DeepSeek Is Building Its Own Chip: What It Means When China's AI Champion Goes Full-Stack

DeepSeek Is Building Its Own Chip: What It Means When China's AI Champion Goes Full-Stack

On July 7, 2026, Reuters dropped a story that sent Nvidia shares sliding roughly 1.6% in premarket trading: DeepSeek, the Hangzhou-based AI lab that rattled Silicon Valley with its R1 model eighteen months ago, is now designing its own custom AI chip. Three people familiar with the matter told Reuters the effort has been underway for about a year, is focused on inference rather than training, and is expected to be fabricated by Semiconductor Manufacturing International Corporation (SMIC) — China's largest domestic foundry.

For a company that built its reputation on doing more with less — squeezing frontier-class performance out of constrained hardware — the move is a logical next step. But it is also a significant one. If DeepSeek succeeds, it would become the first Chinese AI lab to close the loop entirely: training its own models, running them on its own chips, and doing all of it without a single component that Washington can cut off.

"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," said analyst Richard Windsor of Radio Free Mobile — a reminder that the chip's strategic value is domestic, not global.

That framing, however, undersells the significance. A chip that works well enough for inference inside China is exactly what DeepSeek needs. And inference, not training, is where the AI industry's compute demand is now concentrated.

Why Inference, and Why Now

The distinction between training and inference matters enormously here. Training a frontier model — the process of teaching it on vast datasets — requires massive, tightly coupled GPU clusters and months of compute time. Inference is what happens every time a user sends a query: the model generates a response, typically on a smaller, more specialized chip. As AI applications proliferate, the ratio of inference to training compute is shifting fast. Industry analysts now estimate that roughly 70% of AI compute demand comes from inference, a share that will only grow as models get embedded into products, APIs, and enterprise workflows.

This is precisely the segment where purpose-built chips — Application-Specific Integrated Circuits, or ASICs — can outperform general-purpose GPUs on cost and power efficiency. It is also the segment where Chinese silicon is already closest to competitive. Huawei's Ascend 910C chips, for instance, are already being used for DeepSeek V4 inference at scale, and a Huawei-led team successfully post-trained the 1.6-trillion-parameter V4-Pro model↗ on a cluster of over 1,000 Ascend 910C chips — a milestone that would have seemed implausible two years ago.

DeepSeek's own trajectory illustrates the hardware division of labour that has emerged. Its R1 model was trained on Nvidia H800 GPUs — chips designed for the Chinese market that Washington subsequently banned. Since then, the lab has leaned increasingly on Huawei for inference while continuing to use whatever Nvidia hardware it can access for training. A custom inference chip would let DeepSeek own that second half of the stack outright.

The SMIC Question

The Reuters exclusive↗ names SMIC as the expected foundry partner, and that choice comes with real constraints. SMIC has been cut off from the most advanced chipmaking equipment — specifically the Extreme Ultraviolet (EUV) lithography machines made by ASML — by a combination of US and Dutch export controls. The foundry is currently operating primarily on a 7-nanometre process using older Deep Ultraviolet (DUV) immersion lithography with complex multi-patterning techniques.

The practical implications:

• Yield rates at SMIC's 7nm node are estimated between 20% and 70% depending on the application, well below what TSMC achieves at comparable nodes — meaning more wafers are scrapped per usable chip.
• Performance ceiling: A 7nm SMIC chip will not match the compute density or power efficiency of chips built on TSMC's 3nm or 2nm processes, which Nvidia and other Western chipmakers use for their latest accelerators.
• Cost structure: Multi-patterning is expensive and slow, inflating per-chip costs and limiting how quickly SMIC can scale production.
• Maintenance risk: SMIC's DUV machines require ongoing calibration and servicing; any future restrictions on foreign technicians could degrade yields rapidly.

None of this makes the project unviable. It makes it a domestic-market play, not a global one — which is consistent with The Next Web's analysis↗ that "a design is not a working chip, a working chip is not a shipping product, and taped-out silicon can still fail on the test bench." The honest read is that DeepSeek is trying to close the last gap in a fully Chinese AI stack, not to compete with Nvidia on the open market.

The Competitive Landscape: Everyone Is Building Chips Now

DeepSeek's move fits a broader pattern. The global AI industry is converging on the view that controlling your own inference silicon is a strategic necessity, not a luxury.

OpenAI made this explicit on June 24, 2026, when it unveiled Jalapeño↗, its first custom inference chip, developed with Broadcom in a nine-month design cycle. The chip is a reticle-sized ASIC with six High Bandwidth Memory modules, optimized specifically for OpenAI's own model workloads. Deployment is slated for late 2026, with full production ramp in 2027-2028. Anthropic has also been exploring custom silicon, Reuters reported in April.

Inside China, the race is equally intense. Alibaba's T-Head and Baidu's Kunlunxin chip units are both preparing for IPOs on the Hong Kong Stock Exchange, a signal that domestic AI silicon is now considered a standalone business, not just an internal cost centre. Baidu's Kunlun chips already power its ERNIE models, and analysts forecast Kunlunxin chip sales will increase six-fold in 2026, reaching approximately 8 billion yuan ($1.1 billion). Alibaba's T-Head processors have been reported to match the performance of Nvidia's H20 — the most advanced chip Washington currently permits for export to China — for certain workloads.

The pattern is clear: every major AI lab, East and West, is concluding that renting compute from a hardware vendor is a long-term liability. The question is not whether to build your own silicon, but how fast you can get there.

DeepSeek's entry into this race is notable precisely because the lab has historically been the most software-centric of China's AI champions. Its reputation was built on algorithmic efficiency — finding ways to train and run models that punch above their hardware weight. The decision to invest in chip design suggests the lab's leadership believes software optimisation alone can no longer close the gap with what purpose-built hardware can deliver.

What This Means for Huawei

The Reuters story notes that a successful DeepSeek chip "could add to challenges faced by Chinese tech giant Huawei" — and that framing deserves unpacking. Huawei currently holds approximately half of China's $50 billion domestic AI chip market, a position it built largely because US export controls eliminated Nvidia as a competitor. DeepSeek has been one of Huawei's most important customers, and the V4 model's optimisation for Huawei's Ascend ecosystem↗ has served as a high-profile reference workload that validated Ascend's capabilities to the broader market.

If DeepSeek begins running its own inference on its own chips, it reduces its dependence on Huawei — and potentially signals to other Chinese labs that vertical integration is the path forward. Alibaba and Baidu are already on that path. A DeepSeek chip would accelerate the fragmentation of what has been, until recently, a relatively captive domestic market for Huawei's Ascend line.

The Funding That Makes It Possible

One reason this project is credible now when it might not have been a year ago: money. In June 2026, DeepSeek closed its first-ever external funding round↗, raising approximately 51 billion yuan ($7.4 billion) at a valuation between $52 billion and $59 billion. The round included Tencent (~$1.5B), CATL (~$737M), JD.com, NetEase, IDG Capital, and the China National AI Industry Investment Fund. Founder Liang Wenfeng personally contributed 20 billion yuan (~$2.9 billion).

The deal structure was unusual: most investors placed capital into a limited partnership managed by Liang rather than investing directly in the company, and agreed to a five-year lock-up with no voting rights. The China National AI Industry Investment Fund was the sole exception, retaining both voting rights and liquidity — a detail that underscores the degree of state interest in DeepSeek's trajectory.

Chip design is capital-intensive and multi-year. The funding gives DeepSeek the runway to pursue it seriously. The lab has already begun quietly recruiting chip-design engineers↗ without public job postings, and is in discussions with chip-design firms, foundry partners, and memory suppliers.

Practical Takeaways for Developers and Buyers

For the global developer community currently building on DeepSeek's APIs, the near-term picture is unchanged. DeepSeek-V4-Pro and DeepSeek-V4-Flash remain the current flagship offerings, accessible via the DeepSeek Platform↗ and through third-party providers. The chip project is early-stage — no prototype, no confirmed tape-out date, no public benchmark. It will be years before any DeepSeek silicon reaches production.

What the announcement does signal:

• DeepSeek is not going anywhere. The combination of $7.4B in fresh capital, a chip programme, and a new data centre build in Inner Mongolia points to a lab that is scaling into a durable institution, not a one-hit wonder.
• The inference cost curve will keep falling. Purpose-built inference chips, whether from DeepSeek, Huawei, Alibaba, or Baidu, are designed to drive down the per-token cost of running large models. That is good news for anyone paying API bills.
• Supply chain risk is real but manageable. Developers relying on DeepSeek's API should note that the lab's hardware dependencies are shifting, not disappearing. The API documentation↗ remains the authoritative source for any service changes.
• The domestic Chinese AI stack is becoming genuinely self-contained. From model architecture to training hardware to inference silicon, China's leading labs are systematically closing the gaps that export controls were designed to exploit.

The chip story is, in one sense, a story about geopolitics: Washington's export controls created the incentive, and Beijing's capital created the means. But it is also a story about engineering ambition. DeepSeek built its reputation by refusing to accept that hardware constraints were a ceiling. Designing its own chip is the logical conclusion of that philosophy — and the global AI industry, East and West, is watching to see whether the lab can pull it off.

---

*Sources and further reading are listed below.*