Huawei Pulls Ascend 960DT Forward to Early 2027
· TexaHub · Semiconductors, Geopolitics & Tech, Regional Tech Ecosystems
Huawei says its next AI accelerator arrives in Q1 2027, ahead of plan, but per-chip performance still trails Nvidia and the new SuperPoD is smaller than first promised.
The announcement
At Huawei Connect in Shanghai, rotating chairman David Wang said the Ascend 960DT will be available in Q1 2027, three quarters ahead of the company's original roadmap, and that the Ascend 960PR will be ready in Q3 2027, one quarter early [1]. Wang also said Huawei is evolving the Ascend line on a one-generation-a-year cycle, with the Ascend 970 and 980 due in 2028 and 2029 [1]. TechCrunch, citing the same event, described the 960DT shift as a two-quarter move from Q3 2027, so reports differ on the size of the acceleration [2].
On specifications, heise reported that the 960DT carries 288 GB across eight memory stacks with 9.6 TB/s of bandwidth and delivers 4 petaflops at FP4 [3]. The Register said the memory is likely Huawei's in-house HiZQ technology, a homegrown alternative to HBM, and that the part offers twice the performance and memory capacity of the 950 series launched earlier this year [4].
Scale over per-chip speed
Huawei's strategy remains to compensate for weaker chips with very large systems. The company says a single Atlas 960E SuperPoD can scale to 4,096 NPUs and deliver 8 EFLOPS of FP8 compute with up to 1 petabyte of HBM capacity [1]; heise gives a maximum of 16 exaflops for the same system, and the two figures appear to refer to different number formats [3]. China tech analyst Rui Ma noted to TechCrunch that the SuperPoD announced is much smaller than what Huawei originally laid out [2]. The Register reported that Huawei expects to support clusters of up to 512,000 NPUs and sees million-NPU superclusters as possible [4].
The per-chip gap is large. heise put Nvidia's Rubin at up to 50 FP4 petaflops and AMD's MI455X at 40, concluding that at least ten 960DTs are needed to match one [3]. The Register's assessment was that neither Nvidia part is for sale in China, so domestic model developers lack better options [4]. Quartz reported that rotating chairman Eric Xu said Huawei lacks sufficient production capacity for international expansion [5].
Why it matters
An earlier 960DT tightens the timeline on which Chinese labs can plan training and inference capacity on domestic silicon. The binding constraint, on Huawei's own account, is manufacturing volume rather than demand, which makes yields, memory supply and interconnect maturity the metrics to watch as Q1 2027 approaches. Claims about delivery dates and system scale remain company projections until hardware ships.
Sources
[1] Advancing the Agentic World, Building a Solid Silicon Foundation — Huawei, 17 Sep 2026
[2] Huawei plans Q1 2027 launch of new AI chip as it takes on Nvidia — TechCrunch, 17 Sep 2026
[3] Huawei packs 4096 AI accelerators into a massive server — heise online, 18 Sep 2026
[4] Huawei's next-gen Ascend NPUs could become China's best option — The Register, 17 Sep 2026
[5] Huawei speeds up Ascend 960 AI chip launch to challenge Nvidia — Quartz, 17 Sep 2026