August 12, 2026, (Inside AI) —
Alibaba Cloud has launched its M890 AI supernode in China, marking a significant step in enterprise-grade AI infrastructure. The first deployment region is Ulanqab, Inner Mongolia.
Enterprise customers can now provision 64-card, high-speed-interconnect computing units directly through the cloud. This eliminates the need to build and maintain private data centers for large-scale AI workloads.
The M890 is purpose-built for inference on mixture-of-experts (MoE) models with up to 10 trillion parameters. Two flagship models, Kimi K3 and Qwen3.8-Max, are already serving traffic on the new instance.
This launch positions Alibaba Cloud to capture demand from Chinese enterprises racing to deploy massive language models. MoE architectures, which activate only a fraction of parameters per query, demand high-bandwidth interconnects that the M890 provides as a cloud-native service.
MoE Inference Demands a New Hardware Class
Mixture-of-experts models have become the dominant architecture for frontier AI systems. By splitting a model into specialized sub-networks, they reduce per-token computation but require rapid communication between accelerators. The M890's 64-card configuration addresses this bottleneck with high-speed interconnects optimized for all-to-all communication patterns typical of MoE routing.
Industry analysts note that 10-trillion-parameter models represent a threshold where on-premise infrastructure becomes cost-prohibitive for most enterprises. Alibaba's cloud-based approach shifts capital expenditure to operational expenditure, mirroring strategies by global hyperscalers like AWS and Google Cloud. However, Alibaba's focus on domestic deployment in Ulanqab reflects China's data sovereignty requirements and the strategic importance of the Beijing-Tianjin-Hebei economic zone.
The Ulanqab Advantage and Market Implications
Ulanqab has emerged as a key data center hub due to its cool climate, abundant renewable energy, and proximity to major northern Chinese markets. Alibaba Cloud operates multiple availability zones in the region, and the M890 launch reinforces its commitment to serving AI workloads from this location.
Competing Chinese cloud providers, including Tencent Cloud and Huawei Cloud, have also invested heavily in AI supercomputing. Tencent's Hunyuan cluster and Huawei's Ascend-based offerings target similar workloads. Alibaba's differentiation lies in the M890's tight integration with its ModelScope platform and the immediate availability of Kimi K3 and Qwen3.8-Max, which are among China's most widely adopted open-weight models.
The M890's launch comes as Chinese AI companies face export restrictions on advanced GPUs. While Alibaba has not disclosed the specific hardware powering the M890, industry sources suggest it likely uses a combination of NVIDIA H800 GPUs and domestic alternatives. This hybrid approach would align with China's push for technological self-sufficiency while maintaining competitive performance.
For enterprises, the M890 offers a turnkey solution to serve models at scale without managing complex hardware. Pricing details remain undisclosed, but Alibaba Cloud typically offers reserved and on-demand options. Early adopters in finance, healthcare, and autonomous driving are expected to test the instance's capabilities for real-time inference tasks where latency and throughput are critical.
Alibaba Cloud's broader strategy involves building an AI ecosystem that spans infrastructure, model development, and application deployment. The M890 supernode is a critical piece of this vision, enabling the company to compete not just on price but on performance for the most demanding AI workloads. As Chinese enterprises accelerate AI adoption, the availability of such specialized cloud instances will likely influence vendor selection and shape the competitive landscape.
