China’s AI champion DeepSeek is reportedly working on its own artificial intelligence chip, marking a major strategic shift from model development to AI hardware as geopolitical pressure, export controls, and inference demand reshape the chip race.
TL;DR
- DeepSeek is reportedly developing an AI chip focused on inference, not training.
- The move could reduce its dependence on Nvidia and Huawei chips, although the project is still at an early stage.
- The effort comes as Chinese AI firms face U.S. chip curbs and Beijing pushes domestic alternatives.
DeepSeek may be taking a big step beyond artificial intelligence model development.
According to Reuters, the Chinese AI startup is developing its own AI chip, citing three people familiar with the matter. The chip is being designed for inference, which is the stage where trained AI models generate responses for users, rather than for training new models.
Source
That detail matters. Training frontier AI models typically requires large clusters of powerful GPUs, while inference is becoming a massive day-to-day computing workload as AI apps reach more users. A dedicated inference chip could help DeepSeek lower costs, improve control over its AI stack, and reduce dependency on outside hardware suppliers.
Reuters reported that DeepSeek’s chip effort began about a year ago and remains in an early phase. The company has reportedly reached out to external partners and held discussions with chip design, foundry, and memory companies. It has also increased private hiring of chip design engineers in recent months, according to the report.
DeepSeek did not respond to Reuters’ request for comment.
If successful, the move would place DeepSeek in the same broad direction as global AI players building custom chips to control performance and cost. Reuters noted that OpenAI unveiled its first custom inference chip, Jalapeno, with Broadcom last month, while Anthropic has also been weighing custom AI chips.
However, DeepSeek’s motivation carries a China-specific urgency.
U.S. export restrictions have blocked Chinese companies from buying Nvidia’s most advanced AI chips. DeepSeek has used Nvidia and Huawei chips for its models, and Reuters reported that the foundation model behind R1 was trained on Nvidia H800 chips, a China-focused chip later banned by Washington in late 2023.
DeepSeek’s own technical report for V3 said the model required 2.788 million H800 GPU hours for full training, with official training costs estimated at $5.576 million, excluding prior research and experimental costs. That efficiency helped DeepSeek become one of the most closely watched AI companies in the world.
The company also gained global attention with DeepSeek-R1, which it said delivered performance “on par with OpenAI-o1” and was released with open-source model weights and technical details.
The latest reported chip push follows DeepSeek’s growing alignment with domestic AI hardware. In April, Reuters reported that DeepSeek returned with a new model adapted for Huawei chips, with Omdia’s semiconductor research director He Hui saying, “Huawei's Ascend chips are the country's best homegrown alternative to Nvidia.”
Yet making a competitive AI chip will not be easy. Chip design can take years, requires major capital, and depends on advanced manufacturing and high-bandwidth memory. Reuters also noted that U.S. restrictions limit Chinese chip designers’ access to top overseas foundries and memory components essential for AI chips.
Topics for more insights:
Markets still reacted quickly. Reuters reported that Nvidia shares slipped about 1.6% in premarket trading after the news, while Barron’s also noted pressure on Nvidia stock as investors weighed the long-term implications of customers and rivals moving toward custom AI silicon.
For now, DeepSeek’s chip project remains unconfirmed by the company and is still early. Yet the direction is clear. The AI race is no longer just about who has the smartest model, but who controls the chips, infrastructure, and supply chain behind it.
