Huawei has decided to halt the global expansion of its AI chip business. Rotating chairman Eric Xu stated that the company currently lacks the production capacity to meet domestic demand within China. While limited shipments are being sent to select markets with high demand, a full-scale international rollout is not planned.
The company unveiled a new roadmap for its Ascend NPU series. The upcoming Ascend 960, 970, and 980 chips are designed to improve training and inference performance. By 2027, the Ascend 960DT is expected to reach 2 PFLOPS in FP8 training, while the Ascend 980, slated for 2029, aims for 7.2 PFLOPS in FP8 and 28 PFLOPS in FP4 performance.
Despite these improvements, Huawei's hardware remains behind Nvidia's current and future offerings in raw compute power. To compete, Huawei is focusing on massive system-level scaling. Its Atlas SuperPoDs can integrate up to 15,488 Ascend 960 NPUs, delivering up to 30 FP8 EFLOPS and 120 FP4 EFLOPS.
The reliance on large-scale clusters leads to significantly higher power consumption compared to Nvidia's architectures. Furthermore, production constraints at SMIC for 7nm and 6nm nodes continue to limit the supply of Huawei's AI hardware. Consequently, the company may prioritize providing cloud access to its hardware to promote its software stack.
