This is how the AI bubble pops | Cory Doctorow

An hour-long interview on The Tech Report, recorded as Doctorow toured The Reverse Centaur’s Guide to Life After AI. His case is that the boom is a bubble driven by a fantasy of “a world without people”, and that it has to pop before it does lasting damage.

He starts with motive. Some billionaires, he says, have stopped seeing other people as real, and like the idea of replacing them. Others simply see a good sales pitch: a worker who never tells the boss an idea is impossible, illegal or dangerous. Either way the promise is a workplace where knowing how to do something no longer gives anyone power. On the technology he is a materialist who thinks a thinking machine is possible, just not this way. Today’s models find statistical correlations without understanding, so they cope badly with anything new, and each generation costs far more for a smaller step. His example is chess: a language model will move a piece onto a square it already occupies, a mistake the first valve-and-relay chess machines never made. Using one there is like clearing a driveway with a blowtorch.

The money is the core. He puts this year’s spending at about a trillion dollars against $50 billion of revenue. Inference costs aren’t falling, he argues, citing Ed Zitron’s leak of OpenAI’s accounts, where a marketing budget as large as Coca-Cola’s probably hides free inference. With almost no cost to switching models, the labs must keep training new ones, rebuilding the same railway on the same ground. He lists ways it could break: Nvidia lending money to the companies that lose it buying Nvidia’s chips, unstable founders, data-centre debt whose later lenders can walk away, and Gulf investors with war damage at home to repair.

On labour he takes two lessons from the Hollywood writers, the one group he says has beaten AI: say that your boss will give your job to an AI that can’t do it, not that the AI will take it, and organise, as the guilds’ multi-employer bargaining lets them. His biggest worry is process knowledge, what workers know that is never written down. Fire them for chatbots and switch the chatbots off, and rehiring won’t bring it back; he puts the cost at a generation. Afterwards he expects open-weight models to keep working and the specialised chips to be mostly scrap, and he tells governments to wait and buy them cheap. Finally he separates AI from enshittification: it is the attempt to invent a new market once squeezing the old one has run out, but also a very good squeezer, through surveillance pricing, wage discrimination and hallucinations as cover.

Some of his examples don’t hold as told. The nurse-pay apps he says set wages by credit-card debt come from a Roosevelt Institute report that says they “may” do so: a possibility, not something it documented. The MacBook Neo, his story of process knowledge, was reportedly built around binned iPhone chips Apple already had, not a foundry selling off a bad batch to a customer. DeepSeek is not a cleaned-up Meta model, and its roughly $6 million covered one training run’s computing; the one-day loss of about $590 billion in Nvidia’s value is right. The leak does show OpenAI’s sales and marketing at $5.73 billion, which Zitron compared with Coca-Cola’s, and Morgan Stanley’s forecast of about $3 trillion in data-centre spending through 2028 checks out. His aside that the Epstein files show Epstein’s margin call set off Lehman’s collapse found no support in mainstream reporting.