Swissquote asks who pays for the leases if the AI race slows down

The Philadelphia chip index fell 6 per cent while the Nasdaq 100 was down 1.7 per cent in early trading ^[6][5], about three and a half times as far ^[24]. Nvidia, which sells the chips, fell 3.5 per cent ^[7]. Lam Research, which sells the machines that make the chips, fell 8 per cent, and Applied Materials 7 per cent ^[8], more than double Nvidia's move ^[25]. Bloom Energy lost 8.9 per cent and GE Vernova 7.6 per cent ^[9], against drops of more than 1.4 per cent each at Meta and Amazon ^[10], a gap of up to about six percentage points ^[26].

Chip equipment and power suppliers sell against orders placed years ahead. According to The Straits Times, the scrutiny on AI sectors was triggered by companies increasingly relying on debt and circular financing to fund their ambitions while raising spending forecasts, at a time when global yields have reached multi-year highs ^[13]. Ipek Ozkardeskaya, a senior analyst at Swissquote, said: "If the AI race slows materially, the key question becomes: Who pays for all that infrastructure? The leases, debt and power commitments remain even if expected compute demand and revenue growth slow. And that could bring credit risk increasingly into the AI story" ^[14].

So far that is a price move and an analyst's question. The report describes no changed lease, breached covenant or renegotiated power contract ^[27]. Steve Sosnick, chief market analyst at Interactive Brokers, said: "If this does lead to sort of a slowdown and a rethink of AI spending, that will have ramifications for the economy and some important sectors of the stock market, because essentially, we've been running hot based on AI spending" ^[15].

The case for slowing sits outside the essay too. Anthropic published a threat intelligence report on Sept 10 describing actors who used its Claude models for activities including weapons development, cyber operations, surveillance and fraud ^[17]. Anthropic researcher Jacob Coxon resigned, stating that the "people building AI earnestly believe that it could kill us all by the end of the decade" ^[18].

For the person who has to answer for an AI line item, the exposure turns on two questions, and neither is about model capability. How long would it take to move the workload if the terms changed at renewal, counting the prompts, the evaluation sets and the data parked with that vendor. And how much of what your product does actually depends on that vendor. A dependency that is cheap to leave and light on value can be left alone. The ones that are expensive to leave and carry the product are where you want a notice period on price changes and a written commitment to keep serving the model version you built against, because a slower roadmap costs a customer very little and a supplier's interest bill costs them something.