DeepSeek, the Chinese AI startup that rattled Silicon Valley with its low-cost models, is now working on its own artificial intelligence chip. That’s according to three people familiar with the matter, as reported by Reuters. The chip is designed for inference, meaning it would handle the job of running AI models and generating responses for users, not training new ones from scratch.
The project is still early. DeepSeek has been reaching out to chip-design firms, foundries, and memory companies, and has been quietly hiring chip engineers without posting jobs publicly. One source said the effort started about a year ago. DeepSeek did not respond to a request for comment.
News of the chip sent Nvidia shares down roughly 2% in premarket trading. That reaction, while modest, reflects how seriously the market takes any signal that a major AI player might reduce its dependence on Nvidia hardware.
To understand why this matters, you need to know how DeepSeek got here. The Hangzhou-based company became a global name in early 2025 when its R1 reasoning model outperformed expectations at a fraction of the usual cost. That model was trained on Nvidia’s H800 chips, a version designed for China before Washington banned its export in late 2023. Since then, DeepSeek has leaned more heavily on Huawei’s Ascend chips. In April 2026, it released a version of its V4 model adapted for Huawei’s hardware, and Huawei confirmed its processors played a role in training V4-Flash, a lighter variant.
But relying on Huawei has its own risks. U.S. export controls already block Chinese companies from buying Nvidia’s most advanced chips. On top of that, separate restrictions limit China’s access to high-bandwidth memory, a key component for inference chips. DeepSeek building in-house hardware is partly a response to that squeeze, and partly a reflection of where the broader AI industry is heading.
DeepSeek would not be alone in this push. Several major AI companies are moving away from total dependence on Nvidia:
- OpenAI released its first custom inference chip, called Jalapeno, last month, built with Broadcom
- Anthropic has been weighing its own chip development, according to a Reuters report from April
- Chinese rivals Alibaba and Baidu have already developed in-house AI chips and are winning market share from Huawei
That last point is significant. Huawei currently holds around half of China’s $50 billion domestic AI chip market, largely because U.S. sanctions pushed Chinese buyers toward it. But that grip is loosening. If DeepSeek joins Alibaba and Baidu in building its own silicon, Huawei faces pressure from multiple directions at once.
The inference chip market is also the fastest-growing part of AI computing right now. As more AI applications go live, the industry’s computing workload is shifting from training, which is expensive and infrequent, to inference, which happens constantly at scale. Chips built specifically for inference can be cheaper and more energy-efficient than the general-purpose GPUs that dominate training workloads. That makes it an attractive target for any company running AI at scale.
Still, building a competitive chip is hard. It typically takes years and significant capital. Foundry access is also a problem for Chinese designers, who are blocked from using the most advanced overseas fabs.
On the funding side, DeepSeek is also changing course. The company spent years rejecting outside investment, but Reuters reported in June that it was preparing to raise $7 billion in its first external funding round, at a valuation between $52 billion and $59 billion. A chip program of any ambition will need that kind of backing.
DeepSeek’s founder, Liang Wenfeng, acknowledged in a 2024 interview that chip export controls were a real challenge for the company. Building its own hardware would be the most direct answer to that problem, though whether it can actually pull it off remains an open question.
