BIS dares to blaspheme as AI bubble fears wane: Mike Dolan

The BIS Breaks a Taboo

It has become almost heretical in markets to question whether artificial intelligence stocks are inflating a speculative bubble. Yet the Bank for International Settlements — the central bank to central banks — has done exactly that, warning that the AI investment frenzy carries echoes of past financial manias even as investor anxiety about a bubble appears to be fading. Mike Dolan's column captures a striking divergence: the institution tasked with safeguarding global financial stability is growing more cautious just as the market grows more complacent.

Why This Matters Beyond Wall Street

The AI boom is not merely a stock-market story; it is underwriting one of the largest capital-spending cycles in modern industrial history. Hyperscalers and chipmakers are pouring hundreds of billions of dollars into datacenter buildouts, custom silicon, and specialized inference hardware. If the BIS's caution is warranted, the consequences would ripple far beyond equity valuations and into the physical infrastructure of AI itself — the chip fabs, the power-hungry datacenters, and the custom accelerator programs that hyperscalers have bet their futures on.

A bubble bursting in AI equities wouldn't just erase paper wealth; it could chill the financing environment for the next generation of AI chips, including custom silicon like Google's TPUs and similar in-house accelerators being developed by Amazon, Meta, and Microsoft. These projects require years of committed capital before returns materialize. A sudden repricing of AI optimism could force a slowdown in capex guidance, delaying next-generation datacenter expansions and squeezing the semiconductor supply chain that has enjoyed unprecedented demand visibility.

The Inference Cost Angle

One underappreciated thread here is inference economics. Much of the market's AI enthusiasm has been built on training-model breakthroughs, but the real commercial test is whether inference — running these models at scale for paying customers — can be delivered profitably. Custom silicon and specialized inference hardware exist precisely to bring these costs down. If investor sentiment cools and capital becomes less patient, companies may face pressure to prove near-term returns on inference infrastructure sooner rather than later, exposing whether current pricing models for AI services are sustainable.

Context and Caution

History offers ample precedent: the dot-com buildout of fiber-optic and server capacity in the late 1990s proved genuinely useful eventually, but not before a painful correction wiped out overleveraged investors. The BIS's intervention should be read less as a prediction of imminent collapse and more as an institutional nudge — a reminder that infrastructure spending divorced from demonstrated revenue can outrun fundamentals. For an industry now defined by trillion-dollar datacenter ambitions, that's a warning worth taking seriously, even if the market isn't yet listening.