What is Huawei Cloud’s Ascend 950 AI cluster service, when does it become available in China and abroad, what computing capacity and scaling does its UnifiedBus architecture offer, and how do Huawei’s claims about overtaking Nvidia in China, chip supply, CUDA’s declining importan

Huawei Cloud’s Ascend 950 service lets customers use a large, interconnected AI-computing cluster through the cloud. It began commercial service in China on September 30, 2026; Huawei says access outside China will follow on November 30. The overseas date is an announced plan, not evidence of broad capacity there. 93

• Capacity and scaling: UnifiedBus links 1,024 Ascend 950 cards into a cluster. Reported Huawei specifications put its peak capacity at about 1 exaflop in FP8 or 2 exaflops in FP4, with 256 TB of addressable memory; Huawei also describes scaling to clusters of up to 100,000 cards. These are stated capabilities, not independently verified performance results. 1311
• Competitive position: Huawei’s Eric Xu says Ascend has surpassed Nvidia in Chinese market share. That is Huawei’s assessment, not a verified finding; restrictions on Nvidia’s highest-end sales have helped create an opening for domestic suppliers. 35
• Launch constraint: Xu also says Huawei cannot make enough AI-computing equipment to satisfy demand in China and is limiting overseas sales. Thus a November overseas launch should not be read as a promise of plentiful overseas compute. 3
• CUDA: The available evidence is insufficient to verify the specific claim that CUDA’s importance is declining. The launch demonstrates Huawei’s effort to offer an alternative to Nvidia, but it does not by itself establish that developers no longer depend on CUDA. 3