摘要：At Huawei Connect 2025 on September 18th, Huawei announced its three-year roadmap for computing power, shifting its strategy to a

At Huawei Connect 2025 on September 18th, Huawei announced its three-year roadmap for computing power, shifting its strategy to a "super node + cluster" architecture. Through multi-chip collaboration to offset process limitations, the company is ushering in a new model of AI computing power in China.

Over the next three years, Huawei will launch four series of Ascend chips, and follow a 1-year release cycle and double compute with each release. Based on these chips, the company will introduce the Atlas 950 SuperPoD and Atlas 960 SuperPoD, supporting 8,192 and 15,488 Ascend chips, respectively. The resulting cluster will exceed 500,000 NPUs, scaling up to one million NPUs, making it one of the most powerful computing clusters worldwide.

This breakthrough stems from Huawei's "course shift" after TSMC cut off supply in 2020—abandoning the obsession with single-chip performance and turning to system-level innovation. At the same time, Huawei has launched the interconnect protocol "Lingqu" to drive the formation of a complete industry chain and a self-sufficient AI ecosystem, helping reduce reliance on external sources.

By the first quarter of this year, the eight computing hubs of China's "East Data, Western Computing" initiative had reached a total computing power of 215.5 EFLOPS, accounting for over 70% of the national total. China's AI industry is now at a critical juncture of a "computing breakthrough." Huawei's strategy not only addresses the issue of computing power supply but also establishes a pathway moving from technological innovation through industrial application to ecosystem building, helping reduce the costs of AI transformation across industries.

As Huawei's home base, the Guangdong-Hong Kong-Macao Greater Bay Area (GBA) has a vibrant AI industry. Leveraging multiple intelligent computing centers, Guangdong benefits from a complete electronic information industry chain and abundant application scenarios. Empowered by Huawei's solutions, the region's computing infrastructure is set to make a qualitative leap, driving broader application of AI technologies and enabling China to gain a stronger voice in the global AI race.

大湾区打造全球最强算力集群

9月18日“华为全联接大会2025”上，华为官宣未来三年算力演进路线，战略转向“超节点+集群”架构，以多芯片协同弥补制程差距，开启中国AI算力新形态。

未来三年，华为将推四系列昇腾芯片实现“一年一代，算力翻倍”，并基于芯片发布Atlas 950/960 SuperPoD超节点，可分别支持8192及15488张昇腾卡，构建的集群算力规模超50万卡、达百万卡，跻身全球最强。

这一突破源于2020年台积电断供后的“换道”——放弃单芯片性能执念，转向系统级创新。华为同步开放“灵衢”互联协议规范，推动形成完整产业链与自主AI生态，助力摆脱外部依赖。

截至今年一季度，我国“东数西算”八大枢纽节点算力总规模达215.5EFLOPS，占全国总算力的70%以上，中国AI产业正处“算力突围”关键期。华为该战略不仅解决算力供给问题，更构建“技术创新→产业落地→生态共建”路径，降低行业AI转型成本。

作为华为扎根地，粤港澳大湾区AI产业活跃。广东依托多座智算中心，凭借完整电子信息产业链与丰富应用场景，在华为方案赋能下，未来算力基础将实现质的飞跃，推动更多AI技术落地成长，助力中国在全球AI竞争中掌握更多话语权。

策划｜记者 吴江

来源：羊城派