What DeepSeek’s AI really means for the market

Open the Editor’s Marking Free

Rulla Khalaf, the FT editor, chooses its preferred stories in this week’s news.

In the past 40 years, the most important trend of finance in technology is the result of the winning structures in the field of technology. We live in a world of network effects and grow back on a scale. Microsoft, online retail (Amazon) and social media (Meta), have gone to one company. It is only a moderate exaggeration that has been important events for investors in the last few decades on one or more of all the stories.

So, it is natural that everyone or assumed, wanted or feared that the trend of the next wonderful technology, artificial intellect, will have an economic structure. The incident in the markets yesterday was not a surprise, neither panic nor the bubbles. It was pricing tref More likely odds that AI is not a winner-game.

It was always clear that the AI ​​revolution would be like an automotive or plane invention. These revolutions led to huge, duration of profitable companies, but also a lot of competition, ensuring most of the value created than the shareholders. Now it appears to be a more important opportunity for AI.

Most readers will now know the news. Deepseek, Chinese company AI, has issued ai model ai model, which is comparable to the best models by companies such as Openai, Anthropic and Meta, but have been trained at a radically lower price and used lower. Price: art GPU chips. Deepseek also met the details of the model community that others can run it in their computers (but others do not necessarily be able to recreate).

This puts three punctures in the winner-long theory. It tempts the idea that the best AI results can only be achieved through the best chips and software of NVIDIA. The idea that only the largest technology companies can afford to build and run high quality AI models. And the idea that only with their AI models can offer great AI applications.

Over the past two years, the NVIDIA Stock Missile Missile became a clean calculation priority. Its chips were better for AI. Deepseek has built R1 with older nvidia with older chips, which US sanctions allow to export China. This suggests that the door is open to other competitors at the Silicone level.

Next: the value. The largest technology companies were more or less isolated from AI’s competition, as it was supposed to make the only way to build a better AI model through more contributions. The finished model would then be under the control of the large company that built it. The entrance barrier was money. That’s why the big technology passed on the cost of the data center. But R1 assumes that companies with less money can operate competitive models soon.

Finally, applications. “The” is the list “where the user really meets AI. The questions were answered, the tasks are over. It was believed that any application would depend on models and servers of larger companies. But modes like R1 can work well enough and make it effective to be effective in order to launch all types of companies that need to be “hired” company such as Openai.

All this has great consequences. But, in addition to Nvidia, the big ones were not seven large technologies yesterday. The Ai Halos of those companies remains a square. The real damage was carried out to the stock of the economy of the developing data center. Constellation, a utility, which includes the Node of the Middle Atlantic Data Center, and Washington State Big Power Power, NRG, Northoast’s and Texas Main Player.

GE Vernova, Eaton and Quanta Services Build Energy Systems for Data Centers. Oracle simply announced a large investment of the data center. Broadcom and Arista provide non-GPU technology to data centers. Everyone was beaten.

Amazon, the alphabet, Meta and Microsoft have all made a lot of investments in AI data centers. Maybe they wasted some of those money, but they could be free to spend less now. Apple and Microsoft, whose natural power could be in the construction of applications for AI models, and not even in a better position.

Nijdia itself is barely on her knees. His property coding language, CUDA, however, is the industry standard. And just because the Deepseek model is more effective, it does not mean that Niv’s models will not help the best calculated power offered by NVIDIA. Before its shares, almost 17 percent fell yesterday, which only brings it to the level of September (very, very high).

We should not overestimate the market response to R1. NASDAQ fell by 3 percent. It was a bad day, not panic. The winner-all AI’s view is wounded, not dead. AI Bubble if that’s what’s up can still pop, but it won’t fly yesterday. Not even close.

Read one better

Located hope.

Robert.armstrong@ft.com aeration of Aiden.reiter@ft.comA number

Ft helpless podcast

Can’t be satisfied unsafe. Listen to Our new podcast15 minutes diving in the latest markets News and financial titles twice a week. Catch the newsletter to past publications hereA number