Huawei Says Ascend Has Overtaken Nvidia in China, Pointing to Supply Rather Than Demand as the Limit

Overview

Huawei’s rotating chairman, Eric Xu, says the company’s Ascend line of neural processing units has overtaken Nvidia in China by market share. According to The Register, Xu made the claim in a Q&A-style release, and the figure is Huawei’s own estimate rather than independently audited data. Xu himself said that “It’s pretty hard to collect data about the market share of Nvidia in China,” before adding that “based on the data we have collected, Ascend has surpassed Nvidia.”

What We Know

• The claim and its origin. A transcript published by Geopolitechs says the Q&A took place at HUAWEI CONNECT on September 17, 2026, with Xu and Dr. Liao Heng, Chief Scientist of HiSilicon, taking questions from journalists.
• Supply is the constraint. Xu said that for the Ascend 950DT, which is aimed at AI training, SuperPoDs “are under testing, and their large-scale supply will start at the end of this year or early next year,” according to the Geopolitechs transcript. He added that “it’s not about how hard we are going to push model providers, but about how much supply capacity we will have.”
• Little exported. Xu said Huawei does not have “enough capacity to satisfy the demand in China” and has no plans “to expand the international market in a fully-fledged way,” as quoted by The Register.
• Scale over per-chip performance. Xu conceded that “our chips may be less advanced,” arguing that “at least their supply is assured,” per The Register. The outlet reports that Huawei is deploying a 256,000-card Atlas 950 SuperCluster and that its newer architecture is designed to scale to as many as one million NPUs.
• Nvidia’s China position. The Register reports that Nvidia has been allowed to sell H200-series parts to approved Chinese customers, but that on Nvidia’s second-quarter earnings call executives said the shipments amounted to less than 1 percent of its datacenter revenues. The outlet also reports that Chinese government officials began pressuring datacenter operators to move away from foreign accelerators.

Software Ecosystem

The Register says DeepSeek reportedly released software tools this week aimed at making Ascend accelerators more accessible to other AI labs. DeepSeek’s DeepGEMM-Ascend repository describes itself as a port of DeepGEMM to the Huawei Ascend platform that is “fully API-compatible with DeepGEMM” and supports BF16, FP8 and FP4 GEMM, MQA logits and MegaMoE. Its news entry for 2026.09.30 lists an initial release “with support for Ascend 950 devices.”

A Shortage Beyond China

Xu also framed the constraint as industry-wide. According to the Geopolitechs transcript, he said optical-module and storage suppliers are selling at a very high premium because of “a very severe shortage of supply compared to market demand,” and that global supply-demand balance “will be achieved somewhere around 2029,” with mainland China probably slower.

What We Don’t Know

• Huawei has not published the underlying data, the time period, or the definition of market share behind the claim, and Xu acknowledged the difficulty of measuring Nvidia’s share.
• It is unclear how much of Ascend’s volume comes from the inference-oriented parts already on the market versus the training-oriented 950DT, which Xu said is still under testing.
• No cited source gives Ascend shipment volumes or revenue.

Analysis

The claim, if accurate, would mean US export policy shifts since April 2025 have not restored Nvidia’s position in China, a trajectory The Register attributes in part to Beijing’s pressure on operators. Xu’s own framing, that capacity rather than demand limits Ascend, suggests the competitive question in China now turns on manufacturing and component supply rather than customer interest. That remains a company assertion, and independent share estimates would be needed to confirm it.