The battle for AI computing power stopped being about a single fast chip some time ago. With the Atlas 960E SuperPoD, Huawei is presenting a complete machine in which up to 4,096 AI processors are meant to work together. The Chinese company is building an alternative to Nvidia’s infrastructure.
Huawei unveiled the system on 17 September at HUAWEI CONNECT 2026. The striking part sits between the chips: optical links designed to move enormous volumes of data. With a large AI model, a fast processor achieves little if it constantly has to wait for its neighbours.
Why thousands of chips must act as one team
A large AI model does not always fit in the memory of a single processor. The work is split up, and the parts then exchange information constantly. Think of a huge workshop: extra staff only help if materials and instructions reach everyone on time.
Huawei is therefore trying to design the entire computer as a single unit. In its Atlas 960E announcement, the company describes UnifiedBus for the interconnects and shared addressing of memory. The machine is also fully liquid-cooled.
Light carries the data further
The optical technology is called NPO, short for near-package optics. The optical components sit close to the chips. Huawei uses its Hi-ONE module for this, quoting a transmission capacity of 7.2 terabits per second.
The Atlas 960E is designed to link up to 4,096 NPUs. A neural processing unit is a processor built for neural-network calculations. Huawei is aiming the new infrastructure at training and running very large AI models. Larger clusters are then built from several of these systems.
Huawei is building the whole toolbox
The approach did not come out of nowhere. With the earlier Atlas 900 A3, Huawei already linked up to 384 Ascend 910C chips. With the SuperPoD family, the company is extending a chain of its own processors, networks and server designs.
That is a different way of looking at the chip race. The question is not only how fast one component calculates, but how much useful work the whole installation delivers. Memory, interconnects, cooling and the power consumption of the complete set-up all count.
Nvidia has more than fast hardware
Nvidia’s position also rests on a complete platform. Through CUDA, developers can run software on the GPU. On top of that, the company supplies libraries and tools that accelerate common calculations. Anyone who has worked with those tools for years does not switch at the push of a button.
Huawei counters with its own CANN environment. The company says it is now being developed further as open source with a community of developers. That should make it more attractive to bring models and applications to Ascend hardware.
The real contest plays out in the data centre
A maximum chip count says nothing yet about what a customer ultimately pays for a usable AI answer. For that, the software has to make good use of the system, the chips have to be available in sufficient numbers, and the whole set-up has to keep running reliably. Big theoretical compute figures are only part of the story.
The interesting development is that Huawei is trying to compete on all of those layers at once. The Chinese challenger is selling a plan for the entire AI factory: from processor to interconnect and programming tools. That makes the fight with Nvidia far bigger than a comparison between two chips.
