Huawei cuts its flagship AI pod to 4,096 NPUs, a quarter of the count it once described
The Atlas 960E fits 8 EFLOPS behind Huawei's own optical engines, and the Ascend 960DT arrives three quarters early, while rotating chairman Eric Xu says supply goes to Chinese customers first and dates China's catch-up with its own demand to 2030.
The Product Desk
What happened
- Huawei said at its Connect conference that the Ascend 960DT will arrive in the first quarter of 2027, three quarters earlier than its published roadmap had it.
- The Atlas 960E SuperPoD it launched alongside holds 4,096 NPUs, delivers 8 EFLOPS at FP8 precision and carries a petabyte of HBM per pod.
- Fitting 5,500 of Huawei's own Hi-ONE near-packaged optical engines, each moving 7.2 Tbit/s, removes 48,000 800G pluggable optical modules from the pod.
- Huawei says that change cuts power draw by more than 550 kilowatts, doubles the time between faults and lifts availability to 99.8%.
- Huawei had previously described an Atlas 960 SuperPoD scaling to 15,488 Ascend 960 chips, well above the pod it put on stage this week.
Why it matters
- What it limits constraint Ascend is only a purchasable option for the customers Huawei chooses to serve, and Eric Xu says that queue starts with Chinese buyers and has no global expansion behind it.
- Who has to decide decision A buyer weighing this has to price the whole pod, because the per-chip comparison with Nvidia's B300 goes the wrong way and the case rests on a utilisation number Huawei generated in its own lab.
- What doesn't add up contradiction The roadmap moved in two directions in one keynote: the chip came in by three quarters while the pod around it lost about three quarters of its chip count.
- What's now possible capability With CANN open-sourced and Ascend listed as a PyTorch accelerator backend, a team can do the porting work now without holding any Ascend allocation.
Next story · Build · Sep 30
DeepSeek's Ascend ports swap a CUDA dependency for Huawei's CANNConfirmed4 publishers
Eight EFLOPS of FP8 spread over 4,096 NPUs is about 1.95 petaFLOPS each [24], and a petabyte of HBM across the same pod averages roughly 244GB per NPU [25]. The Register puts the 960DT at up to 288GB and four petaFLOPS of FP4, about half the FP8 and a third of the FP4 compute of Nvidia's B300 [9]. Nvidia cannot sell the B300 in China [10].
Huawei's answer to that gap is that the interconnect sets throughput. Yang Chaobin, who runs its ICT business group, said as little as 20% of the compute in a conventional 100,000-NPU cluster goes to the model, with the rest idle while data moves [12]. UnifiedBus folds more than ten protocols into one and cuts latency from seven microseconds to two [11]. In a Huawei lab simulation, not a measured customer cluster, pods of 4,000 NPUs reach 2.75 times the model FLOPs utilisation of eight-NPU servers [13].
The optics claim converts into per-chip terms too. More than 550 kilowatts saved across 4,096 NPUs is about 134 watts per NPU [26]. The 99.8% availability Huawei quotes works out at roughly 17.5 hours of downtime a year [27]. Huawei has filed an implementation agreement on the near-packaged optics with the Optical Internetworking Forum, a standards body [20].
Stacked, the pods reach the cluster figures Huawei put on stage. A 512,000-NPU two-tier Clos network needs 125 Atlas 960E pods [16], and at 550 kilowatts a pod the optical change would avoid about 69 megawatts of pluggable module power at that size [17]. The million-NPU configuration has not been built, and The Register describes it as theoretical [3].
A build plan has to absorb the shrink. Some 4,096 chips is about 26% of the 15,488 Huawei had described earlier, a factor of 3.8 [21]. "The chip itself is coming WAY earlier, but the SuperPoD they announced is much smaller than what they originally laid out," China tech analyst Rui Ma wrote [7].
Supply is a separate problem. Rotating chairman Eric Xu told reporters Huawei is prioritising Chinese customers and has no broad global expansion plan [4], and he put the moment China catches up with its own hardware demand at 2030 [22]. Huawei executives told Bloomberg the company now holds a bigger AI chip market share than Nvidia inside China and cannot make enough of them [6]. His colleague David Wang opened the Shanghai conference by saying AI "may well be the final technological revolution in human history" [23].
Will Huawei sell to you before 2030? And is your training job's utilisation limited by the interconnect or by the chip? A Chinese customer with allocation and a large training run is who the Atlas 960E was built for. A customer with allocation and inference-shaped work is paying for wiring it will not use. Outside China the roadmap is a data point in an Nvidia negotiation and a reason to start porting early: Ascend already supports more than 90 third-party open-source projects [15].
What to watch
- Whether the 4,096-NPU figure for the Atlas 960E holds when the Ascend 960DT ships in the first quarter of 2027.
- Whether the Optical Internetworking Forum turns Huawei's Hi-ONE filing into a spec other vendors can build to.
- Any Ascend sale outside China before 2030, against Eric Xu's no-global-expansion line.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence55
- Adoption50
- Hype gap+30
- Incentives70
- Confidence60
Perspective Coverage
4 publishers- Builder
- Builder 26%
- Operator
- Operator 38%
- Investor
- Investor 36%
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
The Ascend 960PR, built for inference, follows in the third quarter.
- [2]
David Wang placed an Ascend 970 in 2028 and an Ascend 980 in 2029, crediting the Tau Scaling Law for the cadence.
- [3]
A two-tier, four-plane Clos network links 512,000 NPUs and a multi-rail topology takes that to a million; Huawei has not built one, and The Register describes the million-NPU configuration as theoretical rather than proven.
- [4]
Rotating chairman Eric Xu said Huawei is prioritising Chinese customers and has no broad global expansion plan.
- [5]
Huawei had previously described an Atlas 960 SuperPoD scaling to 15,488 Ascend 960 chips.
- [6]
Huawei executives told Bloomberg the company now holds a bigger AI chip market share than Nvidia inside China but cannot make enough of them.
- [7]
"The chip itself is coming WAY earlier, but the SuperPoD they announced is much smaller than what they originally laid out," China tech analyst Rui Ma wrote.
- [8]
The Atlas 960E SuperPoD holds 4,096 NPUs, delivers 8 EFLOPS at FP8 precision and carries a petabyte of HBM per pod; Huawei calls it the industry's first to use near-packaged optics.
- [9]
The Register puts the Ascend 960DT at up to 288GB of memory and four petaFLOPS of FP4, roughly half the FP8 and a third of the FP4 compute of Nvidia's B300.
- [10]
Nvidia cannot sell the B300 in China.
- [11]
Yang Chaobin, who runs Huawei's ICT business group, said UnifiedBus folds more than ten protocols into one, raising bandwidth from hundreds of gigabytes per second to terabytes and cutting latency from seven microseconds to two.
- [12]
Yang Chaobin said that in a conventional 100,000-NPU cluster as little as 20% of the compute goes to the model, with the rest idle while data moves.
- [13]
Huawei's own lab simulation compares two ways of building such a cluster: pods of 4,000 NPUs reach 2.75 times the model FLOPs utilisation of eight-NPU servers.
- [14]
CANN, which does for Ascend roughly what CUDA does for Nvidia, has moved to community-driven open source, and external developers make up 61% of the CANN community, outnumbering Huawei's own for the first time.
- [15]
Ascend supports more than 90 third-party open-source projects, and Huawei says it is now an official PyTorch accelerator backend, the first Chinese compute platform listed on the PyTorch website.
- [16]
A 512,000-NPU network would need 125 Atlas 960E pods.
- [17]
At 550 kilowatts saved per pod, a 125-pod network avoids about 69 megawatts of pluggable optical module power.
- [18]
Huawei's own near-packaged optical engine, Hi-ONE, moves 7.2 Tbit/s per unit with a built-in light source, and fitting 5,500 of them removes 48,000 800G pluggable optical modules from the pod.
- [19]
Huawei says the near-packaged optics cut power draw by more than 550 kilowatts, double the time the system runs without a fault and lift availability to 99.8%.
- [20]
Huawei has filed an implementation agreement on the near-packaged optics technology with the Optical Internetworking Forum, a standards body.
- [21]
The announced 4,096-NPU pod is about 26% of the 15,488-chip configuration Huawei had described, a factor of 3.8 smaller.
- [22]
Eric Xu put the moment China catches up with its own hardware demand at 2030.
- [23]
AI "may well be the final technological revolution in human history," Huawei rotating chairman David Wang told Huawei Connect 2026 in Shanghai.
- [24]
Each NPU in the Atlas 960E averages about 1.95 petaFLOPS of FP8.
- [25]
The pod averages roughly 244GB of HBM per NPU.
- [26]
The claimed 550 kilowatt saving is about 134 watts per NPU.
- [27]
Availability of 99.8% is about 17.5 hours of downtime a year.
- [28]
Huawei said its next AI processor, the Ascend 960DT, arrives in the first quarter of 2027, three quarters earlier than its roadmap said.
Sources
4 independent publishers whose own reporting we read for this story.
- digitaltrends.comHuawei just laid out its 2027 AI chip plans, and Nvidia is clearly the target - Digital Trends
1 article · September 16, 2026
- straitstimes.comHuawei set to unveil China’s best answer to Nvidia AI chip reign | The Straits Times
1 article · September 16, 2026
- techcrunch.comHuawei plans Q1 2027 launch of new AI chip as it takes on Nvidia
2 articles · September 17, 2026
- thenextweb.comHuawei’s answer to Nvidia isn’t a faster chip, it’s a much bigger machine
1 article · September 18, 2026
Topics and entities
Topics
- AI cluster interconnect
- US-China Chip Export Controls
- AI Accelerators
- Co-Packaged Optics