Huawei plans three-year campaign to overtake Nvidia in AI chips

On Thursday, the company trumpeted its grand plan with all the drama of an Nvidia launch. Xu took the stage to present the next generation of AI chips, twinned with its upgraded “SuperPod” designs — a term borrowed from Nvidia’s own playbook that refers to a data center platform that encompasses computing, storage, networking, software and infrastructure management technologies.

In theory, the technique lets Huawei link as many as 15,488 of its Ascend-branded AI chips using self-developed UnifiedBus interconnect protocol, a new technology also formally unveiled on Thursday.

That’s the rough equivalent of overwhelming an enemy force through sheer numbers, and its prowess is further enhanced by far faster data transmissions between individual chips — as much as 62 times quicker than Nvidia’s upcoming NVLink144 technology, Huawei claims. Meanwhile, Nvidia’s current-generation NVLink72 technology allows the company to connect 72 Blackwell graphic processing units and 36 Grace central processing units together.

“Huawei’s willingness to publicly articulate its AI roadmap represents a strong signal of confidence in the resilience of its future local foundry supply,” Bernstein analysts led by Qingyuan Lin said in a note published on Monday. “These developments indicate that Huawei has secured reliable manufacturing capabilities to support its ambitious AI plans, marking a significant milestone in building a robust local semiconductor ecosystem capable of withstanding global supply chain disruptions.”

Huawei’s announcement coincided with a recent plethora of revelations about advances in AI chips led by Chinese firms from Alibaba Group Holding Ltd. to Baidu Inc. The drumbeat of news is remarkable given most of the country’s firms have for years kept their cutting-edge technology a secret, to avoid drawing Washington’s scrutiny. They’re emerging in rapid succession as Beijing puts chip policy at the center of delicate talks with the US.

Washington for years has tried to ringfence China, for fear that US technology will further its economic and military ambitions. In response, Beijing has exhorted the country’s tech firms to climb the value chain. The Chinese government recently launched a series of moves targeting Nvidia, including a directive to stop Chinese firms from buying certain components supplied by the company.

Earlier this year, China’s Xi showed that his country is prioritizing support for strategic sectors when he met with a small group of entrepreneurs, including Huawei founder Ren Zhengfei. Senior Chinese officials have also repeatedly pledged to harness the entire nation’s resources to speed homegrown technologies breakthroughs.

Chinese chips generally remain inferior to those from Nvidia or Advanced Micro Devices Inc., both of which get to use the world’s most cutting-edge chipmaking processes offered by TSMC. Bernstein analysts noted that a single next-generation Ascend 950 can only offer 6% the performance of Nvidia’s upcoming VR200 superchip.

But in recent years, Chinese firms have found innovative solutions to get around power and performance ceilings. Huawei talked Thursday about how it can group or cluster up to a million chips, theoretically bridging that gap.

To build the so-called super cluster, Huawei plans to ramp up connection and memory access speeds within chips. It claims a self-designed high-bandwidth memory architecture that can beef up a processor’s ability to transfer data between different components. That’s despite US restrictions that cut Huawei’s ties to memory industry leaders like SK Hynix Inc.

On Thursday, Huawei also touted its technology for connecting multiple chips with blazingly fast speeds for sharing data, a core competency for the company because of its long history as the leader in communications equipment for telecom operators. Xu claimed the Ascend 970 slated for 2028 will sport a 4 terabit-per-second interconnect speed. Nvidia currently offers an interconnect speed of 1.8 terabit-per-second.

Here’re the key details from Huawei’s conference:

It remains to be seen whether any of Huawei’s claims will hold up over time — and whether their designs can be produced at scale. Huawei surprised the tech world by introducing a 7-nanometer chip to power its Mate 60 Pro smartphone in 2023, but it hasn’t been able to advance beyond that level since. Xu on Thursday also did not elaborate on how the company will manufacture the new chips — a potentially serious choke point.

Nvidia’s dominance of the market is such that even AMD and Intel Corp. are relegated to also-rans in the hot AI space. Elsewhere, companies like ASML Holding NV dominate high-end chip equipment. TSMC makes most of the world’s most sophisticated processors — and is technically barred from doing business with a swath of Chinese firms including Huawei sanctioned by the US government.

“Huawei’s new chips, in our view, are uncertain, since its plan last year to roll out Ascend 910D using 5nm has not materialized due to poor yield,” Jefferies analysts led by Edison Lee said in a note on Thursday, saying that a lack of advanced chipmaking equipment remains China’s biggest hurdle in reducing dependence on Nvidia chips.

Yet Xu didn’t sound fazed though he did hint at prolonged technological bottlenecks. The executive talked constantly about how Huawei intended to supplant Nvidia’s place in the industry.

“We believe only by relying on ‘SuperPod’ and cluster technology can we achieve a breakthrough in the constraints we face in chipmaking process technology and supply endless computing support for our country’s AI development,” Xu told the official Xinhua news agency last week.