DeepSeek is reportedly intensifying its efforts to develop AI chips in-house, driven by persistent challenges posed by overseas computational power blockades. The goal is to establish an independently controllable foundational computational power framework.

On July 7, Sina Tech reported that foreign media sources indicated DeepSeek is developing its own AI chips to diminish reliance on companies such as NVIDIA.

Sources reveal that DeepSeek's in-house AI chips are primarily designed for inference tasks, rather than for training AI models. "The company has been collaborating with external partners, engaging with chip design, foundry, and memory chip firms," the source disclosed, noting that DeepSeek initiated its self-developed chip project approximately a year ago.

DeepSeek has not yet issued a response to inquiries about its self-developed AI chips.

According to IT Home, Richard Windsor, an analyst at Radio Free Mobile, commented, "NVIDIA's market share in China has essentially fallen to zero, and this situation is expected to continue. Unless DeepSeek can access the most advanced manufacturing processes, its chips are unlikely to gain traction in overseas markets. Therefore, this move is unlikely to have a significant impact on NVIDIA."

DeepSeek has long been recognized for driving advancements in AI technology rather than prioritizing commercialization. Sources indicate that the company's chip project is still in its nascent stages. In recent months, DeepSeek has been discreetly increasing its recruitment of chip design engineers without publicly advertising job openings.

Zhidx reports that DeepSeek has long focused on computational power and chips. In interviews with Any surge in 2023 and 2024, Liang Wenfeng, founder and CEO of DeepSeek, stated, "Our real challenge has never been funding but the export ban on high-end chips," "For the research institute, the demand for computational power is unquenchable," and "We will consciously deploy as much computational power as possible."

Currently, DeepSeek utilizes both NVIDIA and Huawei chips. NVIDIA's H800 previously underpinned many of DeepSeek's model trainings, but the U.S. government prohibited its export at the end of 2023. Since then, DeepSeek has progressively increased its reliance on Huawei's Ascend chips.

In April, DeepSeek unveiled the V4 model, compatible with Ascend chips. Huawei confirmed that Ascend chips contributed to part of the training process for the V4-Flash lightweight model. Reuters previously reported that following the model's release, Chinese tech companies significantly boosted their orders for Huawei's Ascend 910B chips.

DeepSeek is not the first major Chinese model company to pursue in-house AI chip development. According to the Hong Kong Economic Journal, tech giants like Alibaba and Baidu are also developing their own AI chips and gradually capturing market share. If DeepSeek successfully launches its own inference chips, competition in China's AI chip market will intensify further.

As AI applications proliferate, industry computational demands are gradually shifting from training large models to inference operations. Global AI companies are also seeking greater control over their hardware supply. OpenAI unveiled its first in-house inference chip, codenamed Jalapeno, developed with assistance from Broadcom last month. Anthropic has also explored in-house chip development.

If DeepSeek's in-house chip development is confirmed, it signals that leading domestic large-scale model companies are accelerating efforts to address upstream computational power deficiencies. If DeepSeek's inference chips are successfully implemented, it will establish a complete industrial closed loop of algorithm models plus in-house computational power hardware, further reducing costs and lowering the barrier for domestic developers to utilize large-scale models.