A second market for the AI trade's downside

The machines are now inside the loan. Nvidia is working with BlackRock, Apollo and Goldman Sachs Private Wealth to raise more than $500 billion against GPUs pledged as collateral — the chips themselves standing behind the debt ^[1] — while the same machines' rental rates are being standardized into indexes built to be traded. A year ago the market for AI's downside was a single hand-written trade. In November 2025, banks asked Saba Capital, for the first time, to write protection against the debt of Microsoft, Oracle, Meta, Amazon and Alphabet, priced by hand because no standardized market existed ^[2]. Sam Altman had already put the era's warning on the record.

someone’s gonna get burned there. — Sam Altman

By August, that market had gotten ahead of the stock market. The spread on the hyperscalers' credit-default swaps — the annual price of that insurance — jumped even while Nvidia's and Oracle's shares held near their highs ^[3]. Credit was repricing AI's debt risk while the stock market held its ground. And the repricing did not stay inside tech: BNP Paribas found the hyperscalers' borrowing spree had pushed insurance costs higher for unrelated blue chips like LVMH, Sanofi and BAE Systems, because every high-quality borrower now competes with the hyperscalers for the same capital ^[4]. Apollo's chief economist Torsten Slok put a number on it — hyperscaler insurance now runs roughly 60 basis points a year more than bank insurance, about 0.6 percentage points of annual cost that didn't exist a year earlier ^[5]. By October the market was building itself in every direction at once. Protection costs for Alphabet, Amazon, Meta, Microsoft and Oracle hit their highest levels since 2021 ^[6], and credit-default swaps began trading on several mega-cap companies for the first time, as debt raises put a default price on names that had always carried only cash ^[7]. Jump Trading opened an over-the-counter desk where data-center developers can hedge against falling chip-rental rates ^[6]. Its pitch was not a warning.

You don’t need to be bearish to want some insurance. — Jump Trading

Hyperbolic, run by a former Citadel Securities researcher, is building a marketplace with GPU futures that settle in the actual chips, and CME Group has a cash-settled futures contract tied to a Silicon Data index sitting under CFTC review ^[8]. The rental rate of a GPU — the buildout's raw input — is becoming a price you can bet on. Aon is assembling a $10 billion insurance pool because data centers packed with billions in equipment have grown hard to insure, and an uninsurable data center cannot borrow ^[9]. The insurer's chief executive put the stakes plainly.

If you create them and you build them on top of each other, you've created concentration risk in a way in which it makes it more difficult to get insurance. — Greg Case

The machine at the center of all this is now doing triple duty: collateral at the bottom of a half-trillion-dollar loan, the thing whose rental rate the new futures and desks price, and the asset the new insurance pools are being built around. None of it has broken the conviction underneath. Even the buyers stepping back in after the selloff take their cue from the credit market now: one portfolio manager redeploying cash into AI infrastructure stocks argued the bond market was signaling equities were near a bottom ^[10]. There is a defect in the new protection, though — unfinished plumbing rather than a punchline. Yahoo's five-year default insurance cost quadrupled over twelve months as it poured money into data centers ^[11]. But the insurance itself may not fire when it is needed: the liability-management deals and lease-backed structures spreading through AI debt may not trigger a credit-default swap payout the way an ordinary bankruptcy does — the same gap that left protection worthless at Altice France and Ardagh in 2024 ^[11]. The same Wall Street that spent a year building protection on the buildout is preparing to underwrite its biggest listing — Anthropic's IPO, valued at up to $2 trillion ^[12]. The insurance arrived first. The machinery that would pay it out is still being assembled.