China’s AI stack is converging on MoE and larger supernodes
teortaxesTex · x · 2026-07-20
The post argues that China’s AI ecosystem has converged on MoE plus wide expert parallelism to make models run on weaker NPUs and with lower HBM bandwidth demands.
It also points to a broader hardware trend: system vendors are now designing much larger “supernode” domains. The attached image highlights several Chinese infrastructure demos, including Biren’s 1024-card scale-up cluster with optical interconnect, MetaX’s next-gen AI SuperNode, SUGON’s 8000 SuperCluster for training/inference/scientific computing, and Alibaba’s Zhenwu-890 / Panjiu AL128 SuperNode with 128 chips per cabinet and high-bandwidth interconnect.
More from Infra
- DeepSpeed shows AutoEP/AutoTP/AutoSP parallelism gains, hosts office hours — PyTorch · 2026-10-27
- Cloudflare Durable Objects validated as top architecture for AI agent WebSocket servers — tristanbob · 2026-10-27
- Gary Marcus: With Q3 data added, Anthropic's compute-efficiency growth shows deceleration — GaryMarcus · 2026-10-27
- Inference is just Kubernetes with a multi-trillion-dollar market, and the communities overlap — matt_slotnick · 2026-10-27
- vLLM eyes Rust migration as inference engines race to rewrite the stack — AccBalanced · 2026-10-27
- Resin lets an agent build standalone LLM inference engines; vLLM and SGLang optimize DeepSeek kernels differently — AccBalanced · 2026-10-27