DeepSeek, one of China’s most prominent AI startups, is developing its own custom inference chip, joining a growing wave of AI companies seeking to reduce dependence on third-party silicon suppliers. According to a Reuters report published July 7, 2026, the Chinese lab has been exploring in-house chip design for roughly a year and has begun reaching out to chip design, foundry, and memory companies, while simultaneously recruiting experienced semiconductor engineers.
What Happened
Reuters reported that DeepSeek has been exploring the development of a proprietary inference chip for approximately one year. Unlike training chips, which require massive parallelism to build AI models from scratch, inference chips are optimized for the phase where a deployed model generates responses for end users — a workload that scales directly with user traffic and commercial revenue. The company has held discussions with chip design, foundry, and memory companies, and has been actively recruiting experienced semiconductor engineers to join its teams.
DeepSeek currently relies on Nvidia and Huawei chips for its operations, including Nvidia’s H800 GPUs — a downgraded export version of the H100 designed to comply with U.S. controls — and Huawei’s Ascend 910B accelerator for certain workloads. Building its own inference silicon would allow the company to sidestep both supply chains and gain full control over the hardware stack that powers its products.
Why It Matters
DeepSeek’s entry into chip design is significant for several reasons. The company became globally famous by releasing its R1 and V3 models at a fraction of the compute cost that Western peers had estimated necessary, suggesting the lab possesses genuine engineering talent capable of tackling novel hardware challenges. Owning the inference stack gives a company control over latency, cost, and ultimately its margin structure — the companies that control their own silicon will have a structural cost and performance advantage over those paying market rates for third-party chips.
The timing also reflects the intensifying global AI chip competition. DeepSeek is not the only AI lab pursuing custom hardware: Anthropic is in early discussions with Samsung to manufacture a custom 2nm AI accelerator, following OpenAI’s June unveiling of its co-designed Jalapeño inference chip built with Broadcom. With Nvidia holding an estimated 74 percent of the AI accelerator market, any large AI lab running at scale has a vested interest in finding alternatives.
Background and Context
For Chinese labs specifically, chip sovereignty carries an urgency that U.S. peers don’t face. Export controls have progressively tightened the supply of high-end Nvidia silicon available to Chinese companies, pushing firms like Huawei to accelerate their Ascend roadmap and encouraging AI labs to minimize external dependencies wherever possible. Chinese AI models have already captured over 30% of weekly token usage by U.S. companies, demonstrating that the gap in capability between Chinese and American labs has narrowed dramatically — even while the hardware gap remains a persistent challenge.
DeepSeek’s initiative follows a broader Chinese push toward AI self-sufficiency at the hardware level. Earlier this year, Meituan released LongCat-2.0, a 1.6-trillion-parameter AI model that was built entirely on Chinese chips — a milestone that demonstrated it is at least theoretically possible to train frontier-scale models without access to Nvidia’s top-tier silicon. Whether that approach can scale commercially and competitively remains an open question, but DeepSeek’s inference chip effort suggests that Chinese AI firms are pressing ahead on multiple hardware fronts simultaneously.
What Comes Next
DeepSeek’s chip effort remains at an early stage. The company has not committed to a specific architecture, fabrication partner, or production timeline, and the discussions with outside partners indicate the project is still in the exploratory phase rather than active tape-out preparation. Given the typical timeline for custom AI silicon — which ranges from two to four years from design to production volume — a DeepSeek chip reaching commercial deployment before 2028 would require unusually fast execution.
Nonetheless, the strategic intent is clear: as inference workloads grow with every new product feature and user interaction, the ability to run that inference on proprietary hardware becomes a meaningful competitive moat. The broader question for the global AI industry is whether this proliferation of custom silicon — from OpenAI, Anthropic, Google, Amazon, and now DeepSeek — will meaningfully erode Nvidia’s dominance, or whether the chip designer’s architectural lead will prove durable regardless of how many labs decide to build their own alternatives.
