Remember when China’s DeepSeek sent tremors through the US artificial intelligence industry and stunned Wall Street? That was last month. To listen to AI executives and investors now, you might think the world has moved on. Nvidia, the hardest hit, has recovered more than half the $630bn it lost.
The speed with which equilibrium has returned owes a lot to the assertion by the biggest US tech companies that they will spend even more than expected on AI infrastructure this year. But it also shows how quickly the investment case for AI has been rewritten. The question is how much this reflects a genuine change in outlook, and how much is just industry spin.
The case for buying Nvidia stock once rested on claims such as those from Anthropic chief executive Dario Amodei, who barely six months ago predicted that the training costs for a cutting-edge large language model would soon reach $100bn. In the wake of DeepSeek, Amodei is still anticipating a huge jump in demand for AI chips — only now, it is for the completely different reason that they are needed for more complex tasks like reasoning, rather than the costs of model training.
No wonder investors are feeling an acute whiplash and a greater sense of uncertainty about the sustainability of the AI boom.
The Chinese company’s breakthroughs increased the risk that even the most advanced large language models will quickly be turned into commodities. This came just as model-builders were facing another existential threat: throwing ever-greater amounts of computing power into training no longer produces the advances it once did.
OpenAI chief executive Sam Altman signalled the obvious strategic response in a post on X this week. No longer will OpenAI release its large language models as standalone products. Rather, they will be packaged together with its other technologies, such as “reasoning”, into more complete systems. From now on, he said, the AI will “just work”, whatever task a user throws at it.
This is a familiar strategy in the tech industry. Moving “up the stack” — building more valuable technologies on the foundation of earlier products as they are commoditised — has long been seen as the way to defend prices and profit margins. If the cost of components that once provided a good margin collapse, so much the better: it brings down the overall cost and leads to faster uptake.
This packaging of AI technologies has important implications for the direction of the whole industry. One is that, as companies such as OpenAI build more complete systems, a gap will open up at the bottom of the market for companies like DeepSeek. DeepSeek.
Anyone wanting to build their own AI-powered software will turn to large language models such as Meta’s Llama and DeepSeek’s R1 — technologies that are released in a version of open source that makes them freely available and cheap. This should open the way for many more tech companies to join in the AI boom. But former Google chief executive Eric Schmidt warned this week it could pose a challenge to the west, making the Chinese company an important global platform in AI.
Another implication is that AI infrastructure suppliers need to quickly adjust their offerings — and their sales pitches. Spending will no longer be so heavily skewed towards big clusters of chips for training ever-larger models.
Nvidia, which soared in value on the boom in training, still has the widest array of silicon for AI and will be working hard to optimise its chips for the many different workloads that will emerge as the market shifts. But the move beyond intensive training should lead to a wider range of technology suppliers fighting over a much more disparate market.
A third implication is that the continuation of the AI boom will depend much more on the actual usage of AI, not just the massive upfront spending that has gone into building models and infrastructure. Much of the computing power that goes into reasoning is a variable cost incurred after a prompt has been entered, rather than the kind of one-off fixed costs that go into training. The AI companies need to show they can provide real value to end customers.
None of these forces are new in an industry that was already under pressure to move faster in commercialising its technology. But the DeepSeek shock has just turned up the pressure.