Your GPU Cloud Runs on Subprime Debt

Ed Zitron's "The Subprime Data Center Crisis" hit the Hacker News front page this week with a thesis engineered to alarm: the AI buildout is 2008 with GPUs, and 80% of the money is going to zero. The essay is long, angry, and — on the structural facts — largely right. Bloomberg puts outstanding AI data-center debt north of $500 billion, with roughly $200 billion of it sitting in private credit funds, and issuance in 2026 alone has already passed $334 billion. The big platforms have doubled their own debt loads to around $350 billion while pushing the largest projects off their books entirely.

Where Zitron loses me is the analogy. This isn't 2008. It's 2001 — and that distinction changes what you, the person actually renting this compute, should do about it.

The machinery is real, and it's weirder than you think

The financing structure worth understanding is the special purpose vehicle. Meta's Hyperion campus in Louisiana isn't being built by Meta in any balance-sheet sense: a bankruptcy-remote entity called Beignet Investor LLC — 80% owned by funds managed by Blue Owl, 20% by Meta — priced $27.3 billion of senior secured notes last October at 6.581%, due 2049, anchored by PIMCO at roughly $18 billion. S&P rated it A+. Meta gets a frontier-scale data center; Meta's balance sheet gets a lease.

The neocloud tier is stranger still. Nvidia is contractually obligated to pay CoreWeave up to $6.3 billion through 2032 for any capacity CoreWeave can't sell — a backstop signed in 2023 but only disclosed in an SEC filing in October 2025. That guarantee is what lets CoreWeave borrow tens of billions to buy Nvidia GPUs. Nvidia's biggest customer category is increasingly companies whose creditworthiness Nvidia itself underwrites. CoreWeave's contracts with OpenAI now total around $22.4 billion; Microsoft signed Nebius for $17.4 billion. Nearly every large check in the neocloud economy traces back to two unprofitable AI labs and the chipmaker they buy from.

So yes: enormous leverage, circular guarantees, off-balance-sheet structures, and revenue assumptions that require OpenAI and Anthropic to keep raising money at historically unprecedented rates. All true, all verifiable.

Why 2001 fits better than 2008

The subprime framing needs contagion — risk hidden in instruments nobody can price, held by institutions nobody can let fail. That's not this. SPV debt is non-recourse by design; if Beignet's economics break, PIMCO's bondholders eat it, not Meta, and everyone buying these notes knows exactly what's in them. The repricing is already happening in the open: nearly 80% of data-center securities issued since early 2025 now trade at wider spreads than at issuance, and Amazon's recent $25 billion bond landed to visibly tepid demand. Markets that reprice gradually and publicly don't produce Lehman weekends.

What this actually rhymes with is the telecom buildout of 1998–2001. Carriers borrowed staggering sums to lay fiber against demand projections ("internet traffic doubles every 100 days") that turned out to be fiction. Global Crossing and WorldCom collapsed, creditors were wiped out, and most of that fiber stayed dark for years. And then the punchline: that stranded capacity became the cheap bandwidth that made YouTube, Netflix streaming, and the entire Web 2.0 era economically possible. The overbuild was a catastrophe for the people who financed it and a subsidy for everyone who came after.

GPUs complicate the parallel in one direction — fiber doesn't depreciate, and an H100 loses relevance in four or five years — but the market has already run this experiment once at small scale. Open-market H100 rental rates collapsed from roughly $8 an hour to under $2 between 2023 and late 2024 when supply caught up. That's what a compute glut looks like: not smoking craters, falling prices.

The risk you're actually carrying

If you build on this infrastructure, the practical exposure isn't systemic collapse. It's counterparty risk, and it's tiered.

Renting from AWS, Azure, or Google Cloud, you're fine — a hyperscaler eating losses on overbuilt capacity is a shareholder problem, not an uptime problem. Renting from a neocloud is different in kind. These are highly levered companies whose debt covenants include debt-service coverage ratios and liquidity minimums, often dependent on one or two anchor customers, holding collateral (GPUs) that depreciates faster than the loans against it amortize. If your provider hits a covenant breach, your prepaid credits are an unsecured claim in a restructuring and your training run is collateral damage.

Concretely:

• Don't lock long against weak counterparties. A three-year reserved contract with a neocloud is a bet on both GPU prices and that company's solvency. With spreads widening and spot prices structurally falling, shorter commitments with renegotiation windows are worth the premium.
• Keep exit costs near zero. Checkpoints and datasets in neutral object storage, not the provider's local disk tier. Containerized training stacks. No dependence on a proprietary scheduler you can't replicate on Kubernetes or Slurm elsewhere. This is cheap insurance against an event that now has meaningful probability.
• Read your provider like a lender would. Customer concentration above 50%, undisclosed backstops surfacing in SEC filings, debt raised against GPU collateral — these are the tells. CoreWeave discloses; smaller shops mostly don't, which is itself information.
• Position for the glut. If the overbuild resolves the way telecom did, 2027–2028 compute gets cheap. Teams that architected for portability get to arbitrage distressed capacity. Teams welded to one provider's stack get to renegotiate from weakness.

The bill comes due, unevenly

Zitron's claim that 80% of this debt goes to zero is a rhetorical flourish, not an analysis — nobody, including him, can price the residual value of a 2026 GPU fleet in 2030. But the softer version of his argument doesn't need it: hundreds of billions in credit is currently priced on the assumption that two money-losing labs can keep raising capital forever, and lenders have started to doubt it in public. Some neoclouds will not exist in three years. Their creditors — increasingly pension funds and insurance annuities routed through private credit — will take real losses.

Developers, though, sit on the other side of this trade. Every dollar of overbuilt, debt-financed capacity is a future subsidy for the people renting it — provided you're structured to walk away from a failing provider and walk into a cheap one. The fiber barons went bankrupt. The internet got fast anyway. Plan accordingly.