Huawei has unveiled a three-year campaign to challenge Nvidia’s grip on AI chips, betting on massive clusters and faster connections to close the gap.
The China-based tech giant claims its new “SuperPod” systems can link tens of thousands of Ascend processors, boasting data transfers 62 times quicker than Nvidia’s next-generation products.
Bloomberg reported on the rare public roadmap that was laid out at Huawei’s annual Connect conference, where executives detailed upcoming chip models and infrastructure designs in a bid to outpace Nvidia’s dominance in AI hardware.
Blazingly fast connections across thousands of chips
Rotating chairman Eric Xu said the push will rely on a new UnifiedBus protocol, designed to knit together as many as 15,488 Ascend processors inside a single SuperPod. Huawei claims those clusters can move information at unprecedented speeds, a headline boast meant to counter the American rival’s edge in raw chip power.
Chip rollout from 2025 through 2027
Huawei laid out a year-by-year schedule for its Ascend line. The roadmap begins with the Ascend 910C chip and Atlas 900A3 SuperPod — expected to arrive in 2025 — capable of linking up to 384 processors.
In 2026, the company plans to introduce the Ascend 950 and scale clusters to 8,192 chips.
By 2027, Huawei expects the Ascend 960 to further push those numbers, tying together as many as 15,488 units in a single system. Each step is designed to show steady progress toward narrowing Nvidia’s commanding lead in AI hardware.
Experts doubt Huawei can match Nvidia’s raw performance edge
Analysts were quick to point out limitations in Huawei’s bold plan.
Bernstein’s Qingyuan Lin told Bloomberg that the roadmap shows “a strong signal of confidence” in China’s local supply chain, but also underscored how far Huawei still lags behind its US rival.
Jefferies analysts led by Edison Lee were more blunt, saying Huawei’s new chips are “uncertain” after last year’s attempt to roll out the Ascend 910D on 5nm collapsed due to poor yields. They warned that a lack of advanced chipmaking equipment remains China’s biggest obstacle to breaking free from Nvidia.
Performance gaps also remain stark. Bernstein noted that a single next-generation Ascend 950 would only deliver about 6% of the computing power of Nvidia’s upcoming VR200 superchip. That disparity highlights why Huawei is betting on clustering massive numbers of chips rather than competing head-to-head on raw power.
Nvidia locks in massive capacity as Huawei plays catch-up
While Huawei is mapping out its roadmap, Nvidia is pushing ahead with record-breaking hardware and its own deals.
The company has unveiled the Rubin CPX platform, a new class of GPU set to launch in 2026 and built to handle massive-context AI tasks such as large-scale coding and video generation.
At the same time, Nvidia signed a letter of intent with OpenAI for a partnership worth up to $100 billion, committing at least 10 gigawatts of compute capacity in future data centers. The deal secures millions of GPUs for one of the world’s most prominent AI players, showing how far ahead Nvidia remains even as Huawei pushes to close the gap.
China bans Nvidia chips as it backs homegrown alternatives
Huawei’s campaign is also playing out against a shifting policy backdrop. Beijing has barred major firms, including TikTok owner ByteDance, from buying Nvidia’s RTX Pro 6000D chips built for the local market, part of a broader effort to sideline US suppliers and boost domestic champions.
Huawei has already benefited. Its work with Zhejiang University on the DeepSeek-R1-Safe model put Ascend processors at the center of training, a swap Chinese officials point to as proof local chips can step in where Nvidia once dominated.
With Washington tightening restrictions and Beijing closing doors at home, Huawei’s roadmap has become a rallying point of China’s bid to break free from US technology.
The rivalry with Huawei isn’t Nvidia’s only move; it’s also putting $5 billion into Intel to shape the next generation of chips.