Unlock free Digest editor
Roula Khalaf, editor of FT, chooses her favorite story in this weekly newsletter.
Remember when Chinese Deepseek sent shivering through the American artificial intelligence industry and stunned Wall Street? That was last month. To now listen to AI and investors executives, you might think that the world continued on. Nvidia, the hardest hit, recovered more than half of the $ 630 billion he lost.
The speed at which the balance returned to the claim by the largest American technological companies he wants spend more than expected on AI infrastructure this year. But it also shows how quickly the case of AI investment is quickly transcribed. The question is how much it reflects a true change of chance and how much is the industry alone.
The case for the purchase of Nvidia shares was once resting on claims such as those from the Anthropic Executive Director Dario Amodei, who was just six months ago just six months ago intended To make the cost of training for a top -notch linguistic model soon reach $ 100 billion. In the midst of Deepseek, Amodea still expects a great jump in demand for AI chips – only now, for a completely different reason, they are needed for more complex tasks such as reasoning, not for the model training costs.
No wonder investors feel an acute blow and a greater sense of uncertainty about the sustainability of AI boom.
The breakthroughs of Chinese companies have increased the risk that even the most advanced large -language models will quickly turn into goods. That came just when the model builders were facing Another existential threat: Throwing increasing amounts of computer power in training no longer makes the progress he once did.
OPENAI CEO of Sam Altman signaled An obvious strategic response in the post on X this week. They will no longer open its big linguistic models as independent products. Instead, they will be packed together with other technologies, such as “reasoning”, into more complete systems. From now on, he said, he will “just work”, regardless of the task the user throws.
This is a well -known strategy in the technological industry. Moving “Gore” – the construction of more valuable technologies based on earlier products while comfortable – has long been considered a way to defend the price and profit margin. If the costs of the components that have once provided a good margin were demolished, so much better: it lowers the total costs and leads to faster intake.
This packaging AI technology has important implications to the direction of the entire industry. One is that, as companies like Openai build more complete systems, it will open a gap at the bottom of the market for companies like Deepseek.
Anyone who wants to build their own software that runs Ai turns to large linguistic models such as Meta Llam and Deepseek R1-Technology that are published in an open code version that makes them freely available and cheap. This should open the way for many more technological companies to join AI boom. But former Google CEO Eric Schmidt warned This week he could represent a challenge to the West, making a Chinese company an important global platform in AI.
Another implication is that the suppliers of AI infrastructure must quickly adapt their offer – and their sales plots. Consumption will no longer be so distorted by large bunches of chips for training versatile models.
Nvidia, who has grown in training at training, still has the widest series of AI silicon and will work hard to optimize her chips for many different loads that will appear as the market is changing. But a move outside of intense training should lead to a wider range of technology suppliers struggling in a much different market.
The third implication is that the continuation of AI boom will depend much more on the actual use of AI, not just on the huge consumption in advance that has entered the construction of models and infrastructure. A large part of a computer power that goes into explanation is a variable cost after the inquiring of inquiries, not a type of disposable fixed costs that go to training. AI companies need to show that they can provide real value to end customers
None of these forces are new in an industry that has already been under pressure to move faster in commercializing its technology. But Deepseek’s shock just increased the pressure.
