The next battle in artificial intelligence is no longer only about the power of one chip. It is about making thousands of processors work like a single machine. At HUAWEI CONNECT 2026, held in Shanghai from September 17 to 19, Huawei unveiled the Atlas 960E SuperPoD. The Chinese group describes the infrastructure as the industry’s first SuperPoD built around NPO, or near-packaged optics. Behind the cascade of technical figures sits a clear industrial message: when a company does not control every part of the global semiconductor chain, it can seek an advantage through system architecture, networking and software.
One computer made of thousands of processors
A SuperPoD is not simply a larger server. It brings together many compute nodes with sufficiently low latency to share data and behave as a coherent unit. This organization has become central to training large models and serving their answers. As models grow, the time lost moving data between processors becomes as important an obstacle as the raw power of each component.
Huawei says one Atlas 960E can connect as many as 4,096 NPUs, the processors specialized for AI tasks, with advertised compute performance of 8 EFLOPS at FP8 precision and 16 EFLOPS at FP4. The system could include one petabyte of HBM memory. These are vendor figures that will need to be tested with real workloads. They nevertheless convey the scale of the ambition: to build a machine capable of training and running models with up to ten trillion parameters.
Optics moves to the center of the product
The most structural change is found in the interconnect. The Hi-ONE engine brings optical components closer to the compute circuits. Huawei advertises transmission capacity of 7.2 terabits per second for each unit and the use of an integrated light source. This NPO approach is designed to reduce the distance traveled by electrical signals, then use light to carry more data with less heat and loss.
According to Huawei, 5,500 Hi-ONE units would replace 48,000 conventional 800G optical modules. The company claims a reduction in power consumption of more than 550 kilowatts and system availability of 99.8%. Again, these numbers come from the supplier’s presentation. The direction is credible, however: networking, cooling and electricity have become major economic constraints in data centers. An interconnect innovation can therefore matter as much as a processor generation.
The complete-system strategy
Huawei is not selling local acceleration alone. The group combines Ascend 960 processors, UnifiedBus networking, Hi-ONE units, OceanStor M900 storage and its CANN software layer. Multiple SuperPoDs can then form a SuperCluster. The company describes an architecture capable of connecting up to 512,000 NPUs, or even one million with a broader topology. That promise places vertical integration at the heart of the offer.
The logic resembles the one now shaping the wider industry: value is moving from the isolated component to the platform. A corporate customer does not choose only a chip. It chooses a programming environment, tools, memory, a network, maintenance and an upgrade calendar. The cost of leaving can become high. Competition is therefore about the ecosystem as much as operations per second.
A response to technology restrictions
The announcement arrives amid US restrictions on Chinese access to advanced chips and manufacturing equipment. Associated Press notes that Beijing is rapidly pursuing technological self-reliance while Huawei seeks to narrow the gap with global leaders. Reuters separately reports that Chinese demand for the group’s computing equipment still exceeds its production capacity. The challenge is therefore not only to design a competitive architecture, but to manufacture enough units and deploy them reliably.
In that context, scale becomes a strategic response. If every individual processor cannot always compete at the same fabrication node, a more efficient interconnect can offset part of the gap at system level. This approach does not erase limits involving manufacturing, yields or high-bandwidth memory supply. It tries to move the basis of comparison toward fields in which Huawei has long experience: telecommunications, networking, optics and infrastructure integration.
Software will determine adoption
A spectacular machine can remain underused if developers struggle to move their models onto it. Huawei says CANN is now under continuous open-source development and that the Kunpeng ecosystem includes 4.16 million developers. The group also highlights official support for Ascend as an accelerator backend in PyTorch. These points are intended to reassure teams accustomed to competing tools and unwilling to rewrite their entire stack.
The promised openness will have to be assessed precisely: repository governance, documentation quality, integration pace, library compatibility and the community’s ability to influence the roadmap. Registration totals do not guarantee meaningful activity. For companies, migration ease, the availability of skilled engineers and stability over several years will matter more than a conference demonstration.
Energy becomes a performance metric
The AI race is often described through parameter counts and compute power. It also depends on electricity grids, cooling systems, buildings and supply chains. A 550-kilowatt saving on the interconnect would be significant, but it represents only part of a SuperPoD’s total consumption. Customers will need to compare energy per completed task, actual utilization and maintenance costs, not just the theoretical maximum.
The Atlas 960E’s fully liquid-cooled design confirms this shift. A data center is becoming a thermal machine as much as a computing machine. The advertised 99.8% availability must also be understood at the level of critical workloads: even a few hours of downtime can be expensive when a cluster supports industrial, financial or public services. Efficiency has value only when paired with resilience.
An accelerated but still forward-looking schedule
Huawei says the Ascend 960DT, intended in part for training, will be available in the first quarter of 2027, three quarters ahead of the original schedule, while the inference-oriented Ascend 960PR will arrive in the third quarter. The Atlas 960E presented this week therefore outlines a platform whose full expansion still depends on future components. Caution requires a distinction between a prototype, a commercial launch and volume availability.
That distinction does not make the announcement irrelevant. Large buyers plan infrastructure several years ahead. A credible roadmap can influence investments, software partnerships and workforce training today. It can also push competitors to respond on optics, shared memory and energy cost.
The battlefield is moving
The Atlas 960E does not yet prove that Huawei has overtaken its rivals. It does show that a comparison limited to one chip is becoming inadequate. Architecture, light, memory, software and energy now form one product. The winner may not be the company with the fastest component, but the one that delivers a system that is available, programmable, efficient and repeatable at scale.
In Shanghai, Huawei chose to tell this new story through thousands of processors linked as one machine. The next stage will be less dramatic and more decisive: proving at customer sites that those promises survive real models, failures, electricity bills and time. That is where infrastructure is measured, far away from the conference stage.