AI is not a bubble - but model makers’ valuations are: David Richards
When I lived in the US in the late 1990s and early 2000s, running my first start-up in Palo Alto, it was fashionable to buy shares on margin, in simple terms with money borrowed from a broker.
Virtually everyone I knew was piling into the dotcoms like there was no tomorrow.
It became so popular that margin debt reached 2.8 per cent of GDP at its peak, and I watched it collapse with the dotcom bust. It turns out these peaks in margin debt to GDP are a key indicator of an impending crash, as in 2007 (pre-financial crisis) and again in 2021. In June margin debt hit a record $1.5 trillion, 4.6 per cent of American output, more than at the top of the dotcom boom or in 2007.
Twice I have argued here AI is not a bubble. I still do. The model makers’ valuations are the bubble. The borrowed money is the petrol that will feed the inferno.
On Saturday Dario Amodei, Anthropic’s chief executive, published an essay called We Must Pace the Frontier; Sam Altman agreed, saying OpenAI would not float this year. By Monday’s close the Philadelphia semiconductor index had lost almost 6 per cent, having begun the day a fifth below its June record; the ten-year Treasury touched 5 per cent for the first time since 2023.
Take the safety talk seriously. Then ask who benefits. Anthropic is reportedly weeks from a $2 trillion Nasdaq listing, and a company that asks the industry to slow while selling shares is asking rivals to slow where it leads.
Yann LeCun, Meta’s chief AI scientist until last year, has heard it before; in 2019 OpenAI, where Amodei ran research, called GPT-2 too dangerous to release in full, then released it nine months later. “I made fun of them then. Everyone should make fun of them now,” LeCun said. Real fear and a convenient message can sit in the same essay.
No patent on a large language model keeps a rival out. Google published the architecture in 2017, patents and all. Every major model since is built on it. China’s Moonshot and DeepSeek give their weights away, and Epoch AI puts the best of them four months behind the closed frontier. Last week Anthropic reported that Chinese labs had pulled nearly 200 million exchanges out of Claude to train their own models. There is no moat. Anthropic’s system card lists five kinds of training data, all scraped, bought, given or made in-house. That is a supply chain, not a patent.
Anthropic began buying secondhand books by the pallet in 2024, slicing off the spines and scanning the pages before pulping the rest, and last year agreed to pay authors $1.5 billion over books it pirated first. What is left to own is the data, and the biggest pile of it, the books, can be bought by anyone with a chequebook and a cutter.
Now the price. Two trillion dollars is about thirty times the $65 billion run-rate Anthropic reportedly reached in July, for a business whose raw material is bought and whose rivals are four months behind. A thirty-times-revenue listing into a 5 per cent bond yield tests what public investors will pay, and SpaceX, which floated in June and sat a fifth below its offer price within seven weeks, is the precedent. When the labs are repriced, and they will be, the suppliers follow. The big cloud companies have promised more than $700 billion of data-centre spending this year; on Monday most rose and the chipmakers fell.
So no, AI is not a bubble. The models will be here on Friday morning. UK businesses should use them.
The labs’ valuations are the bubble, the chipmakers are already correcting, and an October listing would put them to the test. Just do not confuse the technology with the tickers. Buy the tools. Be careful with the shares.
David Richards MBE is managing partner of Yorkshire AI Labs