DeepSeek Develops AI Chip to Cut Nvidia, Huawei Reliance

DeepSeek, a Chinese startup, is developing its own artificial intelligence chip. This initiative aims to reduce the company's reliance on chips from Nvidia and Huawei, which it has used to train and operate its globally recognised models.

Credit: UNSPLASH

The new chip is specifically designed for inference, the phase of AI computing where a trained model generates responses for users, rather than for the initial training of new models. This expansion into semiconductor development would represent a significant strategic shift for DeepSeek, widely regarded as China's leading AI organisation.

And the move could introduce new challenges for Chinese tech giant Huawei. Shares of U.S. based Nvidia experienced a decline of approximately 1.6 per cent in premarket trading following the news. Analyst Richard Windsor of Radio Free Mobile stated that Nvidia holds no market share in China and will likely remain in that position, further commenting that DeepSeek has almost no opportunity to sell silicon outside of China without access to leading edge manufacturing, adding that the development does not affect the chipmaker.

DeepSeek gained global attention over a year ago after releasing two highly efficient AI models that quickly circulated worldwide, surprising observers in Silicon Valley and Washington. For a considerable period, the company was primarily known for its emphasis on breakthroughs in AI models rather than focusing on the commercialisation of its technology.

But Huawei's existing chip offerings, though significantly less advanced than Nvidia's top tier products, have benefited from a U.S. ban on exports to China. This ban has enabled Huawei to secure roughly half of the domestic AI chip market, valued at USD 50 billion, supplying DeepSeek and several other major industry players. Huawei's dominance in this market is, however, already showing signs of weakening as rival Chinese tech firms, including Alibaba and Baidu, are also engaged in developing their own AI chips and subsequently increasing their market share.

DeepSeek's entry into the chip development race remains in its early stages. The company has engaged with external partners and held discussions with organisations specialising in chip design, foundry services, and memory production, three sources familiar with the matter indicated. This endeavour commenced approximately one year ago, according to one of these individuals.

The Hangzhou based company has also privately increased its hiring of chip design engineers in recent months, avoiding public job postings on hiring platforms, two sources reported. All three individuals familiar with the situation declined to be identified due to the non public nature of the information. Despite its role as a key player in China's AI ambitions, DeepSeek has consistently maintained a low public profile.

So, in developing an in house chip, DeepSeek aligns with a broader trend among global AI developers. Organisations such as OpenAI and Anthropic are also seeking greater control over the hardware underpinning their models and aiming to lessen their reliance on Nvidia's products. OpenAI, for example, introduced Jalapeno, its first custom inference chip developed with Broadcom, last month, while Anthropic has been considering building its own AI chips, Reuters reported in April.

This effort carries an additional strategic dimension for DeepSeek due to U.S. export controls. These regulations prevent Chinese companies from purchasing Nvidia's most advanced chips, and the Beijing government has been urging its technology organisations to develop domestic alternatives. DeepSeek founder Liang Wenfeng stated in a rare 2024 interview with a Chinese media outlet that chip export controls presented a significant challenge for his company.

DeepSeek has historically utilised both Nvidia and Huawei chips. The company confirmed that the foundational model supporting R1, its reasoning model whose low cost performance contributed to a downturn in U.S. tech stocks in Jan. 2025, was trained using Nvidia's H800, a chip manufactured for the Chinese market and subsequently prohibited by Washington in late 2023.

And since then, the company has increasingly relied on Huawei's offerings. In April, DeepSeek launched its V4 model, which was adapted for Huawei's Ascend chips. Huawei confirmed that its processors were partly used in the training of V4 Flash, a lighter version of that model. Following this launch, orders for Huawei's Ascend 950 chips from various Chinese tech conglomerates reportedly surged.

A dedicated DeepSeek inference chip would target the fastest growing sector of AI computing demand. As AI applications proliferate, a greater proportion of the industry's computational tasks is shifting from training models to running them, a process that relies on specialised chips. These chips can often be more cost effective and consume less power than general purpose graphic processing units.

However, success is not guaranteed in this endeavour. Designing a competitive AI chip typically necessitates several years and a substantial capital investment. Manufacturing presents another hurdle, as U.S. restrictions prohibit Chinese designers from accessing the most advanced overseas foundries. Furthermore, separate U.S. curbs have restricted China's access to high bandwidth memory, a component deemed crucial for AI inference chips.

But this push into chip development coincides with DeepSeek's first acceptance of external capital. The company was slated to secure USD 7 billion in its initial funding round, which would value it between USD 52 billion and USD 59 billion, according to reports in June. This represents a reversal of its long standing policy of declining outside investment.

• DeepSeek, a Chinese AI startup, is developing its own inference chip to lessen reliance on Nvidia and Huawei.

• This strategic shift moves DeepSeek into semiconductor development, an area previously not its focus.

• The effort is in early stages, involving external partner discussions and private hiring of chip design engineers.

• U.S. export controls on advanced chips have prompted Beijing to encourage domestic alternatives, adding to DeepSeek's motivation.

• The company is also reportedly poised to raise USD 7 billion in its first funding round, reversing its stance on external investment.

Source: Reuters