Huawei accelerates AI chip roadmap, unveils Ascend 960 supernode to challenge Nvidia

Huawei used the opening keynote of its annual Huawei Connect conference in Shanghai on September 17 to pull forward the launch of its next flagship AI accelerator by three quarters, part of a broader push to close the computing gap with Nvidia as Beijing leans harder on domestic chipmakers to reduce reliance on U.S. silicon.

David Wang, Huawei's deputy chairman and rotating chairman, told the Huawei Connect 2026 audience that the company's Ascend 960DT processor, originally slated for the fourth quarter of 2027, will now ship in the first quarter of that year. A companion chip, the Ascend 960PR, moves up one quarter to the third quarter of 2027. Both are built to roughly double the performance of Huawei's current-generation Ascend 950 line.

What Huawei announced

In a keynote titled "Advancing the Agentic World, Building a Solid Silicon Foundation," Wang laid out a multi-year chip cadence that now runs on an annual release schedule: Ascend 970 in 2028 and Ascend 980 in 2029, with Huawei saying each generation will roughly double on the last. He framed the push in sweeping terms, telling the audience he expects artificial intelligence to be "the final technological revolution in human history."

Alongside the accelerated roadmap, Huawei unveiled the Atlas 960E SuperPoD, which it describes as the industry's first AI supernode built on near-packaged optics, a technique that mounts optical engines directly next to the processor instead of using conventional pluggable modules. The company also introduced Hi-ONE, the optical interconnect product underpinning that design, which Wang called "the industry's first NPO product ready for mass production, delivering the largest transmission capacity." Huawei separately detailed the OceanStor M900, a new memory-storage system aimed at speeding up AI inference, in a release carried on PR Newswire.

The numbers behind the pitch

• Atlas 960E SuperPoD: scales to 4,096 NPUs, 8 EFLOPS of FP8 compute and 1 petabyte of HBM capacity, using roughly 5,500 Hi-ONE optical engines in place of about 48,000 conventional 800G modules — a change Huawei says cuts power draw by more than 550 kilowatts.
• Hi-ONE interconnect: rated at 7.2 terabits per second of transmission capacity per engine.
• OceanStor M900: up to 64 petabytes of capacity per cluster and 40 terabytes per second of aggregate bandwidth, which Huawei says cuts AI-inference access latency from milliseconds to roughly 60 microseconds and doubles token throughput in typical workloads.
• Ecosystem scale: Huawei says its Kunpeng computing platform now counts 4.16 million registered developers and more than 7,200 partners, while its CANN software layer has roughly 5,200 monthly active developers, 61% of them outside Huawei.

How Huawei got here

The accelerated timeline lands against a backdrop of four years of on-again, off-again U.S. export restrictions on advanced AI chips sold into China. Washington has alternately tightened and eased licensing rules for Nvidia's China-bound accelerators — including the H20 and H200 — since 2022, with Beijing responding by pressing domestic cloud operators and AI labs to reduce their dependence on foreign silicon. Huawei has positioned its Ascend line, first launched in 2019, as the leading domestic alternative, betting that stacking large numbers of less-advanced chips into tightly networked "SuperPoD" clusters can offset a persistent gap in raw transistor performance against Nvidia's newest GPUs.

Huawei has reported that Chinese foundation models are approaching 10 trillion parameters and that daily inference-token consumption inside China is already in the hundreds of trillions, numbers it uses to argue that domestic compute demand — not just national-security politics — is driving the buildout. The announcement also arrives roughly a week before President Trump is scheduled to host Chinese leader Xi Jinping in Washington around September 24, a summit where technology and chip-export policy are expected to be on the agenda alongside trade.

Who is watching closely

The immediate audience is Chinese cloud providers and AI labs currently juggling a mix of Nvidia and Huawei hardware. Reporting from TrendForce indicates Chinese AI developer DeepSeek is planning to deploy at least 160,000 Ascend 950DT chips at a data center in Inner Mongolia, one of the clearest signs yet of a major domestic model developer committing at scale to Huawei silicon rather than treating it as a stopgap.

Analysts remain split on how much the acceleration actually closes the gap with Nvidia. As TechCrunch reported, independent semiconductor analyst Rui Ma flagged that the SuperPoD system Huawei detailed this week tops out at 4,096 chips, well short of the 15,488-chip configuration the company had previously described for a comparable cluster — a discrepancy Huawei did not directly address on stage. Coverage from DigiTimes likewise noted that, despite the faster cadence, major Chinese AI labs continue to rely heavily on Nvidia GPUs for the heaviest model-training workloads, reserving domestic chips largely for inference.

"We're evolving our Ascend chip series on a one-generation-a-year cycle," said David Wang, Huawei's deputy chairman and rotating chairman, in his Huawei Connect keynote. "In 2028 and 2029, we will roll out the Ascend 970 and 980 chips, respectively."

What happens next

Huawei says the Atlas 960E SuperPoD and the upgraded TaiShan 950 SuperPoD will begin rolling out through 2027 alongside the Ascend 960DT and 960PR chips, with Huawei Connect running through September 19 in Shanghai featuring further sessions on the company's software and open-source ecosystem. The real test will come once the 960-series chips ship in volume next year and Chinese labs report, in practice, whether Huawei's optical-interconnect and SuperPoD approach can substitute for raw per-chip performance at the scale needed to train frontier models — rather than just run them. That question will also sit in the background of the Trump-Xi summit later this month, where chip-export policy is likely to remain one of the more contested items on the agenda.