Large language models share an open secret: they’re trained on the same data and generally detect the same patterns. Engineers race to produce updated versions only for their competitors to catch up to them within a few weeks. This makes for high production costs but low switching costs for customers.
It also makes for a fragile business model.
This logic explains why many fear a bubble. A site called isaiprofitable.com adds up major AI companies’ announced revenue and subtracts expenditure — and comes up with a large negative number. AI companies have excited us with promises of revenue, overlooking whether those exceed the cost of producing them. Policymakers, in Europe and elsewhere, should be wary of pouring public money into gigafactories or other schemes to build copycat large language models just as investors are beginning to question the likelihood of them making profits.
When ChatGPT 3.5 came out, OpenAI’s strategy looked like a classic three-stage capital expenditure play:
- Build up a monopoly too expensive to replicate.
- Raise prices.
- Profit.
Even if AI models turn out to be as useful as their proponents claim, this leaves us with a problem. AI chatbots are the supplier of an expensive, but largely undifferentiated commodity.
Although models trained on proprietary datasets (for example, a pharma company’s bank of proteins) might be able to differentiate themselves, the large language models have not. Despite trillions of dollars of investment, they are all based on the same underlying technology and the same data.
Worse, from the point of view of a monopolist, a competitor can release the weights in their model, meaning nobody else needs to start from scratch. Imagine if you could download a model car factory, run it through a 3D printer (admittedly, a big and expensive one), and start making Mercedes. The car industry would look different.
AI model developers are also squeezed to obtain critical supplies. They need to spend their money on chips. Semiconductor manufacturers are flourishing, so well that they have been funding their own customers to buy their products.
Get the Latest
Sign up to receive regular Bandwidth emails and stay informed about CEPA’s work.
Society benefits, at least for now. AI code has become trivially cheap. One million tokens from one of the more advanced open-weight models currently costs me $0.60. Manufacturers can differentiate themselves by writing different “harnesses” (programs to help AI models integrate their output into structured tasks, like software development), but so can anyone with access to the models’ API programming interfaces.
As chipmakers and end-users capture the bulk of the value, AI model developers are left with little. This will not make the returns that will satisfy venture capitalists.
Circular financing by chipmakers is not completely mad: if it pumps up the prospects of an AI lab, they can attract further investors, whose money will buy more chips. Suppose NVIDIA invests $50 billion in OpenAI, and this allows OpenAI to attract $50 billion from a large Japanese bank. If 90% of this is invested in chips, NVIDIA will have invested $40 billion but gotten $80 billion back. Not a bad return, but at some point, the markets will notice.
By subsidizing chip-buying through AI gigafactories, Europe is making the same mistake as private investors, but with taxpayers’ money. Europe thinks there’s an AI “race” between countries that make AI models and those that do not. But a model’s “weights” are easy to copy. Gigafactories won’t give the people who operate them the chance to charge much for their output. They will become another subsidy to chipmakers, who are already making huge profits.
Europe should instead repeal the EU AI Act, which has chilled investment in the European AI sector. European AI entrepreneurs like me worry that their models will soon reach the AI Act’s thresholds, causing them to be classified as “systemic” and subject to onerous compliance requirements. This threat scares away investors — and encourages European entrepreneurs to move across the Atlantic Ocean.
The US should also be careful. It needs to stop attacking its research capacity, starting by cancelling the new fee (set at over $100,000) for each H1-B visa that allows tech companies to recruit foreign researchers.
If there is a case for public money, it should be to fund research into the next generation of AI technologies, not simply to throw public money into a sector that has already received too much private cash.
Garvan Walshe is a former foreign policy adviser to the British Conservative Party and chair of Unhack Democracy. He is Co-founder of Kronkite, a media tech startup that aims to bring money back into quality journalism.
Bandwidth is CEPA’s online journal dedicated to advancing transatlantic cooperation on tech policy. All opinions expressed on Bandwidth are those of the author alone and may not represent those of the institutions they represent or the Center for European Policy Analysis. CEPA maintains a strict intellectual independence policy across all its projects and publications.
A Roadmap for Europe-US Tech Cooperation
Learn More
Read More From Bandwidth
CEPA’s online journal dedicated to advancing transatlantic cooperation on tech policy.
Read More
