DeepSeek Develops Its Own AI Chip to Compete with NVIDIA

In just over a year, DeepSeek has gone from being a rarity in the Chinese industry to a recurring name in the global conversation about artificial intelligence. It first surprised with models like R1 and V4, which compete in performance with GPT-4, but with a computational efficiency that defied expectations. Now, according to information from Reuters citing three sources familiar with the matter, the startup is developing its own AI chip, designed specifically for inference tasks. This move, not yet confirmed by the company, could redefine the AI supply chain and challenge NVIDIA's dominance in accelerators.

What happened?

DeepSeek, the Chinese startup that shook the artificial intelligence industry with its efficient models, is now targeting hardware. According to Reuters citing three sources familiar with the matter, the company is developing its own AI chip, designed specifically for inference tasks — that is, running already trained models — and not for training. The project is in an early stage and DeepSeek has not responded to requests for comment. The news, published on July 7, 2025, does not specify timelines or manufacturing partners, but indicates that the design is being led by a team with previous experience in chips from Huawei and other Chinese semiconductor companies.

Why is this important?

Historically, DeepSeek has relied on chips from NVIDIA (such as the H800) and Huawei to train and run its models. However, U.S. export restrictions have limited access to advanced GPUs, pushing the company to seek alternatives. A proprietary inference chip would allow it to reduce operational costs, improve response speed, and decrease dependence on external suppliers. Moreover, if DeepSeek manages to design an efficient chip, it could directly compete with NVIDIA in the AI accelerator market, fragmenting a dominance that is currently nearly absolute. According to industry data, NVIDIA controls approximately 80% of the AI chip market, but its strength lies in training; in inference, alternatives like Google's TPUs or Amazon's Inferentia have already gained ground. DeepSeek could leverage its software expertise to optimize hardware for its own models, creating an integrated ecosystem similar to Apple with its M chips.

Context and background

DeepSeek is not the first AI company to venture into hardware. Google has its TPUs, Amazon its Trainium and Inferentia, and Microsoft has co-developed chips with AMD. However, DeepSeek's case is particular because it starts as a startup with limited resources compared to tech giants. Its success in software — with models like R1 and V4 that compete with GPT-4 — has given it credibility to venture into silicon. The decision also reflects geopolitical tension: China seeks self-sufficiency in semiconductors, and DeepSeek becomes a key player in that strategy. It is worth recalling that in 2022, the United States imposed restrictions on the export of advanced chips to China, leading companies like Huawei to develop their own solutions. DeepSeek already uses Huawei's Ascend chips for some of its models, but a proprietary design would give it greater control and could integrate better with its software, reducing latency and energy consumption. Furthermore, the startup has proven to be a pioneer in efficiency: its R1 model achieved results comparable to GPT-4 with only a fraction of the computational cost, according to a study published on arXiv in 2024. This focus on efficiency could translate to hardware, where a specialized inference chip could offer much higher performance per watt than general-purpose GPUs.

Market consequences

If DeepSeek manages to launch a competitive inference chip, the consequences will be multiple:

• For NVIDIA: it would lose an important customer and gain a competitor in a key segment. NVIDIA dominates both training and inference, but its strength is in training. A cheap and efficient inference chip could erode its deployment share, especially if DeepSeek offers it to third parties. However, NVIDIA has already responded with its own inference solutions, such as the L40S chip, and its CUDA ecosystem remains a difficult moat to overcome.
• For Huawei: DeepSeek has been a customer of its Ascend chips; if it develops its own, Huawei could lose a strategic partner. However, Huawei could also benefit if DeepSeek outsources manufacturing to its foundries, such as SMIC.
• For the Chinese ecosystem: it would boost technological sovereignty, reducing dependence on foreign chips. Companies like Baidu, Alibaba, and Tencent are also developing their own chips, but DeepSeek could accelerate the adoption of domestic hardware in AI.
• For global startups: it could lower the cost of inference, facilitating the adoption of AI in commercial applications. If DeepSeek achieves a chip with significantly lower cost per query, it could democratize access to advanced models, especially in regions where cloud costs are prohibitive.

What readers should know

The project is not confirmed and may not materialize. DeepSeek faces enormous technical and financial challenges to design and manufacture competitive chips. Moreover, the focus on inference limits its immediate impact, as training remains the most expensive bottleneck. However, if DeepSeek achieves its goal, it could redefine the AI supply chain and demonstrate that startups can also challenge NVIDIA on its own turf. Recent history shows that companies like Graphcore and Cerebras have tried to compete without success, but DeepSeek has the advantage of controlling both software and hardware, which could enable vertical optimizations. Readers should closely follow the next steps: hiring of semiconductor engineers, possible patents, and alliances with Chinese manufacturers like SMIC. For now, caution is mandatory, but the mere announcement has already caused drops in NVIDIA's stock and renewed interest in Chinese alternatives.