Huawei Pulls Ascend 960DT Forward to Q1 2027, Shows a 4,096-NPU SuperPoD

On 17 September 2026, at Huawei Connect in Shanghai, rotating chairman David Wang said the Ascend 960DT training chip will be available in the first quarter of 2027, three quarters ahead of its original roadmap, according to Huawei's own keynote summary. The inference-oriented 960PR follows in the third quarter of 2027, one quarter early, with the Ascend 970 in 2028 and 980 in 2029. Huawei also showed the Atlas 960E SuperPoD, scaling to 4,096 NPUs with 8 EFLOPS of FP8 compute. For a company building under US export controls, that is a fast cadence, and it is the domestic alternative Beijing is pointing to while it holds back Nvidia imports.

What happened

Wang's keynote set out a "one-generation-a-year cycle" for Ascend, in his words, and claimed Ascend 960 development is on track to double performance. He did not give per-chip FLOPS, memory capacity, memory bandwidth or a process node for any of the new parts; the keynote describes future gains in terms of compute doubling and improved memory and interconnect, without numbers.

The system-level claims were more specific. The Atlas 960E is described as Huawei's first SuperPoD built on near-packaged optics (NPO). Huawei says it scales to 4,096 NPUs, delivers 8 EFLOPS of FP8, holds up to 1 PB of HBM, and replaces the 48,000 800G optical modules a conventional design would need with 5,500 of its own Hi-ONE optical engines, cutting power by more than 550 kW. Each Hi-ONE engine is rated at 7.2 Tbit/s.

Huawei also claimed more than 1,000 of its existing Atlas 900 A3 SuperPoDs are deployed, said the Atlas 950 is in large-scale commercial use, and described an "Agentic SuperCluster" architecture for up to 512,000 NPUs, or up to a million with multi-rail topology. TrendForce, citing Chinese media, put the liquid-cooled Atlas 960 SuperPoD at the third quarter of 2027 and an air-cooled Atlas 860 SuperPoD at the second quarter.

The numbers

Why it matters

Huawei's argument is not that one Ascend beats one Nvidia GPU. It is that a much bigger machine, wired with its own optics and its UnifiedBus interconnect, can make up for weaker individual chips. Wang's keynote cites an internal simulation in which a 100,000-NPU cluster built from 4,096-NPU SuperPoDs reaches 2.75 times the model FLOPs utilisation of a same-size cluster built from 8-NPU servers, and says intra-cluster communication eats over 40% of training time on traditional designs. That is a vendor simulation, but the underlying point is one every distributed-training engineer recognises: past a certain size, the network is the computer.

The timing matters politically. Seven weeks after the keynote, Reuters reported that about 600,000 Nvidia H200s were waiting in Hong Kong because Beijing had not approved large-scale imports, and TNW's summary of that report says Beijing wants Alibaba, ByteDance, Tencent and DeepSeek to use Chinese-made chips, especially for inference. A credible annual Ascend roadmap is what makes that policy defensible at home.

Demand is not the problem. TrendForce, citing Bloomberg, reports DeepSeek plans to deploy at least 160,000 Ascend 950DT chips at a new data centre in Inner Mongolia, and that Huawei is exploring overseas markets including Malaysia and Egypt.

What sceptics say

The biggest gap is disclosure. Huawei gave no per-chip specifications for the 960 family, no process node and no memory supplier. Everything above is a company claim, and none of it has been independently benchmarked.

The SuperPoD also shrank. Analyst Rui Ma, quoted by TNW, put it plainly: "the SuperPoD they announced is much smaller than what they originally laid out." Huawei had earlier described an Atlas 960 SuperPoD of 15,488 chips; the 960E tops out at 4,096. The million-NPU cluster is a design ceiling, not something Huawei says it has built.

Supply is the binding constraint. TrendForce notes that leading Chinese labs still rely heavily on Nvidia GPUs for training, with domestic chips gaining ground mainly in inference, and TNW, citing executives who spoke to Bloomberg, reports that Huawei cannot make enough AI chips. Pulling a roadmap forward is easier than pulling manufacturing capacity forward, particularly when advanced lithography and HBM are both under export controls.

What it means for engineering and DevOps teams

My take: for teams outside China, Ascend is not a procurement option, and this news changes little about what you buy in 2027. What it does change is the competitive pressure on the vendors you do buy from. A second ecosystem growing on an annual cadence, with its own interconnect and its own software stack, is one more reason Nvidia, AMD and Google are racing on rack-scale systems rather than individual chips.

There are two transferable lessons. First, the interconnect argument applies to you too: if your training or large-batch inference jobs spend a big share of time in collective communication, a faster GPU will not fix it, and better topology awareness in the scheduler might. My explainer on GPUs on Kubernetes covers where DRA and topology hints fit. Second, optics and power are now first-class design constraints. Huawei's headline numbers are about optical modules and kilowatts as much as FLOPS, and the same arithmetic shows up in AI data centre power planning everywhere.

If you run models that might be deployed on Chinese infrastructure (for a joint venture, or open-weight models your users self-host there), keep your serving stack portable. Models trained on one stack and served on another is becoming a normal pattern, not an edge case.

FAQ

When is the Huawei Ascend 960 released?

Huawei says the Ascend 960DT will be available in Q1 2027, three quarters earlier than its original roadmap, and the Ascend 960PR in Q3 2027. The Ascend 970 is planned for 2028 and the 980 for 2029.

How does Ascend 960 compare with Nvidia?

Huawei has not published per-chip specifications for the 960 family, so a direct comparison is not possible from primary sources. Huawei's pitch is system-level: large SuperPoDs connected by its own optics and UnifiedBus, rather than per-chip parity.

What is the Atlas 960E SuperPoD?

It is Huawei's first SuperPoD using near-packaged optics, scaling to 4,096 NPUs with a claimed 8 EFLOPS of FP8 compute and up to 1 PB of HBM. Huawei says it uses 5,500 Hi-ONE optical engines in place of 48,000 conventional 800G modules.

Related: GPUs on Kubernetes Explained (2026): Device Plugin, MIG, Time-Slicing, DRA, AI Power in 2026: 485 TWh, a 0.3 GW Stargate, a 116 GW Turbine Backlog, and 1,800 MW That Dropped Off the Grid in Seconds and LLM Inference Engines 2026: vLLM vs SGLang vs TensorRT-LLM vs Dynamo.