Huawei Ascend is Huawei’s family of AI processors for training and inference, which Huawei calls NPUs. The current line runs from the Ascend 910 series, deployed in Huawei’s Atlas 900 A3 SuperPoD, to the Ascend 950PR and 950DT scheduled for 2026, and Huawei has published a roadmap through the Ascend 980 in 2029. Huawei’s announcements center on large clustered systems built from these chips, the Atlas SuperPoDs.
On this page
The Ascend line-up as Huawei describes it
Huawei laid out its chip roadmap at Huawei Connect 2025 on September 18, 2025, in a keynote by Eric Xu, and updated it at Huawei Connect 2026 on September 17, 2026, in a keynote by David Wang, Deputy Chairman and Rotating Chairman.
| Chip | Timing as Huawei states it | Specifications Huawei gives |
|---|---|---|
| Ascend 950PR | Q1 2026 | 950 series: 1 PFLOPS FP8, 2 PFLOPS MXFP4, 2 TB/s interconnect; HiBL 1.0 memory |
| Ascend 950DT | Q4 2026 | 950 series compute; HiZQ 2.0 HBM, 144 GB at 4 TB/s |
| Ascend 960DT and 960PR | Q1 2027 and Q3 2027 (2026 update) | Double the 950’s compute, memory and bandwidth (2025 roadmap) |
| Ascend 970 | 2028 | Doubles FP8 and FP4 compute over the 960; 4 TB/s interconnect (2025 roadmap) |
| Ascend 980 | 2029 | Compute to keep doubling (2026 update) |
The 910 series came first. Huawei said in its 2025 keynote that “the Ascend 910C chip has become better known in the industry as we’ve scaled up the deployment of our Atlas 900 A3 SuperPoD.”
Huawei names the two 950 variants by workload: “PR” for prefill and recommendation, “DT” for decode and training. The 950PR uses HiBL 1.0, a memory Huawei developed and describes as “more cost-effective than HBM3E and HBM4E”; the 950DT uses Huawei’s HiZQ 2.0 HBM. The 960 adds HiF4, Huawei’s own 4-bit data format.
What changed in the 2026 update
At Huawei Connect 2026, David Wang said development of the Ascend 960 “has exceeded the company’s expectations, with performance doubling as planned.” The 2025 roadmap had put the 960 in the fourth quarter of 2027. The 2026 keynote split it into two parts: the “Ascend 960DT will be available in Q1 2027, three quarters ahead of the company’s original roadmap,” and “the Ascend 960PR will be ready in Q3 2027, one quarter ahead of schedule.” Wang added: “In 2028 and 2029, we will roll out the Ascend 970 and 980 chips, respectively,” and said Huawei expects compute to keep doubling along with memory bandwidth, capacity and interconnect bandwidth. Huawei attributes this to what it calls the “Tau (τ) Scaling Law.”
Atlas SuperPoDs and SuperClusters
Huawei builds Ascend into large interconnected systems. A SuperPoD is a set of Ascend NPUs connected over Huawei’s UnifiedBus interconnect so they work as one system; a SuperCluster links many SuperPoDs. At Huawei Connect 2025, Huawei introduced UnifiedBus 2.0 and the following systems:
| System | Timing as Huawei states it | NPUs | Compute | Memory and interconnect |
|---|---|---|---|---|
| Atlas 950 SuperPoD | Q4 2026 | Up to 8,192 Ascend 950DT | 8 EFLOPS FP8, 16 EFLOPS FP4 | 1,152 TB; 16 PB/s |
| Atlas 960 SuperPoD | Q4 2027 | 15,488 Ascend 960 | 30 EFLOPS FP8, 60 EFLOPS FP4 | 4,460 TB; 34 PB/s |
Huawei also announced the Atlas 950 SuperCluster, with more than 500,000 Ascend NPUs, and the Atlas 960 SuperCluster, with more than one million.
At Huawei Connect 2026, Wang said “over 1,000 Atlas 900 A3 SuperPoDs have been deployed, and Atlas 950 SuperPoD is seeing large-scale commercial use.” He introduced the Atlas 960E SuperPoD, which “can scale up to 4,096 NPUs, delivering 8 EFLOPS of FP8 compute performance, with up to 1 petabyte of HBM capacity,” with “99.8% system availability.” He also said that a two-tier, four-plane Clos network lets a SuperCluster interconnect up to 512,000 NPUs, or up to one million with a multi-rail topology.
Huawei’s FP8 and FP4 figures are its own and are given here exactly as Huawei published them. For accelerators from other vendors, see the AI chips catalog.
How Huawei builds it
Huawei says it designs around the manufacturing it can use. Eric Xu said Huawei builds its systems “with the semiconductor manufacturing process nodes that are practically available to the Chinese mainland.” For the 950 generation it has also developed its own memory, HiBL and HiZQ, alongside the UnifiedBus interconnect that links chips into SuperPoDs.
On software, Huawei pledged at Huawei Connect 2025 to fully open-source its CANN software, its Mind series toolchains and its openPangu foundation models by December 31, 2025. Huawei’s general-purpose Kunpeng CPU line is a separate product; Huawei said at Huawei Connect 2026 that Kunpeng has more than 4.16 million developers and 7,200 ecosystem partners.
US export controls and Ascend
The U.S. government addresses Ascend chips directly. On May 13, 2025, the Commerce Department’s Bureau of Industry and Security (BIS) issued guidance on General Prohibition 10 that names the Huawei Ascend 910B, 910C and 910D. BIS said these chips “were likely developed or produced in violation of U.S. export controls,” and that using them without authorization “risks violating U.S. export controls and may subject companies to BIS enforcement action.” BIS said it would not pursue parties that obtain such a chip solely for technical analysis or evaluation, such as destructive testing.
In the other direction, U.S. rules require a license to export advanced AI chips classified under ECCN 3A090 to China. A BIS rule of January 15, 2026 moved license applications for some chips below set performance limits, such as Nvidia’s H200 and AMD’s MI325X, from a presumption of denial to case-by-case review for China and Macau. Separately, BIS guidance of May 31, 2026 says a license is required to export such chips to entities headquartered in Country Group D:5 or Macau, wherever they are located. Company profiles for Huawei and other chipmakers are in the AI companies directory.
