Huawei chairman claims Ascend NPU sales overtook Nvidia in China

Huawei rotating chairman Eric Xu used a Q&A release this week to claim that the company's Ascend line of neural processing units has outsold Nvidia inside China. No third party has verified the figure. "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," Xu said, according to The Register, which reported the remarks on 30 September.

The claim lands as the policy ground keeps shifting beneath the US chip export control regime, and as Nvidia's own numbers point the other way in revenue terms. During the company's Q2 earnings call, executives said shipments of its H200 accelerators to China amounted to less than 1 percent of datacenter revenues, per The Register's account. That is the gap between unit share in a protected market and dollars in the bank: a company can lead on cards shipped into China while the same cards barely register in its global revenue line, and the two figures are not measuring the same thing.

DeepSeek opens the software stack

The hardware claim would be hollow without software. That is where the week's more concrete development sits.

On Wednesday, Hangzhou-based DeepSeek open-sourced six software modules tailored for Huawei's Ascend AI chips, the South China Morning Post reported on 30 September. The company described the release as a step toward an "independent and controllable" software ecosystem, mirroring its earlier open-source tools for Nvidia hardware. Among the releases is an Ascend-compatible version of TileLang, a programming language designed to streamline development of high-performance computation kernels. TileLang lists Nvidia as its primary back end, but now officially supports Huawei's Ascend 950 accelerators with "native code generation, automatic scheduling, and synchronisation," according to the project's GitHub page as cited by the SCMP. That is a developer-tools move, not a benchmark, and its effect on export controls is indirect at best.

Xu was candid about the limits. "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," he said. On international expansion, he was blunter: "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." The Register notes Huawei is deploying a 256,000-card Atlas 950 SuperCluster and designing a newer architecture to scale to as many as one million NPUs.

Washington's on-again, off-again controls

The policy record behind that supply anxiety is messy. The Trump administration imposed new licensing requirements on Nvidia's and AMD's China-spec accelerators in April 2025, reversed course that summer, and later reached an arrangement to cut the US government in on a percentage of certain China-bound chip sales. By late last year it allowed Nvidia to sell H200-series parts to approved Chinese customers for the first time, The Register reported. Beijing had already begun pressuring datacenter operators to move away from foreign accelerators.

That pressure is now visible in the lobbying war around the annual defense bill. Nvidia and AMD have asked the White House to defuse China chip curbs in the defense legislation, according to reports from TradingView and Benzinga on 29 September, while ASML issued a public warning about China chip export limits, simplywall.st reported the same day. The headlines are running ahead of any confirmed legislative text.

The export control question has also drifted away from the diplomatic table. US export controls on chips were not a key topic at the Xi-Trump summit, the South China Morning Post reported on 26 September, and US trade representative Greer said Washington did not discuss chip export controls with China, per a 27 September account. Those are background, but they explain why the chip fight is being fought in company announcements rather than communiques.

Where the American side is moving

US vendors are hedging in two directions. OpenAI and Synopsys signed a multi-year partnership to build a specialized chip-design model called GPT-Synopsys, combining OpenAI's AI technology with Synopsys' electronic design automation tools, the-decoder.com reported on 30 September. OpenAI is already working with Broadcom on chips built for running AI models, and the recently unveiled Jalapeno chip is described as competitive against similar specialized parts.

Startups are chasing the same yield problem. Emergence AI is moving its neuroformal AI technology into active deployments with fabless semiconductor companies, EE Times reported on 30 September, with co-founder Satya Nitta saying "you cannot make chips any faster, but what you can do is definitely get more chips per wafer yielding with AI." Nitta said the approach pairs LLMs with symbolic AI so that "algorithms propose, and symbolic AI verifies." He said IDM customers are asking the company to extend the work into the fab.

Meanwhile, the politics of AI safety got a voluntary layer. President Trump and the leaders of the largest US AI companies signed a set of voluntary safety standards at the White House, CBS News reported on 29 September, with Trump calling the accord "morally binding" and saying it is "almost like a constitution, in a way." The companies committed to "four layers of controls and audits" including internal evaluations, external audits and board reviews. Trump also signed an executive order directing the federal government to use the term "Super Intelligence" instead of artificial intelligence, and said he plans to name an AI czar within three or four days.

None of that changes the arithmetic Xu is describing. Huawei's pitch is not that its chips are better, but that they cannot be switched off by a US licensing decision. DeepSeek's open-source modules are an attempt to make that pitch credible to developers. Whether Ascend has actually outsold Nvidia in China remains a company claim without an independent audit, and Nvidia has not published a China-specific unit number to test it against.