BIS and Oracle Filings Amplify Warnings on AI Capex Bubble

It is one thing when short sellers call an AI bubble. It is another thing when the Bank for International Settlements, the Basel institution that coordinates the world's central banks, puts the same warning in its annual report, and the hyperscaler underwriting the largest single AI infrastructure bet in the world spells out the downside in its own SEC filing. That is the pileup The Register is drawing attention to this week.

The BIS report, released at the end of June, puts the current AI capex cycle in the same lineage as the British railway mania of the 1800s, the canal investment bubbles, and the dot-com crash. The through-line the report draws is that each of those episodes "attracted a lot more capital than the resulting industry could actually produce." The numbers behind the concern are not small: combined capital expenditure of the world's five largest hyperscalers is expected to exceed $1 trillion from 2025 through the end of 2026, outpacing earnings and free cash flow to the point where some are issuing debt to close the gap. The BIS itself warns that "disappointment in returns could trigger a sudden pullback in financing and turn the capex boom into a protracted investment bust."

The Oracle piece of the story is where the abstract macro warning turns concrete. Oracle signed up as one of the primary financiers of Stargate, the roughly $500 billion OpenAI data center build-out, to the tune of about $300 billion over an indeterminate period. Its latest SEC filing, as The Register previously detailed, now explicitly flags customer non-payment and non-performance as a risk, and states that "if customers do not renew their contracts, we may be unable to re-lease, repurpose or assign such capacity on acceptable terms, if at all." It also cites the challenge of securing reliable and cost-effective power for AI data centers as demand outstrips supply. OpenAI is not named in those risk factors, but it is the counterparty the market is pricing.

The honest caveat is that neither the BIS nor Oracle is calling a crash. The BIS is describing a pattern and a tail risk; Oracle is doing what public companies are required to do, which is disclose downside scenarios. The reporting also does not give you the covenants inside the Stargate lease structure, the exact split of hyperscaler capex between debt and cash flow, or which other providers carry OpenAI-like single-customer concentration. Take the specifics as reported, not settled.

What is worth watching from here is who gets rewarded for restraint. Enterprises running disciplined AI budgets, teams betting on smaller open-source models, and credit analysts who can price hyperscaler exposure honestly all look better positioned if the BIS scenario even partly plays out. The interesting question for the second half of 2026 is not whether AI is real, but whether the balance sheets underwriting it can survive a bad quarter of revenue expectations.