September 18, 2026, (Inside AI) — Huawei Cloud will start commercial operations of its Ascend 950 AI cluster cloud service in China on September 30, with global expansion set for November 30, CEO Zhou Yuefeng announced at Huawei Connect 2026. The service, built on Huawei's UnifiedBus interconnect, packs 1,024 Ascend 950 processors into a single cluster delivering up to 1 EFLOPS of FP8 and 2 EFLOPS of FP4 computing power, plus 256TB of globally addressable memory. More than 1,000 Ascend supernodes are already deployed, and the company says its next-generation Ascend 960 roadmap is ahead of schedule.
The news lands as global demand for AI compute continues to outstrip supply, and as US export controls push Chinese firms to build self-sufficient alternatives. Huawei's move signals that its domestic AI stack, from chips to cloud services, is moving from pilot projects into full commercial scale. The company is positioning the Ascend 950 as a turnkey system for training and running very large models without the need for additional cluster-level engineering.
Inside The 1,024-Processor Cluster
The Ascend 950 cluster is not a single chip but a tightly coupled system. Huawei's UnifiedBus interconnect links 1,024 AI processors so they behave like one machine. That design avoids the communication bottlenecks that often slow down distributed training. The system offers 1 EFLOPS in FP8 precision, a format widely used for training large language models, and 2 EFLOPS in FP4, a lower-precision mode that speeds up inference. The 256TB of globally addressable memory means every processor can access the same pool of data, which simplifies model partitioning and reduces idle time.
Huawei claims the system supports end-to-end training of large-scale models without extra adaptation. In practice, that means customers can load a model and start training without rewriting code for the cluster's specific topology. The company has not disclosed pricing, availability limits, or performance benchmarks against competing systems such as Nvidia's DGX SuperPOD or Google's TPU pods. Inside AI could not independently verify these performance claims.
Zhou Yuefeng framed the launch as a milestone for Huawei's cloud and AI strategy. "Huawei Cloud will begin commercial operations of its Ascend 950 AI cluster cloud service in China on Sept. 30 and expand it to global markets on Nov. 30," Zhou Yuefeng, CEO of Huawei Cloud, said at Huawei Connect 2026.
The global rollout on November 30 will test Huawei's ability to support customers outside China, where regulatory and supply-chain challenges are more complex. Huawei has not specified which countries will be included in the first wave.
Ascend 960 Roadmap Moves Ahead
Huawei also said its Ascend 960 roadmap is ahead of schedule. The Ascend 960DT is planned for the first quarter of 2027, while the 960PR is planned for the third quarter. The company did not detail the differences between the two variants, but the naming suggests different target workloads or form factors. The accelerated timeline may reflect strong demand for the 950 series or confidence in Huawei's supply chain.
More than 1,000 Ascend supernodes have already been deployed, and Ascend 950 supernodes have entered commercial use, according to the company. That deployment figure includes earlier generations, but it shows Huawei has built a substantial install base. The company is betting that its integrated hardware and cloud service will appeal to Chinese firms that face limited access to Nvidia's most advanced GPUs due to US export restrictions.
Huawei's push comes as China's government encourages domestic AI infrastructure. Several Chinese cloud providers and research institutes have been testing Ascend-based systems for tasks like natural language processing and computer vision. The commercial launch of the 950 cluster cloud service could accelerate that adoption by offering a managed service rather than requiring customers to build their own clusters.
Still, questions remain about software compatibility. Huawei's CANN toolkit and MindSpore framework are designed to work with Ascend chips, but many developers rely on Nvidia's CUDA ecosystem. Huawei has been investing in migration tools, but switching costs remain a barrier. The company's claim of "end-to-end training without additional cluster-level adaptation" may address some of those concerns, but independent benchmarks are scarce.
The Ascend 950 launch also highlights the growing divergence in global AI infrastructure. While US firms build on Nvidia and AMD, Chinese firms are increasingly turning to domestic alternatives. Huawei's success could influence other countries that seek to avoid dependence on US technology. For now, the company is focused on delivering the service on time and expanding its ecosystem.
Huawei Connect 2026, held in Shanghai, drew thousands of attendees and featured dozens of AI announcements. The Ascend 950 cloud service was the centerpiece. Huawei's stock is not publicly traded, but its cloud and AI business has become a key growth driver. The company reported strong revenue in its latest quarter, though it does not break out AI-specific figures.
As the November 30 global launch approaches, the industry will watch whether Huawei can replicate its domestic momentum abroad. The company's ability to attract international customers will depend on performance, price, and geopolitical factors. For now, the Ascend 950 cluster cloud service stands as Huawei's most ambitious AI infrastructure play to date.
