Huawei claims Ascend outsells Nvidia in China as export controls reshape the market

In a Q&A-style release this week, Huawei's rotating chairman Eric Xu said its Ascend line of neural processing units had exceeded Nvidia in Chinese market share. "It's pretty hard to collect data about the market share of Nvidia in China, but based on the data we have collected, Ascend has surpassed Nvidia," he said, according to The Register, which published the remarks on 30 September. No third-party firm has confirmed the figure.

Xu framed the shift as a matter of survival rather than pure performance. "Chinese people have a keen sense of urgency, and will not accept a future in which others decide whether or not we can have access to certain products," he said. That line is the clearest statement yet of how US export controls have become a sales argument inside China. It also sidesteps the question of raw performance, where Nvidia's top parts still lead on most published benchmarks, and where Huawei has not released comparable throughput numbers for Ascend. The claim, in other words, is about availability and political reliability, not about who builds the fastest chip.

The policy backdrop is messy. The Trump administration imposed new licensing requirements on Nvidia's and AMD's China-spec accelerators in April 2025, then reversed course that summer, later reaching an arrangement to cut the US government in on a percentage of certain China-bound chip sales. By late last year, the administration said Nvidia could sell its more potent H200-series parts to approved Chinese customers for the first time, The Register reported.

Supply, not just speed

Huawei's pitch is blunt: the chips may be less advanced, but they will not be banned at short notice. "Even though our chips may be less advanced, at least their supply is assured, so that you don't have to worry about chip supply day in and day out," Xu said. That is a direct answer to the whiplash Chinese buyers experienced in 2025, when H20 sales were blocked and then partially restored.

The numbers behind Nvidia's China business remain small. During Nvidia's Q2 earnings call, executives said shipments of H200 accelerators to the region amounted to less than 1 percent of its datacenter revenues. Beijing, meanwhile, has pressured datacenter operators to move away from foreign accelerators in favour of homegrown alternatives, according to The Register.

Huawei is compensating for any per-chip deficit with scale. It is deploying a 256,000-card Atlas 950 SuperCluster, The Register reported, while a newer architecture is designed to scale to as many as one million NPUs for training and inference on multi-trillion-parameter models. Xu was careful not to overstate the export threat: "We don't have enough capacity to satisfy the demand in China. We don't have plans to expand the international market in a fully-fledged way."

Software, not silicon, has been the harder wall. Liao Heng, chief scientist of Huawei's HiSilicon, said the first barrier Ascend encountered was the ecosystem, not the hardware. "Two years ago there was still a tremendous software barrier for Ascend," he said, pointing to what is often called the CUDA moat. Nvidia's dominance rests partly on an estimated four million developers worldwide who build with CUDA, a figure cited by The Decoder on 30 September.

DeepSeek opens the toolbox

The most concrete move against that moat came from DeepSeek on 30 September. The Chinese AI developer released open-source programming tools for Huawei's Ascend chips, according to a post on its official WeChat channel, with Reuters reporting the details. The centrepiece is TileLang, a language developed by researchers at Peking University that DeepSeek has used for about a year and which it argues is simpler to program than CUDA while still extracting full hardware performance.

DeepSeek says Huawei "fully supported" the work. The two companies also optimised a so-called supernode, a cluster of 128 Ascend 950 chips. The release includes libraries for computation and for moving data between chips, and all of it is open source. DeepSeek first tested TileLang on older Nvidia chips; it is now the company's main tool for its AGI work, according to The New York Times, as cited by The Decoder.

That matters because software lock-in is the part of Nvidia's position that hardware rivals have struggled to replicate. AMD has not crossed the moat even when its silicon looked competitive on paper, The Decoder noted. Research firm SemiAnalysis, after testing OpenAI's Jalapeño inference chip, called the CUDA moat "potentially dead" because OpenAI gets new models running on its own hardware so quickly. Jalapeño beat Nvidia's Blackwell on performance per watt in most scenarios tested, though SemiAnalysis cautioned that the tests covered only relatively easy-to-optimise workloads of about 8,000 input tokens and 1,000 output tokens.

Huawei's chips were not part of that comparison. In an earlier analysis of DeepSeek V4, SemiAnalysis noted that Huawei's CANN software stack was the only one besides CUDA to support the model on day one, a detail that undercuts the idea that Chinese software is uniformly behind.

Design tools move too

The AI-in-chip-design push is not confined to China. On 30 September, OpenAI and Synopsys signed a multi-year strategic partnership to build a specialised model called GPT-Synopsys, according to The Decoder. Synopsys makes electronic design automation tools, the software engineers use to design chips. The model is meant to reason about chip design and verification and to operate Synopsys' tools directly, with engineers delegating objectives and approving output. Early tests with semiconductor customers are already underway, and the two companies will market the product together and share revenue.

Tom's Hardware is running an AI Chip Design Week from 28 September to 2 October, with coverage noting that Cadence, Synopsys and Siemens all now offer agentic AI for chip design, largely built on Nvidia's stack and with varying claims of autonomy. Synopsys unveiled seven AgentEngineer agents on its new Autopilot platform, with general availability planned for the end of 2026, according to Tom's Hardware.

The controls themselves remain contested. On 24 September, Tom's Hardware reported that Dylan Patel, head of semiconductor analysis firm SemiAnalysis, accused AMD of effectively dumping export-controlled military-grade silicon in China after a Chinese firm, Puzhi Electronic Technology, crowdfunded a system based on AMD's Zynq UltraScale+ XCZU47DR RFSoC. Patel wrote on X that the chip is quoted at $1,000 in China against a $36,000 US list price, and called for a treason investigation. AMD denied selling or shipping the chips directly into the restricted region, telling Tom's Hardware that any shipment into restricted regions violates its policy and that it investigates suspected diversion cases.

For all the export-control activity, the picture that emerges is less about a single choke point than about parallel supply chains. Huawei says it cannot meet domestic demand and will sell only a limited number of chips abroad; DeepSeek is giving away the software layer; and US design-tool vendors are folding AI into their own products. The controls have not stopped Chinese substitution. They have given it a deadline.