DeepSeek Is Building Its Own AI Chip: China's AI Champion Makes a Strategic Pivot Into Custom Silicon

Posted on 18th Jul 2026 06:07:03 in Artificial Intelligence, Machine Learning

Tagged as: DeepSeek, AI chip, custom silicon, inference processor, semiconductor, Nvidia, Huawei, artificial intelligence, China AI

Chinese AI startup DeepSeek is developing its own custom artificial intelligence chip, marking the company's most significant strategic shift since its models went viral and shook global technology markets in early 2025. According to an exclusive Reuters report citing three people familiar with the effort, the Hangzhou-based company has spent the past year quietly designing a processor purpose-built for AI inference — the stage where trained models generate responses for users — as it seeks to reduce dependence on both Nvidia and Huawei hardware.

The development, confirmed by multiple independent sources, places DeepSeek alongside OpenAI, Anthropic, Google, Amazon, and other AI leaders investing billions in custom silicon. It also signals that China's most celebrated AI lab now sees hardware as inseparable from its core model strategy — and that it no longer trusts the supply chain to meet its needs.

From Model Lab to Chip Designer: A Year in the Making

DeepSeek's chip effort began approximately one year ago and remains at an early stage, the sources told Reuters. The company has been discreetly reaching out to external partners across the semiconductor ecosystem — chip design firms, foundries, and memory suppliers — while privately recruiting chip-design engineers without advertising positions on public hiring platforms.

This quiet approach is characteristic of DeepSeek. Despite becoming a standard-bearer for China's AI ambitions, the company has kept an unusually low profile, rarely speaking to the press and declining to comment on the Reuters report. Founder Liang Wenfeng gave a rare interview in 2024 in which he acknowledged that US chip export controls posed a significant challenge. A custom chip, if successful, would be DeepSeek's most direct answer to that challenge yet.

The timing is not coincidental. DeepSeek was slated to raise $7 billion in a maiden funding round in June 2026, according to Reuters, valuing the company between $52 billion and $59 billion and reversing its long-standing policy of rejecting external investment. That capital injection gives the company resources to fund what is typically an expensive, multi-year semiconductor design programme.

Why Inference — And Why the Economics Are Shifting

The chip DeepSeek is designing is specifically aimed at inference, not training. The distinction matters. Training an AI model — the process of feeding it billions of examples so it learns patterns — still relies on enormous clusters of general-purpose GPUs. But once a model is trained, every question a user asks, every document a coding assistant completes, and every decision an autonomous agent makes must be processed through inference hardware. As AI applications spread across chatbots, enterprise tools, coding assistants, and autonomous agents, inference has become the fastest-growing and most economically significant segment of AI computing.

Specialised inference chips can run these workloads at a fraction of the cost and power consumption of general-purpose GPUs. For a company like DeepSeek — whose models serve millions of users globally and whose January 2025 release briefly shifted the industry's attention toward cheaper alternatives — inference economics are central to competitiveness. Chinese AI models' share of global web traffic rose from roughly 3% to 13% within two months of that release, according to Jon Peddie Research, and while the surge has cooled, the signal was unmistakable: if performance is close enough, users choose based on price.

DeepSeek has used both Nvidia and Huawei chips to date. The foundation model behind R1 — the reasoning model whose low-cost performance triggered a rout in US technology stocks in January 2025 — was trained on Nvidia's H800, a China-market chip that Washington subsequently banned in late 2023. The company has since leaned increasingly on Huawei's Ascend platform, releasing a V4 model adapted for Huawei silicon in April 2026. A custom inference chip would add a third path — one DeepSeek controls directly.

The Manufacturing Wall: Designing Silicon Is the Easy Part

For all the ambition, the hardest part of DeepSeek's chip project may not be the design. As Hardware Busters noted in its analysis of the Reuters report, modern design tools — some of them AI-assisted — have made chip layout faster and more accessible than ever. The bottleneck lies downstream.

One company, TSMC, handles the overwhelming majority of leading-edge semiconductor manufacturing globally. Samsung absorbs some overflow, Intel is years into a foundry turnaround, and every fabrication plant depends on lithography machines that only ASML in the Netherlands knows how to build. On top of wafer capacity, advanced packaging technology — the CoWoS-style processes that stitch compute dies to high-bandwidth memory — is similarly concentrated in a handful of suppliers.

"If you're just starting to design a chip right now, you won't see silicon for three years," Bernstein analyst Stacy Rasgon told Axios. That timeline matters because three years is a geological age in the AI market, where frontier models can be displaced within months.

DeepSeek also faces specific geopolitical constraints. US export controls bar Chinese chip designers from accessing the most advanced overseas foundries, while separate restrictions have cut China's access to high-bandwidth memory — a component critical to AI inference processors. These are not obstacles DeepSeek can engineer around with clever model design; they are physical supply-chain walls.

China's $50 Billion AI Chip Market Gets a New Contender

US export restrictions created an opening for Huawei in China. The Shenzhen-based technology giant now supplies roughly half of China's estimated $50 billion domestic AI chip market, serving DeepSeek alongside most major Chinese technology companies. Orders for Huawei's Ascend 950 processors surged after the DeepSeek V4 launch, Reuters reported.

But Huawei's lead is not guaranteed. Alibaba and Baidu have each invested heavily in their own AI processors, giving China's largest cloud providers direct control over the chips running their AI platforms. DeepSeek's entry would add another domestic competitor to a market that is becoming less dependent on any single supplier with each passing year.

"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," analyst Richard Windsor of Radio Free Mobile told Reuters. His assessment points to the defining reality of this effort: it is primarily about the domestic Chinese market and about supply assurance, not about challenging Nvidia on the global stage.

Jon Peddie Research, which tracks 150 companies offering 292 AI processors, noted that the main competitive impact may land on Huawei rather than Nvidia. Chinese hyperscalers developing their own processors weaken Huawei's position in the one market where US export controls had given it room to grow.

The Global Custom Silicon Arms Race

DeepSeek's move is part of a broader industry trend. OpenAI last month unveiled Jalapeno, its first custom inference chip developed in partnership with Broadcom. Anthropic is reportedly weighing its own chip programme, according to Reuters reporting from April 2026. Google has built multiple generations of its Tensor Processing Units. Amazon's Trainium and Inferentia chips power AWS. Microsoft is developing its own accelerator silicon. Meta has a custom chip programme. Apple has been designing its own processors for years and is reportedly extending that expertise to AI servers.

The common thread is not necessarily performance. Nvidia's GPUs and its CUDA software ecosystem remain difficult to displace for many workloads. Rather, it is about economics and leverage. "Custom silicon can shave inference costs, but Nvidia's networking and software stack is genuinely hard to walk away from," Hardware Busters wrote. "Designing your own chip mostly buys leverage in the negotiating room — something in my pocket when I'm sitting across the table from Jensen."

For DeepSeek, the calculation has an additional geopolitical layer. Even if US-China relations were to normalise — an unlikely scenario given current tensions — the experience of the past three years has taught Chinese AI companies that supply-chain dependence on foreign chipmakers is a strategic vulnerability. Building domestic alternatives, however challenging and however slow, is increasingly seen as a non-negotiable cost of competing in AI.

The Jon Peddie Research analysis also flagged a longer-term opportunity that extends beyond China's borders. Much of Africa and Asia remains AI-poor — not because demand is absent, but because cost, cloud access, infrastructure, language support, and sovereign-control concerns make frontier US AI difficult to adopt at scale. Low-cost Chinese models running on domestic Chinese silicon could become attractive in those markets if they reach a threshold of being good enough, cheap enough, and accessible. For countries and companies outside the United States, the question is no longer only who has the best model — it is who can be trusted to keep providing access.

DeepSeek's chip effort will not produce silicon tomorrow, or next year, or likely the year after. But the direction of travel is now unmistakable. China's AI champion has concluded that the models of the future will need hardware built specifically for them — and it has started building.

Sources

• Reuters — Exclusive: China's DeepSeek developing its own AI chip, sources say (July 7, 2026)
• Hardware Busters — DeepSeek Is Building Its Own AI Chip — But Designing Silicon Was Never the Bottleneck (July 9, 2026)
• Technology.org — DeepSeek Builds Its Own AI Inference Chip (July 8, 2026)
• TechStartups — DeepSeek is building its own AI chip to cut reliance on Nvidia and Huawei (July 7, 2026)
• Jon Peddie Research — DeepSeek moves toward its own AI chip (July 8, 2026)