Full Breakdown
Huawei Accelerates AI Chip Rollout and Unveils Massive “Supernode” Architecture
By Drooid · · How we work
Accelerated Launch of the Ascend 960 Series
On September 17 2022, Huawei Connect in Shanghai announced that its Ascend 960 AI chip family will be ready earlier than planned. The Ascend 960DT, aimed at AI model training, is slated for Q1 2027, three quarters ahead of the original Q4 2027 schedule. The inference-focused Ascend 960PR is expected in Q3 2027, one quarter ahead of its prior timeline. Rotating chairman Eric Xu said the company is already in “extensive discussions” with Chinese AI model developers and expects many to begin training on the new chips next year. He added that Huawei lacks sufficient capacity to meet domestic demand and will not pursue a full-scale international rollout at this stage.
Peerium Computing Architecture and UnifiedBus
Huawei introduced its Peerium Computing Architecture, which relies on the proprietary UnifiedBus interconnect to link processors, memory, storage, and networking equipment. The architecture enables “Supernodes” that aggregate thousands of AI processors into a single logical computer. Huawei’s Atlas 960 SuperPoD, the first system to use near-packaged optics (NPO), can connect up to 4,096 AI processors per node, according to company materials. The firm also disclosed a roadmap for larger clusters—“superclusters”—that could eventually incorporate up to 1 million processors.
Data & Statistics
- Atlas 950 SuperPoD can connect up to 256,000 accelerator cards; the newer Atlas 960 SuperPoD targets 4,096 cards per node.
- Huawei reports more than 1,000 SuperPoD systems deployed to roughly 370 customers.
- Over 5,200 developers are active each month on software for Ascend chips, and more than 40 AI models have been trained on Huawei’s platform.
- China’s intelligent-computing capacity reached 2,185 EFLOPS by the end of June 2026, a 177 % year-on-year increase (MIIT).
Official Statements & Responses
- Eric Xu emphasized that the chip push is part of China’s drive for “full self-sufficiency for chips,” noting that reliance on foreign suppliers “cannot determine our fate.”
- David Wang announced the accelerated timeline at the conference and highlighted the “biggest opportunity” Huawei sees in AI infrastructure.
- Guo Ping, chairman of Huawei’s supervisory board, described AI as the company’s “biggest opportunity” and said its computing and connectivity infrastructure will play a role comparable to Nvidia’s.
- Wang Tao stated that the Ascend 960 “delivers twice the performance” of the previous generation and that UnifiedBus and Hi-ONE optics together enable a SuperPoD to link 4,000–4,096 processors.
Conflicting Reports & Gaps
Sources differ on the exact processor count per Atlas 960 SuperPoD: some describe 4,000 processors, others 4,096. Huawei’s own materials cite the higher figure, but analyst Rui Ma noted the earlier Atlas 960 announcement referenced a smaller scale. No independent verification of performance metrics such as the claimed 8 EFLOPS FP8 compute or 1 PB HBM capacity has been provided.
Why It Matters
Huawei’s accelerated chip schedule and focus on system-level interconnects aim to offset the performance gap with Nvidia’s GPUs. By aggregating large numbers of modest-power chips, Huawei hopes to meet demand for training models with 10 trillion parameters and beyond, while reducing reliance on foreign semiconductor technology constrained by U.S. export controls.
What’s Next
Huawei plans to release the Ascend 970 in 2028 and the Ascend 980 in 2029, maintaining an annual chip-generation cadence. The company also intends to ship near-packaged optics modules in the coming quarters and continue standard-setting work on NPO technology through industry bodies such as the Optical Internetworking Forum. A high-level meeting between President Donald Trump and President Xi Jinping is scheduled for September 24 in Washington, DC, underscoring the geopolitical backdrop to Huawei’s AI push.