Last week, semiconductor stocks like Nvidia (Nvda 3.17%)),, Advanced micro-apparents (Dmla -9.76%))And Micron technology (Mu 0.60%)) Immersed on the news that a Chinese start-up called Deepseek had found how to form artificial intelligence models (AI) for a fraction of the cost of its American peers.
Investors were concerned about the fact that the innovative approach of Deepseek would trigger a collapse of demand for graphic processors (GPU) and other components of the data center, which are essential to the development of AI. However, these concerns could be exaggerated.
Meta-platforms (Meta 0.59%)) is a huge Buyer of Nvidia and AMD ia chips. On January 29, CEO Mark Zuckerberg made a series of comments that should be music to the ears of investors who have material actions on AI.
Image source: Getty Images.
Depth background
The high level of Chinese Fund Hedge used IA To build trading algorithms for years. He established Deepseek as a distinct entity in 2023 to capitalize on the success of other AI research companies, which quickly increased.
Last week’s stock market panic was launched by the Deepseek V3 Great language model (LLM), which corresponds to the performance of the latest GPT-4O models of the first start-up of the American AI, OPENAIthrough several landmarks. It is not a concern for nominal value, except that Deepseek claims to have spent only $ 5.6 million in V3 training, while Optai burned more than $ 20 billion Since 2015 to reach its current scene.
To make things more worrying, Deepseek does not have access to the latest GPUs in the Nvidia data center, as the US government has prohibited them from being sold to Chinese companies. This means that the start-up had to use older generations such as the H100 and the underlying H800, indicating that it is possible to form main AI models Without the best equipment.
To compensate for the lack of calculation performance, Deepseek innovated on the software side by developing more effective algorithms and data input methods. In addition, he adopted a technique called distillation, which consists in using a successful model to form his own smaller models. This quickly accelerates the training process and requires much less computer capacity.
Investors fear that if other IA companies adopt Deepseek’s approach, they will not need to buy as much Gpus of NVIDIA or AMD. This would also crush demand for Micron Data Center memory solutions.
NVIDIA, AMD and Micron Power the AI revolution
Nvidia GPUs are the most popular in the world to develop AI models. The company 2025 of the company has just ended on January 31 and, according to management advice, its income Probably more than doubled at a record of 128.6 billion dollars (the official results will be published on February 26). If recent quarters are something, around 88% of this income will come from its data center segment thanks to GPU sales.
This incredible growth is the reason why Nvidia added 2.5 billions of dollars to its market capitalization Over the past two years. If the chip request was to slow down, a large part of this value would likely evaporate.
AMD has become a worthy competitor in Nvidia in the data center. The company plans to launch its new MI350 GPU later this year, which is expected to compete with the latest Blackwell chips from Nvidia which have become the OR stallion for the treatment of IA workloads.
But AMD is also one of the main IA flea suppliers for personal computers, which could become a major growth segment in the future. As LLM becomes cheaper and more effective, it will ultimately be possible to execute them on smaller chips inside computers and devices, reducing dependence on external data centers.
Finally, Micron is often overlooked as a flea company AI, but it plays an essential role in the industry. Its HBM3E (high bandwidth memory) for the data center is the best in the class with regard to capacity and energy efficiency, which is why Nvidia uses it in its last Blackwell GPU. Memory stores information in a ready condition, which allows the GPU to receive it instantly if necessary, and as the workloads of the AI are so strong of data, it is an important part of the material puzzle.
Image source: Getty Images.
Mark Zuckerberg may have put recent concerns in bed
Meta Platforms spent $ 39.2 billion over chips and data center infrastructure in 2024, and plans to spend up to $ 65 billion this year. These investments help the company further progress its LLAMA LLMS, which are the most popular open source models in the world, with 600 million downloads. Llama 4 should be launched this year and CEO Mark Zuckerberg Think that it could be the most advanced in the industry, even surpassing the best closed source models.
On January 29, Meta held a conference call with analysts on her fourth quarter of 2024. When Zuckerberg was asked about the potential impact of Deepseek, he said that he was probably too early to determine what it means for capital investments in fleas and data centers. However, he said even if it leads to less capacity requirements AI training The workloads, this does not mean that companies will need fewer fleas.
Instead, he thinks that the capacity could move away from training and to inferencewhich is the process by which the AI models deal with user inputs and form responses. Many developers are moving away from training models using endless amounts of data and rather focusing on “reasoning” capacities. This is called time to scale, and this implies that the model takes more time to “think” before making an outing, which results in better quality responses.
The reasoning requires more inference calculation, so Zuckerberg thinks that companies will always need the best data center infrastructure to maintain an advantage over competition. In addition, most AI software products have not yet obtained the consumer adoption, and Zuckerberg acknowledges that the service of many users will also need additional capacity of the data center over time.
So, although it is difficult to put exact figures on how Deepseek’s innovations will reshape the chip request, Zuckerberg’s comments suggest that there is no reason for NVIDIA action investors, AMD and Micron. In fact, there is Even an optimistic case for these stocks in the long term.