How GLM built its own inference infrastructure
Points and comments are a snapshot, not live.
GLM built production inference on 100,000 Chinese accelerators partially using an AI agent.
Z.ai deployed GLM-5.3-Flash on a cluster of over 100,000 Chinese-made AI accelerators, achieving production readiness in under two weeks. An Infra Agent powered by GLM-5.3 performed much of the optimization work, including memory optimizations (W8A8 quantization, mixed-precision cache, Layer Split) and an Encode-Prefill-Decode disaggregated architecture, tripling throughput relative to baseline. The agent relied on dense, attributable feedback rather than sparse end-to-end metrics. Specific fixes included correcting TF32 precision in KDA kernel context parallelism and releasing the Python GIL in DeepEP to avoid blocking KV Transfer concurrency.
What commenters are saying
Commenters focused on the scale and implications of the Chinese hardware deployment. One noted most had guessed this when GLM offered 100 trillion free tokens. Another called it an asteroid-level event for Western labs, citing China's cheap power and domestic chip capability. Several discussed GLM's pricing: legacy users got highly subsidized plans, now replaced by tiered token quotas (Max plan ~$168/month for ~$1,100 of GLM-5.3 tokens at 95%+ cache hit rates). Another claimed GLM illegally routed user requests through Anthropic's Opus in a distillation attack, leaking Chinese state secrets; a reply countered that breaking TOS is not per se illegal.