Huawei pulls its next AI chips forward by up to three quarters

Huawei is pulling its next generation of AI chips forward by up to three quarters, rotating chairman David Wang announced Thursday at Huawei Connect 2026 in Shanghai — a sign the Chinese tech giant is racing to close the gap with Nvidia and AMD faster than its own roadmap originally called for.

What's Moving Up and By How Much

The Ascend 960DT, a chip aimed at decoding and training workloads with expanded memory capacity and bandwidth, is now targeted for readiness in the first quarter of 2027 — three quarters earlier than its previous schedule. The Ascend 960PR, built for inference prefill and recommendation workloads, moves up one quarter to the third quarter of 2027. Both chips follow what Huawei calls "Tao's Law," an internal design philosophy that emphasizes doubling compute scale while significantly boosting memory bandwidth, capacity, and interconnect bandwidth simultaneously, rather than treating those as separate upgrade tracks.

Alongside the chips, Huawei detailed two SuperPoD system configurations: the liquid-cooled Atlas 960 SuperPoD, arriving in the third quarter of 2027 as the first Huawei system to adopt near-packaged optics (NPO) interconnects, and the air-cooled Atlas 860 SuperPoD, slated for the second quarter of 2027. Wang said Huawei will maintain a "one-generation-per-year cadence" for Ascend going forward, with the Ascend 970 slated for 2028 and the 980 for 2029.

The Competitive Context

The accelerated timeline lands Huawei's chips roughly alongside competing roadmaps from Nvidia and AMD. Nvidia's Rubin is set for second-quarter 2026 mass production with HBM4 memory, followed by Rubin Ultra with HBM4e in 2027 and the Feynman architecture in 2028. AMD's MI455X reaches mass production in the third quarter of 2026, with MI500 following in 2027 and MI600 in 2028. Huawei's pulled-forward 960 series now lands in roughly the same window as Rubin Ultra and MI500 — closer competitive timing than the original roadmap implied.

Huawei disclosed it has shipped more than 1,000 SuperPoD-class systems to over 370 customers to date, and separately, DeepSeek has reportedly committed to deploying 160,000 Ascend 950DT chips at a data center in Inner Mongolia — a concrete sign of domestic AI labs shifting inference workloads toward Chinese silicon even as they continue relying heavily on Nvidia GPUs for the more demanding task of model training.

Why Speed Matters Here More Than Specs

The significance of this announcement isn't a single spec sheet — it's the acceleration itself. Pulling a chip's readiness date forward by three quarters under export-control pressure and supply-chain constraints signals either that Huawei's manufacturing and design processes have matured faster than expected, or that competitive and geopolitical pressure has made the company willing to compress its own margin for error. Huawei is also expanding its AI infrastructure business internationally, with disclosed moves into Malaysia and Egypt, suggesting the company is positioning Ascend as a genuine alternative for markets where US export restrictions limit access to Nvidia and AMD hardware, not just as a domestic substitute for Chinese customers who have no other option.