In partnership with
Huawei announced the Ascend 960 SuperPoD at Huawei Connect 2026 in Shanghai on September 17.
The company describes the system as a supernode using near packaged optics, or NPO.
Its announcement places the chip inside a larger computing design rather than presenting it as an isolated accelerator.
AI made PMs faster. Multiplayer mode is still broken.
A PM can summarize research, draft a PRD, and mock up a prototype before lunch. The hard part starts when the team has to decide what actually gets built.
Jira Product Discovery gives product teams one place to capture insights, prioritize ideas with consistent frameworks, and build living roadmaps stakeholders can rally around.
And because it’s connected to Jira, the context behind every decision stays with the work—so developers and their agents know not just what to build, but why.
AI helps PMs move faster. Jira Product Discovery helps the whole team build with confidence.
Huawei says the system is intended for large model training and inference.
Inference is the stage when a trained model produces a prediction or response.
Training and inference place different demands on memory, communication, scheduling, and power.
The announcement therefore centers on the links between processors and memory.
Huawei says its upgraded Kunpeng supernode can support as many as 4,096 nodes.
It also describes a 256 terabyte unified memory pool for that system.
Those figures are company specifications, not independent benchmark results.
The same announcement introduces an OceanStor M900 cluster with a UnifiedBus connection.
Huawei describes the design as a layer three point five system with one hop direct links and petabyte scale key value cache.
Key value cache stores intermediate attention data used during model processing.
The Future of AI in Marketing. Your Shortcut to Smarter, Faster Marketing.
This guide distills 10 AI strategies from industry leaders that are transforming marketing.
Learn how HubSpot's engineering team achieved 15-20% productivity gains with AI
Learn how AI-driven emails achieved 94% higher conversion rates
Discover 7 ways to enhance your marketing strategy with AI.
Large language models can spend substantial time moving that data between computation and memory.
An interconnect that reduces movement or waiting may affect how many accelerators work together.
Huawei also presents a path from one supernode to larger clusters.
It says multiple Ascend 960 supernodes can connect through its optical network or RoCE.
RoCE is a networking method for carrying remote direct memory access over Ethernet.
The company says its cluster architecture can scale to hundreds of thousands of cards.
It further describes a possible one million card configuration using multiple topology layers.
These are design claims from Huawei’s event materials.
They do not show that such a configuration is currently deployed or commercially available.
The event also matters because access to advanced AI chips is restricted by export controls.
The Associated Press reported that Huawei is pursuing domestic computing capability while China faces limits on advanced chip access.
That context changes the meaning of the announcement.
Huawei is not only presenting a component.
It is presenting a connected hardware and software stack intended to support model development inside a constrained supply environment.
The company says CANN, its computing architecture for neural networks, is moving toward more open community operation.
An open software layer could reduce the distance between hardware availability and developer access.
However, an opening statement about software does not measure adoption or compatibility.
Developers still need compiler support, libraries, debugging tools, model ports, and stable documentation.
They also need evidence that applications behave consistently across different cluster sizes.
Large systems require power delivery, cooling, networking, monitoring, repair procedures, and data movement controls.
AP described the Atlas 960 SuperPoD as a successor to an earlier model with improvements in training and inference.
That description adds outside context but does not validate Huawei’s performance claims.
Reuters reporting also described Huawei’s plans for additional Ascend chips in 2027 and its emphasis on UnifiedBus.
Those future products remain plans until they are delivered and independently evaluated.
The technical development is best read as an infrastructure announcement.
Its significance lies in combining accelerators, optical links, shared memory, cache, software, and cluster management.
The useful question is whether the complete system can deliver repeatable results across real workloads.
Until that evidence appears, Huawei’s specifications should remain separate from verified performance conclusions.
This infrastructure story remains open until independent tests show how the architecture behaves outside company demonstrations.
200+ Proven Ways to Make Money With AI in 2026
The next wave of millionaires will be people who figured out how to make AI work for them.
The window to get ahead is still open. But not for long.
Here are 200+ proven ways to make money with AI in 2026.
Sign up for Superhuman AI, the free daily newsletter read by 1M+ professionals, and get instant access to all 200+ ways to profit from AI this year.
