far enough the bank absorbs them. There are innocent reasons to do it. But when a founder monetises a third of his position while his company borrows heavily against a single customer's promises, the innocent explanations have to work harder than usual.
Meanwhile, prices are moving against the borrowers. Every frontier-model generation has shipped cheaper per token than the last, and well-capitalised entrants keep undercutting the leaders. The lazy objection is that a price war only matters if it destroys an incumbent. It doesn't need to, it only needs to disperse growth. The debt schedules assume demand compounds at something like today's prices; a borrower who planned for 100% annual growth and delivers 40% is still growing, and still insolvent on schedule.
What should a business signing AI contracts do now?
If you're buying AI services in the UK, none of this is somebody else's problem. Britain has comparatively little frontier compute of its own, punishing industrial energy prices make domestic capacity expensive to build, so UK AI capability is overwhelmingly imported, contracted through the local subsidiaries of providers whose obligations sit on the balance sheets described above. A credit event in Texas surfaces in your renewal terms in London.
So discount forward projections heavily and weight the observables: credit actions, filed risk factors, free cash flow, insider behaviour. On that basis, sensible hygiene looks like shorter contract terms, genuine model portability and multi-vendor architectures that survive a supplier's distress, the kind of groundwork we cover in AI readiness before you build. If your AI supply chain runs through a counterparty whose lender just got downgraded for exposure to it, that belongs in your risk register, and a technical strategy review should surface it before procurement signs anything.
What would change my mind? Audited OpenAI revenue growing faster than its commitments. S&P revising Oracle's outlook upward. Builders keeping their equity instead of selling it. GPU resale values holding their ground through the next chip generation. Any of those would be genuine evidence the structure is sturdier than it looks, and I'd update accordingly.
Until then, notice the asymmetry nobody has priced. If the commitments get paid, the upside is already in the valuations. If they don't, the downside cascades through Oracle's balance sheet, private equity's newest asset class and every lender holding depreciating silicon as security. For a business buyer, that asymmetry is close to a free option: contract flexibility costs a little margin today, while lock-in to a fragile counterparty could cost you your AI capability tomorrow. Take the option.
Questions people ask
How do I assess an AI vendor's financial stability before signing a multi-year contract?
Read the expensive signals, not the marketing: credit rating actions and their stated rationale, the risk-factors section of annual filings, free cash flow trends, customer concentration, and whether insiders are selling or pledging shares. A vendor's forward projections are advertising; its regulatory filings are sworn statements. If a supplier's own filings warn about customer non-payment, plan your exit clauses accordingly.
Are GPU-backed loans riskier than other infrastructure lending?
They carry a specific wrong-way risk: the collateral's resale value depends on the same demand cycle as the borrower's revenue, so the chips are worth least at exactly the moment the loan goes bad. The 2000–02 telecoms crash produced recoveries of cents on the dollar on fibre with a multi-decade useful life; GPUs are repriced by every new chip generation, so distressed recoveries would likely start lower. Watch second-hand GPU prices as an early-warning indicator.
Why does it matter that data centre owners are selling majority stakes?
Because the sellers have the best information. Builders and operators know their assets' economics better than any buyer, and choosing to monetise majority positions while the narrative is at its peak is a revealed-preference signal. It doesn't prove a top, but it shifts the probabilities, especially when the buyers are private-equity funds under structural pressure to deploy capital into anything labelled infrastructure.
Written by an AI editorial persona of Abyshire's proprietary editorial system and reviewed by our team.