Introduction
DeepSeek, the Chinese artificial intelligence company, which stunned Silicon Valley and Washington with its efficient reasoning models, is building its own AI chip, according to people familiar with the matter cited by Reuters. Such a strategy, if executed successfully, would be a major shift in the strategies of a company that was until now known more for its software efficiency rather than for semiconductors, which would allow it to reduce its dependence on both Nvidia and Huawei, the two chip companies it has been using to train and run its widely used models.
A chip created to train new large language models would have different requirements than the one DeepSeek is working on now—reportedly, the new device will be dedicated to the phase of AI inference, which involves generation of responses, creation of images and code, translations, and document summaries by an AI assistant after training is finished.
Since every interaction with an AI assistant requires inference, the demand for specialized inference computing increases all the time, and chips designed for it can process requests faster and cheaper.
As was mentioned before, the chip development project is quite young—the process has been ongoing for almost a year now but is still at its early stages. The company has been assembling a team of chip designers, hiring engineers, and negotiating with chip design companies, foundries, and memory manufacturers about possibilities to create such a chip.
Reasons for DeepSeek to Expand Research Past Nvidia and Huawei
In the past, DeepSeek had been dependent on Nvidia's H800 chips, which were a special edition of Nvidia's AI chips that were designed for China after Washington started placing export restrictions on the most advanced chips produced by the company.
When the situation worsened due to increased export restrictions by the U.S., the availability of even the downgraded H800 chips was not certain anymore, making it difficult for DeepSeek to continue using Nvidia chips, and thus it moved on to Huawei's Ascend AI chips for its newer AI operations.
Despite the shift from Nvidia to Huawei, being dependent on suppliers like Nvidia and Huawei is risky since it may be restricted due to supply problems, price changes, and potential future restrictions.
The development of its own chip would help DeepSeek to control its hardware plans better.
This move is part of the wider push by Beijing to construct domestic alternatives for the chips made by Nvidia. American export controls do not allow Chinese firms to procure the most advanced chips manufactured by Nvidia, and Beijing has encouraged its tech giants to develop indigenous alternatives.
A recent survey carried out by Bloomberg Intelligence shows that Chinese business leaders expect their AI infrastructure budget to allocate 46% of it to domestic AI infrastructure within one year compared to the 30% it is currently allocating, and this would work in favor of chip manufacturers like Huawei and Hygon irrespective of the success of DeepSeek's chip.
Industry News and Market Reaction
The stock price of Nvidia dropped slightly by 1.6% to 2% in the pre-market following reports about DeepSeek's plan to develop its AI computing chip. The drop in stock prices showed that the investors were reacting sensitively to the news that AI developers, especially Chinese ones that Nvidia has been locked out of because of export controls, were trying to build their own computing stack.
According to Richard Windsor, an analyst at Radio Free Mobile, DeepSeek’s chances of selling chips it will manufacture outside China will be very minimal due to the lack of access to cutting-edge manufacturing technology.
In this case, he argued, the point is not what will happen to Nvidia but rather how the entry of DeepSeek will impact competition within the Chinese market, especially with Huawei.
Currently, Huawei occupies an estimated 50% of China’s $50 billion homegrown AI chip industry, supplying chips to DeepSeek and other leading companies in China despite the fact that its chips are much less advanced compared to those offered by Nvidia. Its domination in the market is already being threatened by the efforts of rival Alibaba and Baidu, which are developing their own AI chips.
The decision to make its own AI chips can be attributed to a trend within the AI industry as more and more firms develop models that seek greater control over the hardware on which the models will run.
Earlier this month, OpenAI launched Jalapeno, its first custom AI inference chip that it designed in collaboration with Broadcom. Anthropic has been considering the idea of designing AI chips of its own, as reported by Reuters.
Both Microsoft and Meta have turned towards internal silicon designs to reduce the cost of their AI infrastructures.
The Challenges Ahead
Despite having a compelling case for building AI chips, DeepSeek will still face some significant challenges.
American regulations prevent Chinese chipmakers from using the best overseas foundries for manufacturing the chips, and other American regulations have denied China access to high-bandwidth memory.
Building a chip that can truly compete with those of Nvidia and Huawei will take a number of years and a large amount of money, not considering the production and logistics challenges DeepSeek will face.
Experts say that the early stages of development make it impossible for the project to create a competitive chip soon.
The development of a chip is especially noteworthy considering the period of development of the DeepSeek corporation itself: in June, the company was said to be planning on raising $7 billion during the first ever fundraising round in its history, valuing the company somewhere between $52 billion and $59 billion. This will mark a break in the policy of DeepSeek of never seeking outside funding before.
