Z.ai, the lab behind the GLM family of open-weight language models, walked away from third-party inference providers and built its own serving stack for GLM instead.
The company published an engineering post on September 17 laying out the reasoning and process behind that move. It was submitted to Hacker News, where it turned into one of the more heavily discussed threads on the site, pulling in 256 points and 209 comments. That level of engagement is notable for a piece about serving infrastructure rather than a new model release.
Building inference infrastructure from scratch is a heavy lift most AI labs skip, leaning instead on open-source serving frameworks like vLLM and SGLang or on cloud inference APIs. Z.ai's decision to do it in-house suggests inference cost and control are becoming a competitive lever for open-weight model makers, not just a backend detail.
Whether that investment shows up as faster or cheaper GLM access for developers is still unproven. For now, the clearest result is how many engineers stopped to read about the plumbing.