Huawei says demand for its AI systems is already outstripping supply, including within China, limiting its ability to serve overseas customers. Even so, the company is accelerating its next Ascend processors while outlining Peerium, a computing architecture intended to connect as many as one million processors in a single AI system.
The near-term roadmap begins with Ascend 950DT, which is in testing and is being discussed with Chinese AI model developers for training workloads in 2027. Huawei now expects Ascend 960DT to be ready in the first quarter of 2027, three quarters earlier than previously planned, followed by Ascend 960PR in the third quarter.
- Ascend 950DT: in testing, with planned AI training use in 2027.
- Ascend 960DT: expected in the first quarter of 2027.
- Ascend 960PR: scheduled for the third quarter of 2027.
- Ascend 970: planned for 2028.
- Ascend 980: planned for 2029.
The strategy is increasingly focused on system-scale performance rather than a direct chip-for-chip contest with Nvidia. Huawei previously deployed CloudMatrix 384, which linked 384 Ascend 910C accelerators. It says more than 1,000 Ascend 910C-based systems have been installed, while the Ascend 950 generation has entered commercial use.
Peerium scales the system
Peerium uses Huawei's UnifiedBus interconnect to combine processors, memory, storage and networking equipment into a shared system. An Ascend 960 SuperNode is designed to support up to 4,096 accelerators, and multiple SuperNodes could form clusters containing hundreds of thousands of chips. Huawei's stated upper target is one million processors in a single computing assembly.
Huawei says communication between machines in conventional clusters can consume more than 40% of the time required to train a large model. The company is positioning Peerium and UnifiedBus as a way to reduce that bottleneck, allowing large numbers of less individually powerful accelerators to operate as a more competitive overall system.
The scale-up plan does not resolve Huawei's supply constraint or the broader software challenge. Eric Xu has said Ascend holds a larger share than Nvidia in China, but Huawei has not provided figures to substantiate that claim. Nvidia's CUDA ecosystem also remains a substantial advantage, as developers may need significant software adaptation to move workloads to another accelerator platform.
A recent TrendForce forecast estimated that domestic suppliers could account for nearly 90% of China's high-end AI accelerator segment this year as the market shifts toward local alternatives. Huawei's next test will be whether it can manufacture enough Ascend hardware and build the developer ecosystem needed to turn its large-cluster ambitions into deployed systems.
