In today’s AI ecosystem, Chinese AI models face significant barriers—from chip restrictions and high costs to inefficient compute clusters and limited scalability. Despite algorithmic breakthroughs, domestic models lack a sovereign, scalable, and cloud-native compute foundation.
At HDC 2025, Huawei introduced a bold solution: the next-generation Ascend AI Cloud, powered by the CloudMatrix384 Super Node. This marks a shift from fragmented, card-based compute to a high-performance, elastic, and fully integrated AI infrastructure—built in China, for China’s AI future.
Why CloudMatrix384 Is a Game-Changer
- Compute Breakthrough: 384 Ascend NPUs + 192 Kunpeng CPUs interlinked via MatrixLink deliver 300PFlops—67% more than NVIDIA NVL72.
- Distributed Inference at Scale: Up to 2300 tokens/sec per card throughput; supports 384 experts in parallel MoE inference; >50% MFU efficiency boost.
- Massive Scalability: Cascade up to 432 super nodes (160,000 cards), achieving 100,000 PFlops compute to support trillion-parameter models.
- Poolable, Elastic Resources: IO and memory pooling, task-based compute slicing for optimal AI workload assignment and resource utilization.
- Robust Performance: 40-day stable training, <10min fault recovery, rapid failover supported by advanced ops and disaster recovery capabilities.
From Infrastructure to Industry Value
Huawei’s Ascend AI Cloud, now adopted by over 1,300 companies globally, enables industries to deploy AI at scale:
- Media: Sina’s “Smart Xiaolang” boosts inference delivery by 50%, model deployment speed doubles.
- AI Startups: SiliconMotion serves 100B tokens/day; MiniGPT’s performance improved by 2.7x using CloudMatrix384.
- Research: Chinese Academy of Sciences builds AI4S research platform atop Ascend cloud.
- Search & Internet: 360’s “Super Search” and Sina Weibo’s smart search are now powered by CloudMatrix384.
From Technical Leap to Strategic Imperative
While single-chip power lags behind globally, Huawei’s collective compute architecture closes the gap. CloudMatrix384 transforms AI compute from a constraint into a catalyst, enabling elastic compute services and creating a new AI foundation for China’s digital economy.
