Huawei is gearing up for a big swing in the AI chip wars. After a choppy 2025, the company plans to double output of its flagship Ascend 910C next year, targeting roughly 600,000 units, according to people familiar with the plans. That’s the clearest sign yet that Huawei — and manufacturing partner SMIC — are easing some of the bottlenecks created by US sanctions and trying to pry open a market Nvidia can’t fully serve right now.
The broader roadmap is even more ambitious. Across the Ascend lineup, Huawei is aiming to ship as many as 1.6 million dies in 2026, up from as many as 1 million this year. Internally, the company has budgeted for yield issues and is counting inventory it has already built, but the direction is unmistakable: scale up fast while China’s appetite for accelerators explodes. Big domestic customers — from Alibaba to upstarts like DeepSeek — need millions of chips to train and run AI services, and Nvidia’s constrained China footprint has left a gap. On its last earnings call, Nvidia said it logged no sales of the H20, a China-tailored part, and Beijing continues to discourage or block use of foreign accelerators on security grounds.
Huawei isn’t pretending it’s overtaken Nvidia. On single-chip performance, it hasn’t. Bernstein estimates the coming Ascend 950 delivers about 6% of the grunt of Nvidia’s next-gen VR200 “superchip.” The company’s top boxes today stitch two dies into one package — a power boost that’s tough to manufacture at scale and one reason supply stayed tight long enough for rivals like Cambricon to sell into the shortfall. Most Chinese hyperscalers still lean on Ascend for inference rather than full-blown training.
That’s the gap Huawei is racing to close. The 910C remains the workhorse, but a successor—nicknamed 910D in the industry — is slated to ramp late 2026, with an even bolder variant Huawei calls 950DT packing four dies into a single chipset. The company has also started talking openly about a three-year campaign to chip away at Nvidia’s dominance, rolling out Ascend 950, 960 and 970 in phases up to 2028, and a homegrown UnifiedBus interconnect designed to lash together as many as 15,488 Ascend chips in one fabric. It’s brute force, networking, and policy support — everything short of a node jump.
On that front, SMIC’s “enhanced” 7-nanometer process remains Huawei’s reality for the 910 line, while Nvidia’s Blackwell parts are fabbed at 4nm by TSMC — roughly two generations ahead. Huawei is exploring a beefed-up 7nm route that, if it works, could push Ascend into credible training workloads for tier-one clients. The caution is baked in: technical hurdles could push commercialization to 2027 or beyond.
For all the chest-thumping, the numbers keep it grounded. Internally, Huawei has told customers it can sell around 200,000 910Cs by year-end while reserving roughly 100,000 for its own cloud and state-linked projects. Next year’s target of 600,000 910Cs would mark a clean break from this year’s sanctions-scarred output. The older 910B is being phased out, while 2026’s total die count is set to surge as next-gen parts trickle in.
The strategic picture is straightforward. Nvidia is wobbly in China because of export controls and politics; Chinese demand for compute is soaring; domestic champions have to fill the gap even if their chips are less efficient. Huawei’s bet is to compensate with volume, packaging and networks, then grind toward better yields and bigger clusters. If it can hit its production marks in 2026, the company won’t have matched Nvidia on performance. But it won’t have to, not yet. It just needs to be good enough — and available.
The original story by Yuan Gao for Bloomberg.
The latest news in your social feeds
Subscribe to our social media platforms to stay tuned