IA: Meta on the war footing to counter Deepseek

If you are interested in the news of artificial intelligence, you could not miss the phenomenon Deepseek. The Chinese AI seems to have taken the Silicon Valley de Court. It is not only its capacities that shock American eggs in the sector, but especially the under 6 million dollars mentioned for R1 training, the most efficient model of the Chinese company. Especially at a time when, freshly invested, Donald Trump has announced a 500 billion pharaonic investment, led by Openai, Oracle, Softbank and a crowd of other actors who count in the middle of the AI.

This XXL project is far from being the only major project on the other side of the Atlantic. Indeed, according to information confirmed by several internal sources to Fortune,, Meta would devote $ 65 billion to artificial intelligence in 2025or a little less than double of its IA budget of 2024. The accent would be put in particular on the improvement of its models like Llama. This financial effort includes the recruitment of thousands of engineers and the construction of a huge data center in Louisiana.

Deepseek, high -end high -cost performance

Obviously, with regard to the astronomical sums mentioned, the announcement of the training cost of the R1 model has the effect of a bomb. It must be said that the Chinese start-up, founded in 2022, already challenges technological giants with models of AI as powerful for certain tasks as GPT-4, O1 or LLAMA 3, but for a very lower budget .

Among the hypotheses most cited to explain this difference, we can mention a more effective approach in model training, the use of targeted datasets, reducing the need for massive raw data, or even partnerships with suppliers of Chinese cloud, allowing lower infrastructure costs.

There is therefore an enigma that the eggs of the American AI must resolve quickly. Because the billions of dollars they have agitated throughout communication in recent months have turned against them when an actor, in addition Chinese, comes to ridicule them with very efficient, but much less expensive models. As evidenced Breakdown on the stock market The main American players in the sector in recent days.

To understand as soon as possible how Chinese researchers got caught up, Meta Really put himself on the warfare. The Mark Zuckerberg firm did not wait for the latest Deepseek announcements to feel the danger. Because, like its Llama models, those of Deepseek are published under an open source license, allowing the community to use them, modify them and improve them without significant restrictions. Also, from last December, Meta urgently mobilized teams, nicknamed “War Rooms”, made up of machine learning researchers and hardware experts. Their mission: to analyze Deepseek publications, recruit former employees and experiment with open source replicas of their models. Concretely, this is a real Reverse Engineering operation of a considerable scale.

Under cover of anonymity, an engineer working in one of these War Rooms describes an atmosphere between fascination and frustration: ” Their techniques challenge our manuals. For example, they manage to stabilize learning with a higher error rate in the initial phase, which accelerates training. ». These investigations have already led to major adjustments. Meta would currently test a lightened version of Llama, requiring 40 % fewer GPU resources while retaining 95 % of its capacities.

A geopolitical battle in the background

Beyond the purely technical aspect, the current success of Deepseek sounds like a camouflet for all American actors. Donald Trump himself described the Chinese AI of “warning” for Statesunian companies in the sector. Western observers agreed to affirm that the Asian giant was at least two years late in this area, especially due to the impossibility for its companies to access the most modern American fleas. But, supported by massive public investments, Chinese companies have quickly developed alternatives based on architectures less energy -consuming, and much less expensive. Deepseek's success shows in any case that in terms of AI, China does not intend to leave without doing anything.