IMF Warns AI Chip Debt Outlasts the Hardware

The IMF's top markets official just told central bankers the AI trade's real risk isn't Nvidia's price-to-earnings ratio — it's hyperscalers issuing decade-long bonds to buy chips that may be scrap metal in three years.

Why June 30 mattered:

• Tobias Adrian, director of the IMF's Monetary and Capital Markets Department, spoke at the ECB's annual forum in Sintra, Portugal — the same venue where the Bank of England warned about agentic trading meltdowns that same day.
• Adrian said borrowing tied to artificial intelligence is emerging as a "larger financial stability concern than high technology stock prices."
• The warning followed the Bank for International Settlements' June 28 annual report naming AI exuberance among four pressure points that "demand attention."

The maturity mismatch:

• Adrian said that from a financial stability perspective, "what is quite worrisome … is that the major tech firms are starting to leverage up themselves."
• "Chips can age 'fairly quickly' while firms are also issuing medium or long-term debt," he added. "If the assets stop earning money, the debt stays."
• AI GPUs can become obsolete within a few years as new hardware ships. Corporate bonds financing those purchases often stretch five to ten years.
• That is a classic asset-liability mismatch — the kind that turns an earnings miss into a refinancing crisis.

The borrowing binge, by the numbers:

• BNP Paribas now estimates 2026 hyperscaler capex at $725 billion — nearly double the mid-2025 estimate — with spending rising faster than operating cash flow.
• Reuters reported on June 29 that Amazon and Alphabet alone issued roughly $60 billion in bonds across multiple currencies in the prior 12 months as they tap every market to avoid U.S. saturation.
• Morgan Stanley told Reuters AI-related issuance could push investment-grade bond sales above $2 trillion in 2026 for the first time ever.
• Barclays data cited by Reuters put AI-related debt at roughly 15% of recent U.S. investment-grade issuance — a share that keeps climbing.

Not a bubble — yet:

• Adrian pushed back on the valuation-bubble narrative. "Valuations are very much related to price-earnings ratios," he said, "and so it's both the prices that have come off a little bit, but it's earnings that continue to surprise on the upside as well."
• He called current dynamics "really quite different from sort of bubble-like behavior."
• The IMF is not calling a top. It is calling a financing structure that only works if AI earnings keep beating expectations.

What breaks the trade:

• Adrian's test is profitability, not price. "It's really the underlying profitability that is the key financial stability issue at this point," he said.
• "As long as profitability continues, as long as corporates and individuals are paying … for the frontier-model costs, that is probably fine," he explained. "But at some point, of course there could be a disappointment to earnings."
• If returns disappoint, leveraged hyperscalers face the same choice every infrastructure boom eventually faces: refinance, restructure, or cut capex — each option rippling through credit markets that have never absorbed this much AI paper.

Why Sintra keeps piling on:

• Basel's central bankers flagged circular financing and shadow leverage on June 28. The Bank of England flagged autonomous agents on June 30. Fitch warned on July 1 that AI labor displacement could erode Western tax bases.
• Adrian's contribution is narrower and sharper: the debt outlives the hardware.
• That is a different failure mode than a stock bubble — and one bond investors, not equity traders, will feel first.

Convina's view: Wall Street spent June debating whether OpenAI deserves a trillion-dollar IPO. The IMF spent June 30 asking whether a Blackwell cluster bought with 2031 bonds will still be earning its keep in 2028. Those are not the same conversation. Adrian is right that earnings have held for now — but strong quarterly prints do not shrink the maturity gap. Hyperscalers are betting Moore's Law will keep revenue ahead of depreciation while they lever balance sheets to levels telecom companies used right before the bust. Until bond covenants price chip obsolescence the way insurers price hurricane seasons, the bull case and the systemic-risk case remain one trade. Adrian just named which leg breaks first.