BitMEX co-founder Arthur Hayes, in his essay “Safety First”, says there’s a debt bubble forming around AI infrastructure. He estimates that the biggest AI labs’ demand for compute is backing over $1 trillion in investment-grade debt, plus hundreds of billions more in riskier obligations. If this setup cracks, Hayes thinks the US may have no choice but to pump more liquidity into the system—basically, turn the money printer back on.

Why AI Demand Isn’t Enough to Support the Debt

Hayes argues that when OpenAI, Anthropic, and SpaceX talk publicly about slowing down their AGI push, it’s not just about safety. The real issue, he says, is economics: there just isn’t enough paying demand for AI services at current prices, especially with Chinese models coming in way cheaper. Meanwhile, the AI labs themselves still aren’t profitable. If they cut back on training or get more efficient with compute, demand for data centers could drop, making it a lot harder to service all that debt tied to them.

Where the Risk Is Piling Up

According to Hayes, a lot of this vulnerability may have shifted into the US insurance industry via private equity funds and captive reinsurance schemes. He says the total value of these reinsurance assets is now $1.54 trillion. The big trigger, in his view, would be a downgrade of AI data center debt: insurers would then have to post real capital, and their captive reinsurance structures might not have the cash to cover it.

Two Scenarios for the US and the Liquidity Play

Hayes sees two ways the US government could respond. One: step in as the buyer of last resort for compute, guaranteeing demand for AI infrastructure in the name of national security and tech competition. Two: let the AI debt crash, then bail out the insurers stuck with toxic assets. Either way, Hayes says, the system will need more liquidity—and it’s not just the Fed’s money printer in play anymore.