Huawei is trying to eat Nvidia’s lunch in China.
The Shenzhen outfit is set to take the largest slice of China’s AI chip market this year, with sales jumping by at least 60 per cent as firms hunt for domestic kit.
The dark satanic rumour mill has manufactured a hell-on-earth yarn claiming that Chinese tech groups have placed large orders for Huawei’s latest Ascend processor, the 950PR, as it pushes into a space long owned by the $5.1tn US chip group.
Huawei expects revenues from its AI chips to hit about $12bn this year based on orders already in hand, up from $7.5bn in 2025. Most of that demand is for the 950PR, which entered mass production in March.
An upgraded 950DT is pencilled in for the fourth quarter, as Huawei keeps shovelling capacity into the effort.
The push comes while Nvidia keeps tripping over regulators in Washington and Beijing, leaving Chinese buyers stuck in a weird limbo.
Nvidia chief Jensen Huang said in March that the company had finally received U.S. licenses to sell its H200 chips to China, yet no shipments have been made due to regulatory issues.
Beijing has told Chinese tech companies to support domestic producers and limit their use of Nvidia chips to overseas, one person said. US regulators require Nvidia chips ordered by Chinese clients to be used only in China, another person said.
That tug-of-war has reportedly meant no customs clearance for the H200 in China, which is required for shipments, so the parts sit in paperwork purgatory.
Huawei’s AI chip revenue forecast could rise further if its production ramp-up beats projections, the people said, which is the polite way of saying it wants to crank the handle harder.
The company makes most of its AI chips with Semiconductor Manufacturing International Corporation, China’s leading fab. Two additional fabrication plants dedicated to Huawei’s needs are scheduled to launch later this year.
Huawei’s latest chips still lag Nvidia’s most advanced products by at least two generations, but it has kept nudging performance and efficiency upward while the US brand fights export-control whack-a-mole.
Huawei is pushing the 950PR as the hardware of choice for domestic companies running inference on large language models used to generate responses.
It is betting inference, which is technically less challenging than model training, that will be the bigger demand engine as AI applications, such as agents, spread into everything that moves.
Huawei is leaning on networking gear to build larger computing clusters, trying to compensate for weaker single-chip performance with brute-force scale.
Earlier this month, DeepSeek said it worked with Huawei’s 950PR for its latest V4 model, which improves inference efficiency at reduced costs. Most of V4’s training still ran on Nvidia chips.
Huang said in a recent interview with podcaster Dwarkesh Patel: “The day that DeepSeek comes out on Huawei first, that is a horrible outcome for our nation. AI models around the world are developed and they run best on non-American hardware.”
A long-running headache for Nvidia’s Chinese rivals is software, because Cuda remains the default global toolkit and it is sticky in the way only developer ecosystems can be.
Huawei’s Cann arrived a decade after Cuda and draws complaints for being difficult to use, which pushes up operating costs. Developers say improvements are coming, but the gap to Cuda remains significant.
Morgan Stanley forecasts China’s AI chip market will reach $67bn in 2030, with 86 per cent expected to be supplied by Chinese players. It puts the market at about $21bn this year from domestic suppliers.
Nvidia previously dominated the Chinese AI chip sector, selling $17.1bn in fiscal year 2025, mostly from H20, a watered-down product built to satisfy US export controls.
Since the US banned sales of H20 to China a year ago, Nvidia has struggled to return with a chip that keeps both sides happy, which is not a design problem so much as a politics problem.
