A roadmap moved forward
Huawei will have its Ascend 960DT accelerator ready in the first quarter of 2027, three quarters earlier than previously planned, the company said on Thursday at the opening of Huawei Connect 2026 in Shanghai. The follow-on Ascend 960PR is now due in the third quarter of 2027, one quarter ahead of its earlier date.
David Wang, deputy chairman of Huawei’s board and its rotating chairman, laid out a schedule that runs well past those parts. “In 2028 and 2029, we will roll out the Ascend 970 and 980 chips, respectively,” Wang said, describing a one-generation-per-year cadence for the line. Huawei said the Ascend 960 generation doubles the performance of its predecessor, a figure the company reports about its own silicon and which no independent evaluation has yet checked.
The timing matters because Huawei is the main domestic alternative for Chinese buyers who cannot get Nvidia’s top accelerators under US export controls. Moving a generation forward by nine months narrows the window in which those buyers have to wait.
The first SuperPoD built on near-packaged optics
Alongside the chip dates, Huawei launched the Atlas 960E SuperPoD, which it says is the industry’s first system built on near-packaged optics, or NPO. A single Atlas 960E scales to 4,096 NPUs, delivers 8 EFLOPS of FP8 compute and carries up to 1 petabyte of high-bandwidth memory.
The optics are the point. Huawei said the system replaces roughly 48,000 conventional 800G optical modules with about 5,500 of its Hi-ONE engines, cutting power draw by more than 550 kilowatts and lifting system availability to 99.8 percent. Each Hi-ONE engine carries 7.2 terabits per second of transmission capacity, and Huawei described it as the first NPO product with a built-in light source to reach mass production.
One caveat came from outside the hall. Analyst Rui Ma, quoted by TechCrunch, noted that the SuperPoD Huawei showed contains 4,096 chips — considerably smaller than the 15,488-chip configuration the company had originally sketched for this generation.
Scaling out with UnifiedBus
Huawei’s larger claim is about what happens when those pods are linked. Interconnected with UnifiedBus over a two-tier, four-plane Clos topology, the company said its agentic SuperCluster can scale to one million NPUs. UnifiedBus is Huawei’s attempt to replace the patchwork of interconnects inside and between racks with a single fabric, and the company has said it intends to publish the 2.0 specification as an open standard.
Two supporting products rounded out the announcement. An upgraded TaiShan 950 SuperPoD supports 4,096 nodes with a unified memory pool of up to 256 terabytes, which Huawei says starts sandboxes thirty times faster and doubles vector search efficiency. A new OceanStor M900 handles petabyte-scale key-value caching for long-context inference and, Huawei says, extends SSD service life sixteenfold.
The ecosystem argument
Wang spent much of the keynote on developers rather than hardware. Huawei put the Kunpeng ecosystem at 4.16 million developers and more than 7,200 partners, and said its CANN compute stack now has 5,200 monthly active developers, 61 percent of them outside Huawei, with more than 90 third-party open-source projects supporting Ascend. “No single company can build an intelligent world alone,” Wang said.
That is the harder part of the pitch. Chinese accelerators have not been short of paper specifications; what they have lacked is the software depth that keeps customers on a platform. The dates to watch now are the first quarter of 2027, when the 960DT is supposed to arrive, and whatever independent benchmark results follow it.
