DeepSeek, the Chinese AI startup, is developing its own semiconductor for inference workloads, three sources told Reuters. The effort, which began roughly a year ago, would reduce the company's dependence on Nvidia and Huawei hardware and marks a strategic shift for a firm known for model breakthroughs rather than chip design.
The chip is designed for inference - the stage where a trained model generates responses - not for training new models. DeepSeek has used both Nvidia and Huawei chips to date. Its R1 reasoning model, which triggered a selloff in U.S. tech stocks in January 2025, was trained on Nvidia H800 processors before Washington banned them in late 2023. Since then, the company has leaned increasingly on Huawei's Ascend chips. In April it released a version of its V4 model adapted for those processors, and Huawei confirmed its hardware was used in part of the training for V4-Flash.
DeepSeek's chip initiative remains at an early stage. The company has reached out to external partners in chip design, foundry, and memory, and has quietly hired chip-design engineers without public job postings, two of the sources said. All three people declined to be identified because the plans are not public. DeepSeek did not respond to a request for comment.
Following a global pattern
With an in-house chip, DeepSeek joins other AI developers seeking greater control over their hardware stack. OpenAI last month unveiled Jalapeno, its first custom inference chip built with Broadcom, and Anthropic has been weighing its own chip plans, Reuters reported in April. For DeepSeek, the move carries added strategic weight because U.S. export controls bar the company from buying Nvidia's most advanced GPUs. Founder Liang Wenfeng said in a rare 2024 interview that chip export controls were a challenge for the company.
The company's efficient models, including the R1 reasoning model that sent U.S. tech stocks tumbling, were trained on Nvidia's H800 chips before Washington banned them. Since then, DeepSeek has leaned more heavily on Huawei's Ascend processors. For IT professionals analyzing these model architectures, Deepseek AI Training resources detail the techniques involved.
Tapping the inference boom
A DeepSeek inference chip would target the fastest-growing segment of AI computing. As AI applications spread, more compute shifts from training to inference, which can use cheaper, less power-hungry chips than general-purpose GPUs. Designing a competitive chip typically takes years and significant capital. Manufacturing presents another hurdle: U.S. sanctions block Chinese designers from advanced overseas foundries and restrict access to high-bandwidth memory, a component critical to inference chips.
Analyst Richard Windsor of Radio Free Mobile said, "Nvidia is at zero in China and staying there. DeepSeek has almost no chance of selling silicon outside of China unless it gets access to leading edge manufacturing." He added that the development does not affect Nvidia.
DeepSeek's changing strategy
The chip push coincides with DeepSeek's first acceptance of outside investment. Reuters reported in June that the company planned to raise $7 billion in a maiden funding round at a valuation between $52 billion and $59 billion, reversing its long-standing practice of rejecting external capital. Huawei, which currently supplies DeepSeek and holds about half of China's $50 billion domestic AI chip market, faces growing pressure as rivals Alibaba and Baidu also develop their own AI chips.
Why this matters for IT and Development professionals
The move by DeepSeek underscores a broader industry shift: AI model developers are increasingly designing their own inference chips to optimize performance and cost. For IT and development teams, this trend means hardware decisions will become more tightly coupled with software stacks. Understanding the trade-offs between general-purpose GPUs and custom ASICs for inference workloads will be critical as companies evaluate their AI infrastructure. AI for IT & Development Courses can provide the foundational knowledge needed to navigate these changes.
Your membership also unlocks:
