At the recent Huawei Connect 2025 event in Shanghai, Chinese tech leader Huawei unveiled ambitious plans for its next-generation Ascend chip series, signaling a significant leap forward in AI hardware and infrastructure design. The new Ascend chips are poised to power some of the most powerful AI clusters in the world, positioning Huawei at the forefront of high-performance computing for artificial intelligence applications.
Introduction: A Pivotal Year for Huawei’s AI Ambitions
Eric Xu, Huawei’s Deputy Chairman, highlighted 2025 as a “memorable year,” marked notably by the launch of DeepSeek-R1 earlier in January. Despite ongoing challenges such as trade embargoes and tariffs that restrict access to advanced semiconductor manufacturing process nodes, Huawei remains committed to overcoming these obstacles through innovative infrastructure design and open-source initiatives.
The New Ascend Chip Series: 950, 960, and 970
Huawei plans to release three series under the Ascend family: the 950, 960, and 970, each delivering stepped improvements in computational power and efficiency.
Ascend 950
- Variants: 950PR and 950DT, built from the same die.
- Support for advanced low-precision data formats including FP8, delivering 1 PFLOP performance, and MXFP8 rated at 2 PFLOPs—where a PFLOP equals 1,000 trillion floating-point operations per second.
- Enhanced vector processing capabilities and more granular memory access, optimized down to 128-byte chunks compared to the previous 512 bytes.
- Interconnect bandwidth of 2 TB/s, 2.5 times higher than the earlier Ascend 910C model.
- Availability: 950PR slated for Q1 2026, and 950DT for Q4 2026.
Ascend 960
- Expected launch in Q4 2027.
- Twice the computing power, memory bandwidth, memory capacity, and interconnect ports compared to the Ascend 950.
- Support for Huawei’s proprietary HiF4 data format, which enhances precision beyond typical FP4 technologies.
Ascend 970
- The flagship chip, targeted for release in Q4 2028.
- Projected highlights include 4 TB/s interconnect bandwidth, 8 PFLOPs of FP4 performance, and substantially larger memory capacity.
- Huawei continues refining the final specifications to push the boundaries of AI processing power.
SuperPods and SuperClusters: Redefining AI Computing Scale
Huawei’s strategy leverages clusters of Neural Processing Units (NPUs) bundled into scalable computing units called SuperPods, which are set to debut in Q4 2026. The Atlas 950 SuperPod, featuring Ascend 950DT chips, will serve as Huawei’s cutting-edge offering for hyperscale AI workloads.
According to Huawei, the Atlas 950 SuperPod is expected to contain 56.8 times more NPUs than NVIDIA’s NVL144 system—NVIDIA’s own SuperPod analogue—delivering nearly seven times the processing power. Even as NVIDIA plans to release their NVL576 system in 2027, Huawei’s new SuperPods aim to maintain competitive superiority in raw AI computing capacity.
General Purpose Computing: Kunpeng 950 Processors and SuperPods
Complementing its AI chips, Huawei is preparing to launch two Kunpeng 950 general-purpose processors in early 2026, configured with 96 cores & 192 threads and 192 cores & 384 threads respectively. These processors will underpin the TaiShan 950 SuperPod—the world’s first general-purpose computing SuperPod—also slated for Q1 2026 release, designed to seamlessly intersect with AI workloads.
UnifiedBus 2.0: Open-Source Connectivity Protocol
The AI and general computing SuperPods will utilize UnifiedBus 2.0, the next-generation open interconnection protocol succeeding UnifiedBus 1.0. Since the Atlas 900 A3 SuperPod’s deployment in March 2025—with over 300 installations globally—UnifiedBus has proven effective at scaling AI compute clusters.
UnifiedBus 2.0’s specifications are immediately available to the developer community, reinforcing Huawei’s commitment to open-source collaboration. This protocol will facilitate super-scaled cluster configurations, comprising multiple SuperPods connected into SuperClusters.
Atlas 950 and 960 SuperClusters
- The Atlas 950 SuperCluster will offer 2.5 times more NPUs and 1.3 times greater computing power than xAI’s Colossus cluster, currently recognized as the most powerful AI computing cluster worldwide.
- Coming in late 2027, the Atlas 960 SuperCluster aims to incorporate over one million NPUs, delivering a staggering 4 ZFLOPS FP4 performance (a ZFLOP equals 1021 floating point operations per second).
Global Context and Industry Impact
Huawei’s announcements arrive at a critical juncture in the AI hardware race. Currently, NVIDIA and AMD dominate the high-performance AI chip market, but the introduction of Huawei’s Ascend chips and innovations in interconnect protocols represent an evolving ecosystem of AI-specific architectures designed to meet the surging global demand for computing power.
Industry analysts estimate that the global AI chip market will exceed $120 billion by 2030, growing at a compound annual growth rate (CAGR) of over 40% (Source: MarketsandMarkets). Huawei’s advancements in Ascend processors, SuperPods, and SuperClusters are poised to influence this trajectory significantly, especially with China’s focus on localized semiconductor development amid geopolitical challenges.
Key Highlights
- Ascend 950 series —launching 2026 with support for FP8 and MXFP8, robust vector processing, and enhanced memory access.
- Ascend 960 series —expected in 2027, doubling computing capacity and introducing the proprietary HiF4 format.
- Ascend 970 series —set for 2028, targeting massive FP4 performance and expanded memory.
- SuperPods and SuperClusters —enabling hyperscale AI through clusters with tens of thousands to over a million NPUs.
- UnifiedBus 2.0—open-source interconnect standard accelerating cluster scalability and openness.
- Integration of Kunpeng 950 general-purpose processors for versatile AI and computing workloads.
Conclusion
Huawei’s strategic roadmap for the Ascend chips and associated AI infrastructure reveals a calculated effort to match and surpass current industry standards amid a challenging global semiconductor landscape. By combining advanced chip design, open-source software models, and scalable hardware architectures like SuperPods and SuperClusters, Huawei aims to meet the rapidly escalating demand for AI computing power across industries.
As more enterprises invest in AI-driven solutions, innovations like Huawei’s Ascend chips and the open UnifiedBus 2.0 protocol will play essential roles in shaping the future of AI hardware and large-scale computational capacities.
Image source: Eric Xu, Huawei