Chinese tech giant Huawei has commercially launched its Ascend 950 chip-based compute cluster service for local developers, with plans for a global rollout in late November. The company confirmed that sales of its domestic processors have surpassed rival Nvidia within the Chinese market amid an accelerating shift toward local alternatives.
- Commercial supercomputing rollout and reaffirmation of domestic leadership
- Exceptional technical specifications and massive architectural expansion
- Domestic sales dominance and the fading of software barriers
- Frequently asked questions
Commercial supercomputing rollout and reaffirmation of domestic leadership
Huawei Cloud announced the availability of its “Ascend 950” AI chip-based supercomputing cluster service for commercial clients in China on Wednesday, September 30, while scheduling its rollout for overseas market customers on November 30. Huawei Cloud CEO Zhou Yufeng officially revealed these two dates during the Huawei Connect 2026 conference held in Shanghai from September 17 to 19.
This step allows AI developers and Chinese enterprises to rent the latest locally developed computing capabilities as an integrated cloud service ready for immediate use. The launch coincides with official confirmations from Huawei leadership that its domestic processors have successfully captured the top sales spot and outperformed American firm Nvidia’s market share inside China, amid government directives supporting technological independence.
Exceptional technical specifications and massive architectural expansion
According to specialized platform TechNode, the new cloud clusters rely on Huawei’s innovative interconnect architecture known as the “Unified Bus.” Each independent cluster unit connects 1,024 integrated processing cards, delivering massive computing power of up to 1 exaflop in 8-bit floating point (FP8) mode, and approximately 2 exaflops in FP4 mode, supported by a massive 256 terabytes of RAM that the entire cluster can address directly and uniformly.
Huawei affirms that the system is capable of training ultra-large AI models from scratch through final inference without the need for any code modifications or additional cluster-level adjustments. The company revealed that it has deployed over 1,000 Ascend supercomputing nodes, confirming that Ascend 950 nodes have already entered actual commercial use.
Meanwhile, the Shanghai Metals Market platform reported that the largest giant cluster, the “Atlas 950 Super Cluster,” which relies on all-optical links within the Unified Bus architecture, has the capacity to expand and grow to 500,000 processing cards. Conversely, The Register published complementary data indicating that Huawei is currently deploying a cluster featuring 256,000 cards, with a modern architectural design allowing future expansion up to a million connected processors in a unified environment.
Domestic sales dominance and the fading of software barriers
This launch coincided with prominent statements from Huawei’s rotating chairman, Eric Xu, in which he confirmed that sales of Ascend chips have surpassed Nvidia sales on Chinese soil. Xu stated, as reported by The Register: “It is very difficult to gather precise data regarding Nvidia’s market share inside China, but according to the reliable figures we have tracked, Ascend sales have already outpaced those of Nvidia.”
Although Xu acknowledged that Huawei’s chips may lag technically behind newer American models, he emphasized that supply chain stability and reliability represent the decisive advantage, affirming: “Even if our chips are theoretically less advanced, guaranteed supplies and uninterrupted continuity are assured.” Despite Washington easing restrictions to allow Nvidia to sell “H200” models to select clients, Nvidia officials stated that those shipments represented less than 1 percent of data center revenue, at a time when Beijing continues to exert pressure to push cloud operators toward exclusive reliance on Chinese processors.
The planned overseas expansion in November will be limited in scope, as Xu explained that the company is focusing on satisfying massive domestic demand and does not plan a comprehensive international expansion, restricting foreign sales to specific clients lacking alternatives. Meanwhile, Liao Heng, chief scientist at Huawei’s HiSilicon unit, confirmed that the software barrier that once hindered the adoption of their processors has begun to gradually fade, pointing out that Nvidia’s proprietary CUDA platform no longer holds the same decisive importance for leading AI labs compared to two years ago.
Frequently asked questions
Question: What is the computing capacity provided by a single Ascend 950 server cluster?
Answer: The cluster consisting of 1,024 cards delivers computing power of up to 1 exaflop in FP8 mode and 2 exaflops in FP4 mode.
Question: Does Huawei plan to introduce these cloud processors to global markets?
Answer: The company intends to provide the cloud service outside China starting November 30 on a limited scale to meet the needs of specific clients.
Question: Why do Chinese developers prefer relying on Ascend chips despite technical differences?
Answer: This is due to ensuring continuous supplies and the stability of domestic manufacturing away from the volatility of American sanctions and trade restrictions.
