Wall Street keeps buying AI chips. It has stopped lending against them.
The debt behind the AI buildout is still secured on paper by its chips, but the people writing the newest loans have stopped trusting that collateral — and the risk the professionals refuse keeps flowing down a chain to whoever will take it for the yield.
Since August, the investors who buy commercial mortgage-backed securities — bundles of loans on office towers, warehouses, and now server farms — have been turning down data-center debt in the middle of an AI boom. The bond desks have a name for the refusal: the "Luddite trade." [1] Wellington Management and other big buyers are steering clear of the sector, citing "property valuation uncertainties and potential supply gluts." Two major offerings, from CyrusOne and QTS Realty, had to widen their pricing to get sold, and Pure Data Centres scrapped a planned €1 billion Nordic bond in favor of bank loans. Some issuers now borrow repayment schedules from asset-backed securities, written specifically to survive the servers going obsolete. [1] In September, Ricardo Semler put the fear in a single image: a building that lasts decades, wrapped around servers that go obsolete in three to fifteen years. [2] Financial analyst Jonathan Weil made the point on October 5 that this does not stop the money. The end of the boom may show up first in the collateral itself — if chip resale values, utilization, and rental rates fall together — but capital can keep flowing even after that collateral becomes hard to value. [3] The newest loans already look that way. Nvidia set up a platform with BlackRock and Goldman Sachs to attract over $500 billion in outside capital while backstopping $125 billion of chip-backed debt. Amazon pitched a vehicle holding $8 billion of chips to borrow against and lease back. Broadcom extended up to $42 billion in convertible notes — debt that can convert into shares — to Anthropic to fund its chips. Analysts read all three the same way: the debt leans on the parent company's credit rating rather than the resale value of the hardware, so a loan can outlive the chips it is nominally secured by. [4] Apollo's chief economist, Torsten Slok, has put a number on how far this has gone. AI-related borrowing already exceeds 40 percent of all new long-term investment-grade corporate debt, he estimates, and the buildout could need over $2 trillion with a funding gap of roughly $1 trillion. [5] Into that gap steps money chosen for appetite more than information. Wealthy investors are rotating out of private credit and into infrastructure finance, drawn by hard assets and sustained demand for data centers — and observers note the positions are illiquid, with buyers pricing in little of the exit risk. [6] Aware Super's chief investment officer, Simon Warner, described the exposure his own fund holds in terms he would clearly rather not be using. [7]
There's been some instances in the last six months or so where that's softened a bit, there's more circular financing, a bit more conduit financing. Nothing that flashes red, but things that certainly flash orange. — Simon Warner
In six months, Warner said, the industry moved from funding data centers out of stable retained earnings to circular, conduit financing — and if one of the pillars stumbled, there could be a correction. [7] At the stated end of the chain sit the labs, asking the government to become a last-resort insurer. OpenAI's Sam Altman has proposed the government carry the industry's economic downside as insurer of last resort, and OpenAI's finance chief, Sarah Friar, has asked for a government backstop or guarantee to get the financing done. The request is proposed, not accepted; President Trump dismissed AI safety concerns and said that whoever wins AI wins. [8] None of this is a bet on an imminent fall. Data-center vacancy hit record lows this summer — Northern Virginia at 0.3 percent — with record new inventory taken up just as quickly as it arrived. [9] GPU prices, propped up this year by a memory shortage, are not collapsing. [10] Even Marathon's Bruce Richards, one of the sector's louder skeptics, insists the technology trade is just fine. [11] The evidence here is the structures being written, not a decline already in the spot data. Read end to end, the chain shows its two ends marking the same chips down while the money between them keeps moving. The bond desk prices its doubt in wider spreads; the lab CEO prices his in a request for a last-resort insurer. The handoff to the state has not been taken up, so for now the chain rests where the yield chose it — with the holders who arrived last.
- 1. CMBS Investors Resist Data Center Debt Amid AI Surge
- 2. Ricardo Semler Warns AI Data Center Boom Risks Obsolete Infrastructure
- 3. Jonathan Weil Warns Funding Gap Could End AI Boom
- 4. Tech Giants Use Complex Financing to Fund AI Chips
- 5. Apollo Economist Warns of $1 Trillion AI Funding Gap
- 6. Investors Shift Capital From Private Credit To Infrastructure Finance
- 7. Aware Super CIO Warns of AI Financing Risks for 2026
- 8. AI Executives Seek Government Financial Guarantees to Sustain Growth
- 9. U.S. Data Center Vacancy Hits Record Lows Amid AI Boom
- 10. AI-Driven Memory Shortage Spikes GPU Prices and Cuts Models
- 11. Bruce Richards Bullish on Technology Infrastructure Trade