Digest opened free editor

Rola Khaleda, FT editor, chooses her favorite stories in this weekly newsletter.

Remember when I sent the Chinese Deepseek tremors through the stunning American artificial intelligence industry and Wall Street? It was last month. To listen to executives and investors now, you may think the world has moved. Nafidia, the most difficult blow, has recovered more than half 630 billion dollars lost.

The speed in which the balance has returned a lot to confirm by the largest American technology companies that they will do Spend more What was expected to infrastructure for Amnesty International this year. But it also shows the extent of rewriting the issue of investment in artificial intelligence. The question is how much this real change reflects in the future view, and how just the industry revolves.

The issue of purchasing NVIDIA shares once is based on the claims of such incoming from the CEO of anthropologist Dario Amani, who barely six months ago Propagate The training costs of the advanced large language model will soon reach $ 100 billion. In the aftermath of Deepseek, Amodei still expects a great leap in the demand for artificial intelligence chips – now only, for a completely different reason from the need for more complicated tasks such as thinking, rather than typical training costs.

It is no wonder that investors feel a sharp occurrence and a greater feeling of uncertainty about sustaining the mutation of artificial intelligence.

The Chinese company’s breakthroughs have increased the risk of the most advanced large language models to be converted into commodities. This came just as the builders of the models faced Another existential threatDill quantities of computing energy in training is no longer present.

CEO of Openai Sam Altman signal The clear strategic response in a post on X this week. Openai no longer echoes its large language models as independent products. Instead, it will be filled with other techniques, such as “thinking”, to more complete systems. From now on, “Amnesty International” will only work, whatever the user’s task.

This is a familiar strategy in the technology industry. The movement of “Up the Stack” – building more valuable technologies based on previous products as commoditated – was long being seen as a way to defend prices and profit margins. If the cost of the ingredients that was once provided is a good margin, the more better it is: it decreases the total cost and leads to faster absorption.

This package of artificial intelligence technologies has important effects on the direction of the entire industry. The first is that, like companies like Openai Build Word Systems, a gap at the bottom of the market will be opened for companies like Deepseek.

Anyone who wants to build their own programs of artificial intelligence will turn into large linguistic models such as Llama’s Llama and Deepseek’s R1-techniques that are released in a version of the open source that make it available and cheap. This must open the way for many other technology companies to join the mutation of artificial intelligence. But the former CEO of Google Eric Schmidt to caution This week may be a challenge for the West, making the Chinese company an important global platform in artificial intelligence.

Another meaning is that Amnesty International’s infrastructure suppliers need to quickly control their offers – and their sales playgrounds. The spending will not be severely deviant towards large groups of chips to train permanent models.

NVIDIA, which has increased the value on the boom in training, still has the widest group of Silicon for Amnesty International and will work hard to improve its chips for many different work burdens that will appear with the transformation of the market. But a step that goes beyond intense training should lead to a wide range of technology suppliers fighting in a more contrasting market.

The third meaning is that the continued mutation of artificial intelligence will depend more on the actual use of Amnesty International, and not only the huge correct spending that went to construction and infrastructure models. A large part of the computing power that enters into thinking is a variable cost that is incurred after the insertion of a wave, instead of the one -time fixed cost type entering the training. Artificial intelligence companies need to show that they can provide a real value to finish customers

No new forces are in an industry that was already under pressure to move faster in marketing their technology. But Deepseek shock raised the pressure.

richard.waters@ft.com