In a landmark announcement that signals a new front in the global artificial intelligence arms race, Chinese technology giant Huawei has unveiled a detailed and ambitious roadmap for its next-generation Ascend AI chips.[1][2] At the Huawei Connect 2025 conference in Shanghai, the company laid out a multi-year plan to produce some of the world's most powerful AI computing clusters, a direct challenge to the market dominance of U.S.-based Nvidia.[3][4][5] The strategy, articulated by Huawei's deputy chair Eric Xu, leans heavily on creating massive, interconnected systems that can deliver world-leading performance at scale, even while acknowledging that China is likely to lag behind in the manufacturing of individual state-of-the-art semiconductors for the foreseeable future.[6][7][8] This move represents a significant step in China's pursuit of technological self-sufficiency amid ongoing international trade restrictions and sanctions.[9][10][11]

For the first time, Huawei has made its long-term chip plans public, detailing a trio of new Ascend chip series—the 950, 960, and 970—scheduled for release over the next three years.[8][1] The company intends to follow a brisk one-year release cycle, promising to double the computational power with each new generation.[7][1] The roadmap will kick off with the Ascend 950 series, set to become available in 2026 with two distinct variants: the 950PR, optimized for inference and recommendation tasks, and the 950DT, designed for model training and decoding.[6][7][8] The Ascend 950 boasts a significant leap in performance, delivering one PFLOP (a thousand trillion floating-point operations per second) of performance for FP8 data formats and offering an interconnect bandwidth of 2 TB/s, a 2.5-fold increase over its predecessor, the Ascend 910C.[6][7] Following this, the Ascend 960, launching in late 2027, will again double the computing power and memory capabilities.[6][8] The pinnacle of the new line, the Ascend 970, is slated to offer 4 PFLOPS in FP8 performance.[7][1] Critically, this hardware push is supported by breakthroughs in homegrown technology, including Huawei's development of its own high-bandwidth memory (HBM), a crucial component previously dominated by foreign firms.[1][12][13]

The core of Huawei's strategy extends beyond the specifications of any single processor. The company's primary focus is on a system-level approach, leveraging massive clusters of its domestically produced chips to achieve overwhelming computational power. During his keynote, Eric Xu conceded that individual Ascend chips might not match the performance of their top-tier Nvidia counterparts, but argued that when aggregated into "SuperPoDs," they can form the most powerful systems on the market.[8][14] These SuperPoDs are large-scale AI infrastructure products, designed to function as a single, powerful computer.[7][15] The upcoming Atlas 950 SuperPoD, scheduled for a 2026 launch, will integrate more than 8,192 Ascend chips, while the Atlas 960 SuperPoD, arriving a year later, will connect a staggering 15,488 chips.[3][8][15] These systems are poised to be scaled up even further into "SuperClusters," with plans for an Atlas 960 SuperCluster to contain over one million Ascend processors.[3][16][4] This approach is a direct and strategic workaround to US sanctions that have cut off Huawei's access to the most advanced semiconductor manufacturing technologies.[8][17]

This ambitious hardware roadmap is unfolding within a tense geopolitical climate and has profound implications for the global technology landscape.[14] The U.S. export controls, designed to slow China's technological progress, have effectively created a protected domestic market where a formidable competitor to Nvidia can flourish.[18] Huawei's push for a self-sufficient AI stack is central to Beijing's national strategy.[9][11][19] The company's current Ascend 910C chip, already being produced by China's top foundry SMIC using 7nm process technology, serves as a proof of concept.[20][21] This chip is notably used for inference in the powerful DeepSeek-R1 AI model, and Huawei's CloudMatrix 384 system, which links 384 of these chips, has demonstrated performance comparable to Nvidia's H100 in certain workloads.[22][23][24][25] While relying on older deep ultraviolet (DUV) lithography presents manufacturing challenges and lower production yields compared to cutting-edge EUV methods, it provides a viable path forward.[20] This dynamic is accelerating a potential bifurcation of the global AI industry into two distinct ecosystems: a Western sphere reliant on Nvidia and its partners, and a parallel, China-centric supply chain built around Huawei's Ascend platform.[11][19]

In conclusion, Huawei's announcement at Connect 2025 is more than a product launch; it is a declaration of technological resilience and strategic intent. The company has laid out a clear plan to circumvent geopolitical restrictions not by trying to build a single superior chip, but by redefining the scale of AI infrastructure. By focusing on the collective power of massive, interconnected clusters powered by an increasingly capable and independent line of Ascend processors, Huawei is positioning itself to meet China's surging demand for AI computing power.[3][10] This bold strategy ensures that while the global semiconductor rivalry intensifies, the future of AI development will be shaped just as much by the architecture of the supercomputer as by the power of the chip inside it.