Frontier AI is getting rejected by businesses | Dean Baker

A 30-minute interview on The Tech Report with Dean Baker, the Center for Economic and Policy Research economist known for warning early about the housing bubble. Host Isaac Pound opens on CNBC’s report from Salesforce’s Dreamforce conference, where business customers said last year’s models are enough. Baker calls that the industry’s real problem, bigger than any argument about superintelligence. If lesser models can do nearly everything a business wants, who pays the hundreds of billions, and soon trillions, that the labs and hyperscalers have committed to?

He runs rough numbers. Meta’s new Muse app is free; even five million people paying $20 a month would bring in about $1.2 billion a year, against obligations in the hundreds of billions. Credit markets are noticing, he says, citing spreads on the hyperscalers’ credit default swaps up about 60 basis points in a year. He expects them to scale back data-centre plans quietly, blaming local opposition rather than announcing cuts. And the labs can’t force upgrades by retiring old models, the way a near-monopoly like Microsoft once could. Open-weight models give businesses privacy, control and protection against poisoned training data, and the small group of heavy users who, the host says, account for most of the labs’ demand are the ones best placed to switch.

Two further claims. A serious but survivable AI failure, in air traffic control or a bank’s ledger, is “a million times” likelier than extinction and would sink the market; that, he speculates while saying he has no inside knowledge, may be why Anthropic wants to go public soon. And the bubble lasts because fund managers face a one-way bet: lose along with everyone else and nobody is fired. He admits timing is the hard part. He moved his retirement savings out of stocks in March 1998, two years early, and first wrote about the housing bubble in 2002, four years before prices peaked.

The checking mostly holds up, with some figures needing care. The Dreamforce story is real (CNBC, 18 September 2026), and its main quoted source says most agent work runs on “last year’s AI”. The 60-basis-point figure matches a note from Apollo’s Torsten Slok, but his comparison is hyperscaler spreads against bank spreads since October 2025, not a simple rise over a year. Baker’s housing paper is dated August 2002. The near-miss he describes happened: an analyst at US Special Operations Command Pacific used a chatbot that wrongly identified a Chinese ship’s cargo as nuclear-weapons parts, and the boarding was called off once experts checked. Some of the host’s figures are off. Ramp’s August index put Anthropic’s top model, Fable 5, at 6% of tokens and 11.4% of spend, priced at $10 per million input tokens and $50 per million output. On 22 September OpenAI halved its prices with two GPT-6 models, but Anthropic’s cut on Opus 5.5 was 20%. And the February order on Claude gave federal agencies six months to phase it out; it did not shut it down within 90 minutes.