DeepSeek Is Building Its Own AI Chip – And It Just Sent NVIDIA Shares Falling

 On the morning of July 7, 2026, three sources familiar with the matter told Reuters something that moved markets within hours of publication.

DeepSeek — the Chinese AI company that shocked Silicon Valley eighteen months ago with a model that matched GPT-4 at a fraction of the cost — is now building its own AI chip.

By the time US markets opened, NVIDIA had fallen 1.5% in premarket trading. Micron Technology was down 4.7%. Western Digital had dropped 6.3%. SanDisk fell 4.6%. A single Reuters exclusive, citing unnamed sources, wiped billions from the market capitalization of the companies that have powered the AI revolution.

That reaction tells you something important — not just about DeepSeek, but about how fragile the assumptions underlying the current AI chip market actually are.

What DeepSeek Is Building — And Why the Distinction Matters

The chip DeepSeek is reportedly developing is not a training chip. This point matters enormously and gets lost in most coverage of the story.

Training a large AI model — the process of teaching it everything it knows — requires enormous, sustained computing power. It is the most compute-hungry stage of AI development, and it is where NVIDIA's most advanced GPUs have been indispensable. Training chips are extraordinarily difficult and expensive to design, and China's access to the manufacturing processes required to produce the best ones has been blocked by US export controls.

DeepSeek's reported chip is for inference — the stage where a trained model generates responses to user queries. Every time you type a message to an AI and it responds, that is inference. It is computationally less demanding than training, more predictable in its requirements, and — critically — much more commercially important in terms of volume. Every inference chip runs millions of user conversations per day. The market for inference compute is growing rapidly and, unlike training chips, is an area where a well-designed domestic solution could realistically compete.

This is not a moonshot. It is a calculated, practical move toward the part of the AI computing stack where DeepSeek can actually build something competitive given China's manufacturing constraints.

The Strategic Logic Behind the Decision

DeepSeek's decision to develop its own silicon makes sense the moment you understand the company's current situation.

DeepSeek has historically relied on NVIDIA's H800 and A100 chips — older models that were not subject to early export restrictions — for training, while increasingly using Huawei's Ascend processors for inference as NVIDIA hardware became harder to obtain. The company's own R2 model was reportedly delayed multiple times when training runs failed on Huawei hardware, forcing DeepSeek back onto NVIDIA GPUs for the training phase. The inference workload, however, has been running on Huawei's Ascend chips — which is precisely the division of labour that DeepSeek's new chip would serve.

By designing its own inference chip, DeepSeek would remove the remaining external dependency in its operational stack. It would no longer need Huawei's Ascend for the billions of user queries its models handle daily. It would control its own hardware roadmap. And it would have the ability to optimize the chip specifically for how its own models work — a significant performance advantage that general-purpose chips cannot offer.

The reported chip would be fabricated by Semiconductor Manufacturing International Corporation, China's largest foundry, rather than Taiwan's TSMC. SMIC has been cut off from the most advanced chipmaking tools by US and Dutch export controls, and is widely reported to be operating on a 7-nanometre process several generations behind the leading edge.

This manufacturing constraint is real and significant. A chip built on SMIC's current process will not match the performance density of chips built on TSMC's most advanced nodes. But for inference — where the computational requirements are more modest and more predictable than training — the gap may be workable. And for a company operating entirely within China, working with what is available is not a compromise. It is the only option.

The Market Reaction — What the Numbers Actually Mean

The stock market's immediate reaction to the Reuters report was sharp enough to be worth examining carefully.

NVIDIA fell 1.5% in premarket trading. Micron dropped 4.7%. Western Digital fell 6.3%. SanDisk was down 4.6%. These are not insignificant movements for large-cap companies — they represent billions of dollars in market value erased within hours of a single news report.

What drove that reaction? Not the chip itself — DeepSeek's chip is at an early stage of development, has not been announced officially, and may be years away from production at meaningful scale. The market reaction was driven by what the chip represents: a signal that DeepSeek is moving from being a consumer of AI hardware to a producer of it.

The companies that fell are overwhelmingly exposed to the inference market. Micron and Western Digital make memory chips — a critical component of inference systems at scale. If DeepSeek designs its own inference chip and optimizes it for its own memory architecture, the implication for US memory suppliers serving the Chinese AI market could be significant over the medium term.

The analyst community was divided in its reaction. Richard Windsor of Radio Free Mobile was direct in his assessment: "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." His point is technically valid — export controls on advanced manufacturing equipment mean DeepSeek's chip, built on SMIC's process, will not be competitive in international markets.

But the more relevant question for markets is not whether DeepSeek can sell chips globally. It is whether a DeepSeek inference chip could reduce Chinese AI companies' purchases of US-designed or US-supplied components for their domestic operations. On that question, the answer is: possibly, over time.

Where DeepSeek Is in This Process

The Reuters report is explicit about one thing that much of the subsequent coverage has softened: DeepSeek's effort to join the chip race remains at an early stage, with the company reaching out to external partners and holding discussions with chip-design, foundry and memory companies.

This is early-stage exploration, not an announced product. DeepSeek has not confirmed the report. No timeline has been provided. No specifications have been disclosed. The gap between "holding discussions with chip-design companies" and "shipping a competitive inference chip at scale" is measured in years and billions of dollars of engineering effort.

What is clear is that DeepSeek has the resources to pursue this seriously. DeepSeek's chip push coincides with the company's first embrace of outside capital. The company was slated to raise $7 billion in a maiden funding round valuing it at between $52 billion and $59 billion. That funding, combined with the existing engineering talent the company has assembled, means the chip development effort is not constrained by resources in the way that a smaller company's would be.

There is also precedent to take seriously. As early as 2025, industry publications reported that DeepSeek was recruiting chip-design talent. The Reuters report is not a sudden announcement — it is the confirmation of an effort that has been building quietly for at least a year.

The Bigger Picture — China's AI Stack Strategy

DeepSeek's chip development does not exist in isolation. It is part of a broader pattern across Chinese technology that has been accelerating since US export controls began tightening in 2022.

Alibaba, Baidu, and other major Chinese technology companies are all developing their own AI chips, motivated by the same fundamental concern: dependence on foreign hardware is a strategic vulnerability in an era of active technology competition between the US and China.

DeepSeek's potential in-house chip is part of a trend where global AI developers aim to gain more control over their hardware and decrease reliance on Nvidia. Other companies, like OpenAI and Anthropic, are also working on their AI chips due to similar motivations.

The difference between DeepSeek and Western AI companies pursuing custom chips is the geopolitical context. OpenAI or Anthropic developing custom silicon is a business optimization — they are trying to reduce costs and improve performance on hardware they can already access. DeepSeek developing custom silicon is partly a business optimization and partly a strategic necessity — the hardware they would otherwise prefer is no longer available to them.

This distinction shapes both the urgency and the constraints of the effort. DeepSeek's chip team is working with a harder constraint set than their Western counterparts, but with a clearer strategic imperative.

What This Means for the Global AI Market

The implications of DeepSeek's chip development — if it proceeds and eventually succeeds — are worth thinking through carefully.

For NVIDIA, the immediate impact is more psychological than operational. China has already been cut off from NVIDIA's most advanced chips. DeepSeek building its own inference chip does not remove NVIDIA from a market it can currently serve — it signals that the Chinese AI market is actively working to reduce the role of US hardware even for the older chips that remain accessible.

For Huawei, the implications are more direct. If successful, DeepSeek's expansion into semiconductor development would mark a major strategic shift for a company widely hailed in China as the country's AI champion, potentially adding to challenges faced by Chinese tech giant Huawei. Huawei has been the primary beneficiary of NVIDIA's exit from the premium Chinese AI chip market. A DeepSeek chip that handles inference better than Huawei's Ascend would represent genuine competition for one of Huawei's most important current revenue streams.

For the broader AI industry globally, the story reinforces a pattern that has been developing for eighteen months: the assumption that frontier AI requires access to a specific set of US-controlled hardware and manufacturing processes is being tested more seriously than anyone expected when DeepSeek first appeared in January 2025.

What to Watch For

DeepSeek has not confirmed the Reuters report. The official confirmation — if it comes — will likely arrive through one of a small number of channels: an official company announcement, a job posting that describes chip-design work explicitly, or a regulatory filing related to the funding round that references hardware development.

The manufacturing partner is the most consequential variable. SMIC's current process capabilities impose real performance limits. If DeepSeek can access more advanced manufacturing — through SMIC's continued development of its process nodes, through an alternative domestic foundry, or through manufacturing arrangements that are not currently public — the timeline and capability ceiling for the chip changes significantly.

The $7 billion funding round is also worth watching for specifics. When the round closes and the terms become public, the stated use of funds will provide the clearest official signal of where DeepSeek is prioritizing capital deployment. If hardware development appears in that list, the Reuters report moves from three unnamed sources to officially acknowledged strategy.

Final Thoughts

DeepSeek building its own AI chip is a logical step for a company that has consistently chosen to control its own destiny rather than depend on external suppliers. The company's entire story is one of working around constraints through engineering innovation — building frontier AI models on a budget that would not fund a comparable effort elsewhere, and now apparently attempting to build the hardware infrastructure to match.

Whether the chip succeeds at competitive scale is genuinely uncertain. The manufacturing constraints are real, the timeline is long, and chip design is a different discipline from model development. Succeeding at one does not guarantee success at the other.

But the intention matters regardless of the outcome. A company with $7 billion in funding, a proven track record of surprising the AI industry, and a clear strategic motivation to reduce hardware dependence is a more serious chip development effort than most coverage of this story has suggested.

The AI hardware market in 2026 is more contested, more geopolitically complex, and less predictable than it was eighteen months ago. DeepSeek's chip announcement is the latest evidence that this trend is not slowing.

Follow Future with AI for ongoing coverage of DeepSeek, the global AI chip market, and every development that shapes how artificial intelligence is built and deployed in 2026. New articles every week — grounded in primary sources, written for people who want real understanding.