Alibaba Cloud launched the M890 supernode instance at WAIC this week. Three details worth pulling apart.

The hardware. ICN Switch 1.0 scales up from 16 to 64 cards with 800GB/s inter-card interconnect. FP8/FP4 low-precision support. One node handles 10-trillion-parameter MoE inference. Training performance is 3x the previous generation for autonomous driving and embodied AI workloads.

The delivery model. This is what actually matters. Alibaba says this is the first time it offers supernode-class compute through the public cloud. Previously, this tier required a custom deployment. Now it has a product listing in Ulanqab (invite-only beta, but the form factor is public cloud). That shift from bespoke engineering to standardized SKU is the pattern I keep watching: compute infrastructure becoming a commodity layer.

The capacity signal. Single-cluster support for 130K cards, extensible to million-card scale. PD-separated deployment for inference workloads. 99.7% availability with minute-level fault recovery. This is Alibaba building the runway for Qwen’s next-generation models while selling excess capacity as a product.

The domestic stack (custom interconnect, CIPU 2.0, rebuilt parallel storage) is quietly reaching the scale where it competes on architecture, not just availability.