Huawei unveils new AI computing architecture to rival Nvidia

Huawei launched the Atlas 960E SuperPoD — which the company calls the world's first near-packaged optics (NPO)-based SuperPoD — at HUAWEI CONNECT 2026 in Shanghai, linking up to 4,096 Ascend NPUs into a single computing unit as the Chinese tech giant intensifies its challenge to Nvidia ↗ 2.78% in AI infrastructure.

The announcement, made during a keynote by Yang Chaobin, Huawei's Executive Director of the Board and CEO of its ICT Business Group, also introduced the UnifiedBus interconnect architecture and a suite of new computing products designed to address what Huawei describes as a critical bottleneck in large-scale AI training.eqs-news+1

Tackling the Compute Bottleneck

Huawei says that in traditional computing architectures, as little as 20% of a 100,000-NPU cluster's computing capacity is actually used by models, with the rest sitting idle during data communications. Training 10 trillion-parameter models requires intermediate data volumes that far exceed the memory capacity of any single accelerator.totaltele+1

The UnifiedBus architecture addresses this by combining more than ten interconnect protocols into a single unified protocol, increasing interconnect bandwidth from the 100 GB/s level to the TB/s level and cutting round-trip latency from 7 microseconds to 2 microseconds. It also enables unified global memory addressing within SuperPoDs and direct peer-to-peer access between CPUs, NPUs, memory, and solid-state drives.totaltele

The Atlas 960 SuperPoD, using Huawei's Ascend 960 chips, is expected to deliver 2.3-fold and 2.5-fold improvements in training and inference respectively for 10 trillion-parameter models compared to the commercially deployed Atlas 950 SuperPoD. More than 1,000 units across the SuperPoD product line have already been deployed, according to Huawei.silicon

From Clusters to Enterprise Appliances

Huawei also unveiled interconnect hardware spanning multiple scales. The UnifiedBus LinkBlade eliminates roughly 196 kilometers of copper cabling within a 4,096-NPU SuperPoD through a cable-free design. The UnifiedBus UBG switch supports a radix fan-out capacity of up to 1,024, enabling what Huawei says could be a million-NPU SuperCluster capable of handling tens of trillions of model parameters.eqs-news+1

Beyond large-scale clusters, Huawei applied UnifiedBus to smaller compute appliances. Two Atlas 650E air-cooled servers can be linked through UnifiedBus for switch-free interconnect across 16 NPUs, enabling small and medium-sized enterprises to run trillion-parameter models locally.totaltele

Accelerated Chip Timeline

Huawei disclosed that it has moved up the launch of its Ascend 960 DT training chip by three quarters, now targeting the first quarter of 2027, while the Ascend 960 PR inference chip has been brought forward by one quarter to the third quarter of 2027. Rotating chair Eric Xu Zhijun told the South China Morning Post that the clustering approach would eventually be adopted industry-wide: "In the years to come, other players would certainly follow suit."silicon

Huawei's UnifiedBus competes directly with Nvidia's proprietary NVLink, while a coalition including AMD Advanced Micro Devices, Inc. ↗ 11.04%, Intel ↗ 7.71%, Google Alphabet Inc. ↗ 0.78%, Microsoft ↘ 0.99%, Meta ↗ 0.84%, AWS, and Alibaba ↗ 1.28% backs the rival Ultra Accelerator Link as an open alternative.silicon