Former BitMEX CEO Arthur Hayes has published an essay titled "Safety First" on the economics of AI and the risks associated with infrastructure financing. He links the "safety first" narrative to the high cost of computing, competition from China, and the potential for declining demand. According to Hayes, demand from leading US labs for computing power underpins over $1 trillion in investment-grade debt and hundreds of billions of dollars in riskier debt and loans. In the event of disruptions, Hayes believes US authorities would have to support either the AI infrastructure or holders of distressed debt, which in any case would increase dollar liquidity and could support the crypto market.

Chinese Models Are About 100 Times Cheaper, Market Chooses Price

Hayes notes that as Anthropic, OpenAI, and SpaceX call for slowing AGI development under the pretext of "safety," American solutions face competition: Chinese models are about one hundred times cheaper. Arguments about superior quality and the "distillation" of American models are gradually losing their appeal for buyers who are focused on the final cost of computational intelligence.

The Debt Chain Lies in Data Centers, Funds, and Insurers

The author examines how debt financing for data centers is connected to funds and insurance companies, warning that if demand for computing power weakens, asset valuations will change. In this scenario, the choice of government support between infrastructure and debt holders leads to increased dollar liquidity.

Labs Are Unprofitable and Rely on Tech Giants

According to Hayes, AI labs as a group are not profitable and rely on profitable technology companies—Nvidia, Broadcom, Google, and Microsoft—which effectively support debt financing for data center contracts and semiconductor purchases.

"Safety First" May Reduce Training Costs

If "safety first" becomes the guiding principle, training costs for models will not disappear but are likely to decrease, and companies will seek to improve computing efficiency, allowing clients to spend less on infrastructure. Hayes previously pointed to interest in the future S-1 form of Anthropic and questions about the economics of servicing a single token, but expects that answers to key metrics may not be forthcoming.