Huawei has laid out its most ambitious roadmap yet for high-performance computing, announcing three new generations of Ascend chips and an expanded vision for large-scale AI infrastructure. The announcement came at the company’s annual Huawei Connect 2025 conference in Shanghai, where executives emphasized resilience, innovation, and open collaboration amid ongoing global semiconductor challenges.
A New Era for Ascend Chips
Eric Xu, Huawei’s rotating chairman, called 2025 a “memorable year” for the company, citing January’s debut of the DeepSeek-R1 AI model as a milestone. He acknowledged that China may remain behind in chip manufacturing process nodes “for a relatively long time,” but said Huawei’s strategy is to strengthen architecture, system design, and software ecosystems.
The new Ascend chip lineup will roll out over the next several years:
- Ascend 950 (2026): Available in multiple versions, the 950 will support advanced data formats such as FP8 and MXFP8, with performance rated up to 2 PFLOPs. It will feature 2 TB/s interconnect bandwidth—2.5 times that of its predecessor, the Ascend 910C.
- Ascend 960 (2027): Offering double the computing power and memory capacity of the 950, the 960 will also introduce Huawei’s proprietary HiF4 data format for improved precision.
- Ascend 970 (2028): Huawei’s most powerful chip yet, expected to reach 8 PFLOPs of FP4 performance and 4 TB/s interconnect bandwidth, with further specifications still in development.
Building SuperPods and SuperClusters
Huawei is positioning these chips to power massive AI and general computing clusters. Beginning in late 2026, the company will launch the Atlas 950 SuperPoD, built on Ascend 950 chips. Huawei claims this system will significantly outpace competitor NVIDIA’s upcoming NVL144, with nearly seven times the compute power.
The roadmap also includes:
- Kunpeng 950 processors for general-purpose computing, available in early 2026, with up to 192 cores and 384 threads.
- The TaiShan 950 SuperPoD, billed as the world’s first general-purpose computing supercluster, launching alongside Kunpeng 950.
- UnifiedBus 2.0, a fully open interconnection protocol designed to link massive systems together, forming next-generation SuperClusters.
By 2027, Huawei plans to release the Atlas 960 SuperCluster, integrating over one million NPUs and delivering 4 ZFLOPs (10²¹ operations per second), which would make it one of the most powerful computing systems ever built.
Open Source and Collaboration
In addition to hardware, Huawei is open-sourcing major parts of its ecosystem, including the openPangu AI models and the Mind SDKs, and releasing technical specifications for UnifiedBus 2.0 to developers worldwide. Xu said these moves are aimed at fostering global collaboration and preparing for “surging demand for computing, both today and tomorrow.”
