At Huawei Connect 2026 in Shanghai, Huawei presented its vision for the future of AI infrastructure, showcasing new Ascend processors arriving earlier than planned and unveiling the Atlas 960E SuperPod, a massive AI machine that leverages innovative optical technology to deliver unprecedented scale and efficiency.
- Ascend 960DT chip launches three quarters ahead of schedule in early 2027
- Atlas 960E SuperPod hosts 4,096 AI processors with 8 EFLOPS at FP8 precision
- Huawei’s UnifiedBus interconnect merges multiple protocols boosting bandwidth and lowering latency
What happened
During Huawei Connect 2026, rotating chairman David Wang outlined a bold strategy to accelerate AI chip releases and introduce new large-scale AI infrastructure. The company announced that its Ascend 960DT processor will be available in Q1 2027, advancing its timeline by three quarters. Additional chips—Ascend 960PR for inference, Ascend 970, and Ascend 980—are scheduled through 2029, all following the Tau Scaling Law for development cadence.
The highlight was the launch of the Atlas 960E SuperPod, a massive computing system featuring 4,096 NPUs and delivering 8 exaFLOPS at FP8 precision. This pod integrates Huawei’s proprietary near-packaged optical engine, Hi-ONE, which dramatically reduces power consumption and increases system uptime by replacing tens of thousands of conventional optical modules with compact, high-speed optical links.
Why it matters
Huawei’s approach shifts the competitive AI compute landscape from chasing faster chips to building larger, more efficient AI machines. By combining extensive NPU counts with cutting-edge optical interconnects, Huawei aims to maximize processor utilization and cluster performance, significantly reducing idle compute time caused by data movement delays. This architectural innovation could redefine efficiency standards in AI training and inference workloads.
Despite the advancements, Huawei’s chip-level performance still trails Nvidia’s top-tier models, but its growing AI market share in China indicates successful strategic positioning. The UnifiedBus interconnect, which consolidates multiple protocols into a high-bandwidth, low-latency fabric, underpins the scalability of this new architecture, enabling theoretical configurations extending to a million NPUs, a scale unmatched by current competitors.
What to watch next
Industry observers will be attentive to Huawei's actual deployment and operational results of the Atlas 960E SuperPod and its multi-pod cluster proposals, especially as scaling from thousands to potentially millions of NPUs remains theoretical at this stage. Achieving stable performance and reliability at this unprecedented scale will be critical for commercial success.
Furthermore, Huawei’s progress in expanding its developer ecosystem, notably through the open-sourcing of its CANN software stack and becoming an official PyTorch accelerator backend, could accelerate adoption and innovation on its platform. Watching how Huawei’s software and hardware ecosystem competes against Nvidia’s entrenched position will be key to understanding shifts in AI compute dominance.