AI infrastructure credit is repricing — who finances your tokens

Two things happened this week that look unrelated and aren't. Microsoft posted a blowout quarter and added roughly $450 billion in market value. Meanwhile the companies that actually rent out GPUs got taken apart. That divergence is the whole story: the market is separating AI revenue from AI infrastructure credit, and if you buy inference from anyone other than a hyperscaler, you are exposed to the second one.

What actually happened

On July 29, Nebius fell about 10% and CoreWeave about 9% in a single session, reported as a credit story rather than an earnings story — credit-default-swap costs on AI-infrastructure borrowers widened as investors questioned whether the buildout can be financed at tolerable rates. Over the preceding month the damage was heavier: Nebius down roughly 43%, CoreWeave roughly 36%, SanDisk over 55%.

Then on July 30, CNBC reported that Situational Awareness — the AI-thesis hedge fund launched by former OpenAI researcher Leopold Aschenbrenner, reported at roughly $20 billion — was forced to unwind its public equity positions after steep losses concentrated in exactly those names. Neither the fund nor its brokers have confirmed the details; treat the specifics as sourced reporting, not settled fact. The direction, though, is visible in the tape.

The structural context is the part worth internalizing. The Bank for International Settlements has documented how this buildout is financed: hyperscaler bond issuance topped $100 billion in 2025, and a growing share of the rest sits off balance sheet — special purpose entities and joint ventures with long-term operating leases and capacity commitments, with private credit funds holding the underlying debt. BIS calls these arrangements economically equivalent to debt while remaining largely invisible on corporate balance sheets, and flags circular financing as the risk.

Why AI infrastructure credit matters for your business

You don't buy GPUs. You buy tokens. But the price and availability of those tokens are downstream of whether someone can refinance a data center — and that someone increasingly isn't a company with Microsoft's balance sheet. When credit tightens for the neocloud layer, it shows up on your side as capacity limits, quiet rate-limit tightening, or a repriced contract at renewal.

Two practical moves. Know who actually runs your inference. Not your vendor — your vendor's provider. If your AI feature is served by a startup reselling a neocloud reselling leased capacity, that's three balance sheets between you and the silicon. Then keep the swap cheap. Route through an abstraction layer, keep prompts and evals portable, and make sure moving providers is a config change and a test run — not a rewrite. We build it that way by default, and the reason is exactly this week.

Contract length is the other lever. In a repricing market, short terms are worth more than the discount you'd get for signing long.

Sources: Bank for International Settlements, CNBC, CNBC on AI debt costs.