- New M890 supernode can run trillion-parameter models, with first instances deployed in Ulanqab
- The launch signals China’s AI chip race is moving from individual chips to high-speed computing clusters
Alibaba Cloud has launched its Lingjun Zhenwu (灵骏真武) M890 supernode instance, giving enterprise customers cloud access to high-performance AI computing without having to build their own data centers.
The service went on sale in Ulanqab, Inner Mongolia, on August 12. Customers can provision 64-card high-speed interconnected computing units through the cloud, with enough capacity to support inference for mixture-of-experts (MoE) models with up to 100 trillion parameters, according to Alibaba Cloud.
The M890 supernode is also the first supernode architecture in China reported to have successfully run AI models exceeding 2 trillion parameters.
Both Kimi K3 and Alibaba’s Qwen3.8-Max have used the infrastructure to provide services.
At the hardware level, the Zhenwu M890 is Alibaba’s next-generation training-and-inference AI accelerator developed by its chip subsidiary T-Head. It natively supports low-precision FP8 and FP4 computing.
Using Alibaba’s proprietary ICN Switch 1.0 chip, the Lingjun Zhenwu M890 expands the scale-up interconnection from 16 cards in the previous generation to 64 cards. Card-to-card bandwidth reaches 800 GB/s, while the shared memory pool reaches 9 TB.
3X higher performance
Compared with the previous-generation Zhenwu 810E, training performance is up to three times higher in workloads such as autonomous driving and embodied AI, while performance in agentic inference can improve by up to 1.5 times.
The supernode runs on Alibaba Cloud’s Lingjun computing platform. Its HPN 8.0 integrated training-and-inference network can support up to 130,000 heterogeneous accelerators in a single cluster and scale toward one million cards.
Alibaba Cloud also says customers can receive a supernode instance within just one hour of placing an order.
Why Ulanqab?
Ulanqab, the first deployment location, is one of Alibaba Cloud’s five hyperscale data-center regions. About 90% of its electricity comes from renewable sources, giving high-density AI computing a lower-carbon power option.
The data center also uses Alibaba Cloud’s modular architecture, with power supply, cooling, security, intelligent systems and fire protection achieving an overall modularization rate of 90%.
The approach is designed to significantly shorten data-center construction and deployment cycles.
Alibaba Cloud plans to roll out the M890 supernode instance in additional regions and more than double its global production capacity for modular data centers this year.
From AI chips to supernodes
The launch illustrates a broader shift in China’s AI infrastructure race: competition is moving beyond the performance of individual AI accelerators toward the ability to connect thousands of chips into high-performance computing systems.
The combination of 64-card interconnection, 800 GB/s bandwidth and a 9 TB memory pool gives domestically developed computing infrastructure the ability to support increasingly large AI models and demanding agentic workloads.
For international observers, the M890 also highlights how Chinese cloud providers are attempting to build an integrated alternative spanning AI chips, networking, supernodes, data centers and cloud services.
Whether China’s domestic supernode ecosystem can achieve comparable performance, software compatibility and scale to global alternatives will be an important test of the country’s long-term AI computing independence.
