There's lots of reasons to believe the "hyperscaler" model of AI can never be profitable. This business model pursued by OpenAI, Anthropic, Google, Meta and others of spending billions on chips and data centres in the hope of recouping it through subscriptions is delusional because of its gigantic expenditures and negative unit economics (the companies lose money with every new customer and every new use, and they lose more money with each generation of their products).
The industry strenuously denies this, of course. They insist that they are only days away from turning their balance-sheets right side up. All they have to do is fix those unit economics, then they can make back the cost of producing their models by selling access to them. The problem is that the evidence for those improving unit economics is weak, while the evidence that they're faking their finances is very strong.
Same goes for the claims that these companies are already profitable. Dig into those claims and you'll learn they depend on a new, special meaning of "profitable" that does not match the generally accepted accounting procedures (GAAP) definition, which is to say, these companies are claiming that they are so cool that their profitability can only be measured using a novel, secret form of mathematics.
But let's give the AI bosses a momentary benefit of the doubt and stipulate that they are on the verge of acquiring positive unit economics, which will let them start to pay off the massive expenditures they incurred by training their models and enter their long-anticipated profitability phase, when the money furnaces they've been running for years turn into money printers, to the delight of the investors who've supplied the vast bales of $100 bills the companies have been shovelling into their models' coalboxes for years now.
Basically, they're saying: "Sure, it cost us a lot to get these rails laid, but now that the railroad is complete we can start running carriages over them and make a profit." Unfortunately (for bosses and investors), this proposition is every bit as dubious as their claims to improving unit economics.
To understand why, just look at what happened the last time Anthropic shipped a major Claude update. Virtually overnight, scores of OpenAI's best customers stopped paying for ChatGPT and started paying for Claude. That's because chatbots have very low switching costs: going from one chatbot to another costs almost nothing:
Everyone using AI knows this to be true. When I walked the floor at the Consumer Electronics Show in Las Vegas last year, I asked every AI-powered gadget maker: "What will you do if your chatbot provider jacks up their prices?" and to a one, they said: "No problem, we've designed this thing so that we can switch chatbots with the click of a mouse."
That means that you can't just "build the railroad and run the carriages over it". The minute you finish your railroad, your rivals will announce that they've got a new, adjacent railroad that's even faster than yours, and you will have to get to work laying another set of tracks to support even faster trains.
This is a disaster all around: the AI companies are locked in a Red Queen's race, a fatal beggar-thy-neighbor doom-loop. The only way they could escape that trap is by signing a non-aggression pact among themselves promising not to compete any more. But there's two giant problems with this: first, it is incredibly, fantastically illegal under antitrust law, because it represents a conspiracy among the dominant players to cease to compete with one another; and, second, it leaves the field open for the further development of Chinese "open weight" models that customers can run on their own modest, low-powered computers, which are presently lagging the US "frontier models" by a mere four months.
Freaking out about superintelligence is a canonical example of ‘criti-hype’, where critics repeat boosters' claims but append ‘(and that's bad)’ to them
Even if you don't trust Chinese models, you can extract their training through a process called distillation and transfer them to models you do trust.
But what if there was a way for the AI companies to get government permission to violate antitrust law and cease to compete with one another, and secure a ban on the use of Chinese open weight models? Turns out, there is a way to call time on the Red Queen's race: merely insist that you are on the verge of teaching so many words to the word-guessing program that it will wake up and devour us all, and call for a ban on "superintelligence".
Once the government stipulates that "superintelligence risk" is an existential crisis, it must grant the hyperscalers a consent decree absolving them from any violations of antitrust law stemming from a conspiracy to halt direct competition with one another.
Freaking out about "superintelligence" is a canonical example of "criti-hype", where critics repeat boosters' claims but append "(and that's bad)" to them.
Remember, the tech giants want to stop competing. Mark Zuckerberg and Sundar Pichai colluded to rig the ad market with a secret program called Jedi Blue. Every year, Google sends Apple a bribe of more than $20bn in exchange for Apple not entering the search market. And the biggest tech companies in the world had a secret "no poach" agreement where they illegally promised not to try to hire one another's top engineers by offering them raises.
The only thing Peter Thiel hates more than the Antichrist (spoiler, he's just talking about Greta Thunberg) is "wasteful competition".
When an industry that is eating itself alive through hyperscaling demands that the government bless a conspiracy to halt competition and ban open-source alternatives, you should be suspicious. When that industry is pursuing a venture that has lost more money than any other venture in human history, you should be very suspicious, especially when its "rogue AI hacking" story turns out to be a story about how a hacking tool did exactly what it was designed to do.
Peter Thiel is right: AI is full of wasteful competition, but not because competition is a waste – rather, it's because the companies are competing to convince people to use their expensive products for the cheapest applications.
Elon Musk's SpaceX IPO is indirectly propped up by X engagement metrics generated by letting the world's stupidest chuds produce mountains of child porn and images of Sonic the Hedgehog with giant boobs using Grok. That is indeed wasteful (and reprehensible).
That doesn't mean we should allow the AI companies to get the government to bless their conspiracy in restraint of trade; rather, it militates for having the government investigate them for securities fraud, trafficking in child sex abuse material, election finance violations, and a long list of other crimes and misdemeanours.
This is an edited version of a post from pluralistic.net. It is published under a CC BY 4.0 creative commons licence
Cory Doctorow, who was born in Toronto and now lives in Los Angeles and London, is the Nerve’s tech columnist. His new book, The Reverse Centaur's Guide to Life After AI, is published by Verso