China’s AI hardware scene now has a 10K-GPU cluster and a 236B MoE model
teortaxesTex · x · 2026-07-22
- The post highlights how quickly parts of China’s AI hardware scene have gone from quiet to visibly ambitious.
- It points to a 10K-GPU cluster and a 236B MoE model trained on 25T tokens, framing it as a leap from “nothing visible” to something near the frontier.
- The author also asks whether anything at this scale has been trained on AMD yet, mentioning Zaya-74B 4AB as the only comparable case they know.
Related event: Moore Threads Pretrains 236B MoE on 10,000-GPU Cluster(2 posts)→
More from Infra
- Localmaxxing says it now tracks 3,500+ local AI runs across 444 model setups — LocalMaxxing · 2026-07-22
- State of Local AI 2026 teaser promises findings after months of work — StefanoGogioso · 2026-07-22
- UB-Mesh proposes a hierarchical full-mesh network for AI training clusters — bookwormengr · 2026-07-22
- NVIDIA open-sources a GPU-accelerated medical physics simulation framework — nordicinst · 2026-07-22
- A pure Triton W4A16 GEMM claims 1.1x-1.3x decode speedups over cuBLAS FP16 — bassrehab · 2026-07-22
- OpenAI details Project Camellia data center with $80M community plan — OpenAINewsroom · 2026-07-22