Reportedly: Huawei Ascend 950DT 1024-card supernode cluster goes live in Wuhan, supporting trillion-parameter model training and agent inference
Reported by IT Home · not yet confirmed by the company or a second independent outlet. We update this page when it is.
Huawei's Ascend 950DT 1024-card liquid-cooled supernode cluster went live in Wuhan, IT Home reported. It delivers 1 EFLOPS FP8 compute and 256TB unified memory, supporting trillion-parameter model training and agent inference. The cluster was built by Hubei Science and Technology Investment and completed in one month.
China context
- Original name
- 华为
- Outside China
- Not stated in the sources yet
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can monitor whether Huawei releases APIs or SDKs for this cluster's capabilities, but no such availability is indicated.
- For investors
- Investors should track whether this deployment leads to increased adoption of Huawei's Ascend ecosystem in enterprise AI, as it may signal growing domestic compute self-reliance.
Translated from Chinese. Quotes and facts link to the original sources.
At the Huawei China Tour 2026 Hubei AI Innovation Development Conference in Wuhan, the Ascend 950DT 1024-card supernode cluster was officially lit up. The cluster uses 1024 Ascend NPUs with full liquid cooling, achieving 1 EFLOPS FP8 compute, 256TB global unified memory, TB-level inter-card bandwidth, and microsecond-level end-to-end communication latency. It leverages the Lingqu architecture for direct inter-card access and supports trillion-parameter large model training and agent inference. The project was completed in one month from start to lighting, and will provide local compute support for large model training and industry AI agent development in Hubei.
The cluster's 1 EFLOPS FP8 compute and 256TB unified memory enable training of trillion-parameter models, a scale typically requiring thousands of GPUs. The Lingqu architecture allows direct inter-card access, reducing communication overhead. The one-month deployment timeline suggests a pre-engineered, modular design.
Hubei-based AI developers and enterprises gain local access to high-end compute, reducing reliance on remote or foreign resources and potentially lowering latency and data sovereignty risks. This deployment strengthens Huawei's position in China's domestic AI infrastructure market against Nvidia-based alternatives.
The cluster provides a local, high-performance compute option for AI training and inference, potentially reducing costs and improving data compliance for enterprises in Hubei. It also demonstrates Huawei's ability to rapidly deploy large-scale AI infrastructure, which may attract further investment.
Watch for announcements of specific large models trained on this cluster or enterprise adoption of agent inference services in Hubei. The cluster's operational performance and utilization rates will indicate whether it meets the demand for trillion-parameter training.