DeepSeek Wants Its Own AI Chip. The Hard Part Isn't the Design.

DeepSeek Wants Its Own AI Chip. The Hard Part Isn't the Design.

China's breakout AI lab is quietly building an inference processor to cut its reliance on Nvidia and Huawei. U.S. export controls on advanced foundries and memory could decide whether it ever ships.

DeepSeek has begun designing its own artificial intelligence chip, people familiar with the effort told Reuters, a step that would loosen the Chinese startup's reliance on the Nvidia and Huawei processors it has used to build and serve its models.

The chip is meant for inference, the stage where a trained model responds to a user, rather than the heavier work of training a model from scratch, the people said. News of the plan pushed Nvidia's stock down about 2% in premarket trading on Tuesday.

A move into silicon would mark a turn for a company that made its name on lean, low-cost model design and has mostly stayed out of the business of selling products. It would also set DeepSeek against Huawei, the supplier Beijing has spent the better part of two years grooming as China's answer to Nvidia.

An inference-first bet

Starting with inference rather than training is a pragmatic call.

Inference chips are easier and cheaper to build. They sidestep the bleeding-edge manufacturing and the dense, high-speed links between processors that massive training runs demand. Inference is also the segment of AI computing growing fastest, as models move out of the lab and into products people use every day. Nvidia's own executives have said the volume of inference work jumped roughly tenfold over the past year. Running a model draws less raw compute than training one, but it rewards low cost per query and tight power budgets, which is what a purpose-built chip can deliver.

DeepSeek has an advantage that fits this approach. Its models already use a mixture-of-experts design that activates only a fraction of their parameters on each query, holding down the compute needed to serve them. A team that understands its own serving patterns in that much detail can shape a chip around them, the same logic OpenAI applied to its first processor.

Huawei has the most to lose

Over the past year, DeepSeek has been one of the loudest endorsements of Huawei's Ascend line.

When DeepSeek released its V4 model in April, it was tuned for Ascend chips and Huawei's own software framework rather than for Nvidia's. Huawei said its processors handled part of the training for V4-Flash, a lighter version of the model, and that its full Ascend cluster line was ready to run V4 the day it shipped. Orders for Huawei's Ascend 950 series jumped afterward, and reporting suggests prices for the chip climbed around 20% on the surge in demand.

Huawei sits on roughly half of China's domestic AI chip market, a market worth about $50 billion. That position was built partly on absence. U.S. export limits cut off Nvidia's most advanced parts, and Nvidia's share of the Chinese market fell from about 95% to 55% across 2025, according to IDC figures cited by Reuters. Chinese suppliers together took 41% of the market by shipments last year, with Huawei accounting for the largest single block.

Huawei's grip is already loosening, though, and Nvidia is not the reason. Alibaba's chip unit has rolled out a processor built for memory-hungry AI agents, and Baidu is training its large models on clusters of its own Kunlun chips. If DeepSeek now designs silicon of its own, it pulls one more marquee customer away from the champion Beijing has been building. The contradiction sits at the center of China's chip strategy. The state wants a single national supplier operating at scale, yet every major lab has its own reason to build hardware in-house. So far the labs are winning that argument, and each defection weakens the case for one champion.

A path others have already taken

DeepSeek would be joining a migration already underway among the world's largest AI developers.

OpenAI unveiled its first custom processor, Jalapeño, on June 24, built with Broadcom and aimed squarely at inference. The company said it took the chip from design to tape-out in about nine months, an unusually short cycle for custom silicon, and used its own models to speed the work. Broadcom's chief executive has put the early cost savings for inference at roughly half that of standard GPUs. Anthropic has been weighing a chip of its own, Reuters reported in April, and Google, Amazon, Meta, and Microsoft have leaned on homegrown accelerators for years. Even ByteDance, the parent of TikTok, has been in talks with Qualcomm this year to design custom chips for its data centers.

The reasoning is consistent across these firms. At the scale they now run, paying Nvidia's margins on every chip becomes a heavy recurring cost, and a processor tuned to one company's models can extract efficiency that a general-purpose GPU cannot. For DeepSeek the math carries an extra layer. U.S. rules block Chinese companies from buying Nvidia's top chips at all, and Beijing has been pressing its leading technology firms to build domestic replacements. DeepSeek's founder, Liang Wenfeng, said in a rare 2024 interview that export controls on chips were one of the company's real constraints.

Each of these programs nibbles at the same slice of Nvidia's business. Inference is the fastest-growing part of AI spending, and chips built for it can undercut general-purpose GPUs on cost and power draw. Nvidia still leads in training and holds a deep software advantage through CUDA, so its position is far from threatened. Its inference franchise, though, is where the custom-silicon wave breaks first, which is part of why the stock tends to slip on reports like Tuesday's.

The manufacturing wall

Designing a competitive chip is the easier half of the problem. Getting it built is where DeepSeek runs into a wall.

The same export controls that pushed DeepSeek toward a domestic chip also limit how that chip can be made. Washington bars Chinese designers from using the most advanced overseas foundries, which rules out Taiwan's TSMC and leaves SMIC, China's leading fab, as the realistic option. SMIC produces Huawei's current Ascend parts on a 7-nanometer-class process built without the extreme-ultraviolet lithography machines that the top foundries depend on, which caps how far the technology can be pushed.

The tighter squeeze is on memory. Inference performance leans heavily on high-bandwidth memory, the stacked chips that feed data to the processor fast enough to keep it working, and inference is limited by memory bandwidth more than by raw math. Separate U.S. curbs have cut China's access to that memory. A strong DeepSeek design could still stall against that shortage, and it is the constraint worth watching most closely.

None of this happens quickly. A competitive AI chip usually takes years and heavy capital to bring to market, and there is no guarantee DeepSeek clears the bar. The effort began roughly a year ago, one person said, and the company has been hiring chip-design engineers quietly, without posting the roles on public job boards.

Some analysts think the barrier is lower than it looks. One Barclays executive has said he would not be surprised to see China produce a low-cost, competitive chip within a year or two, much as DeepSeek did with its models when few expected it.

The prize is large enough to justify the gamble. One projection has China's AI chip market growing from about $21 billion in 2024 to nearly $200 billion by 2029.

From research lab to company

The chip push lands at the same moment DeepSeek is reinventing itself as a business.

In June the company closed its first outside funding round, about 51 billion yuan, or roughly $7.4 billion, at a valuation of between $52 billion and $59 billion. It was the largest single AI raise in China's history. Tencent led with around $1.5 billion and the battery maker CATL put in about $735 million, while Liang wrote the biggest check himself at roughly $2.9 billion. NetEase and JD.com joined as well, along with a state-backed AI fund.

The terms were unusual. Investors placed their money into a limited partnership that Liang manages rather than into DeepSeek directly. They accepted a five-year lockup and gave up voting rights. The structure lets the founder keep tight control even after taking billions from outside.

That reverses a stance DeepSeek held for years, when Liang refused outside capital and kept the company off any path toward an IPO or commercial products. The pressure to change came partly from talent. As rival Chinese labs watched their valuations climb, DeepSeek's small core of researchers held stock options that were not moving, and the company needed money to compete for engineers and to fund the data centers it is now building for itself.

DeepSeek is stepping into commercialization more openly, too. Its full V4 model is due in mid-July with a pricing scheme that charges different rates at peak and off-peak hours, a signal of a company starting to think about the cost of serving models at scale rather than only about research.

For now, the company is saying nothing. DeepSeek kept its customary silence on the report and did not respond to a request for comment, in line with the low profile it has held even as it became the public face of China's AI ambitions.