Huawei Atlas 960E is Huawei’s new near-packaged-optics SuperPoD, designed to connect as many as 4,096 Ascend NPUs as one tightly coupled system. Announced at Huawei Connect in Shanghai on September 17, it replaces tens of thousands of conventional optical modules with 5,500 Hi-ONE optical engines, according to Huawei.
Huawei Atlas 960E: what changed
| Announcement | September 17, 2026 | Huawei |
|---|---|---|
| Maximum NPU count | 4,096 Ascend NPUs | Huawei |
| Claimed FP8 performance | 8 EFLOPS | Huawei |
| Claimed HBM capacity | Up to 1 petabyte | Huawei |
| Optical design | 5,500 Hi-ONE engines; no 48,000 800G modules | Huawei |
The launch is best understood as an interconnect story, not simply another accelerator announcement. Large AI systems lose useful compute time when processors wait for data, and conventional pluggable optics add power, components and failure points as clusters expand. Huawei is moving optics closer to the switching system and using its UnifiedBus fabric so the group behaves more like one computer.
Huawei says the Hi-ONE engine carries 7.2 terabits per second and includes its own light source. In the company’s configuration, 5,500 engines replace the 48,000 800-gigabit optical modules that would otherwise connect the NPUs. Those are vendor claims rather than independent benchmark results, but they define the engineering wager clearly: fewer optical parts should mean less power and a simpler failure surface.
The company rates a fully configured Huawei Atlas 960E at 8 exaflops of FP8 performance and up to one petabyte of high-bandwidth memory. Associated Ascend 960DT and 960PR processors are scheduled for 2027. That makes this a roadmap-backed infrastructure launch; buyers should separate available components from future silicon when comparing proposals.
Associated Press described the launch as part of Huawei’s effort to narrow the AI infrastructure gap created by restrictions on advanced US technology. TechNode independently reported the near-packaged-optics design. Neither report supplies customer benchmarks, so claims about training speed, inference throughput and reliability should remain attributed to Huawei until production deployments publish results.
The most useful buyer metric may be delivered work per watt. Huawei claims its optical layout saves more than 550 kilowatts and doubles fault-free operating time compared with the conventional module design, reaching 99.8% availability. Operators should ask whether that figure covers planned maintenance, switch failures and software recovery, not only component uptime.
Software portability is the second constraint. Huawei says Ascend is supported as a PyTorch accelerator backend and that CANN is moving toward community-led open development. Enterprise teams still need to test model conversion, kernel coverage, observability and recovery tools with their own workloads before treating nominal compute as usable capacity.
For Indian cloud and enterprise buyers, the immediate relevance is architectural. Near-packaged optics could reduce the electricity and cabling burden of dense AI facilities even where Huawei hardware is not an option. Procurement teams should compare interconnect energy, spare-part counts and recovery procedures alongside chip benchmarks.
Everyone else is reporting a large Chinese AI system; we are explaining why the optical component count is the real design change. Huawei Atlas 960E will matter if its tighter fabric converts claimed scale into sustained model performance without creating a proprietary operations burden.
Related Lapaas Voice coverage
Read Axelera Europa AI accelerator shipments and AWS Bahrain cloud resilience for adjacent context.
Frequently asked questions
What is Huawei Atlas 960E?
Huawei Atlas 960E is a SuperPoD that combines Ascend 960 processors, UnifiedBus and near-packaged optics to operate thousands of NPUs as one scale-up system.
Why does near-packaged optics matter?
Moving optical engines closer to the switching fabric can reduce pluggable modules, power use and connection complexity in very large AI clusters.
When are the Ascend 960 chips available?
Huawei says Ascend 960DT is planned for the first quarter of 2027 and Ascend 960PR for the third quarter of 2027.
Sources
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
