Huawei announced on Thursday its plans to launch four new generations of its Ascend AI chips over the next three years, marking a departure from its customary secrecy and signaling its ambition to rival semiconductor leader Nvidia for the first time.
The Chinese technology giant, a key driver in the nation’s push for an independent semiconductor industry, disclosed the details at its annual Huawei Connect conference in Shanghai. The move highlights the company’s efforts to reduce dependence on a supply chain largely controlled by the United States.
Following the first-quarter debut of the Ascend 910C, Vice Chairman Eric Xu said Huawei aims to introduce two variants of its successor, the Ascend 950, in 2026, followed by the Ascend 960 in 2027 and the Ascend 970 in 2028.
Xu emphasized, according to the company, that computing power has always been fundamental to artificial intelligence, and is even more critical to China’s AI development.
Xu further revealed that the Ascend 950 will feature Huawei’s proprietary high-bandwidth memory, overcoming a longstanding technological bottleneck that previously forced reliance on South Korean and U.S. suppliers.
In addition, Huawei plans to introduce new high-performance computing supernodes, the Atlas 950 and Atlas 960, touted as the world’s most powerful. These will support 8,192 and 15,488 Ascend chips, respectively, a significant leap from the current Atlas 900 (CloudMatrix 384), which runs on 384 of Huawei’s 910C chips.
According to industry research group SemiAnalysis, Huawei’s latest products exceed Nvidia’s GB200 NVL72—which uses 72 B200 chips—on some metrics.
Huawei stated the new system’s “supernode” architecture enables ultra-high speed interconnection between chips, representing another competitive push in the increasingly heated global AI chip market.
