Zhipu AI GLM‑5.3‑Flash runs on 100k Chinese chips for cheap AI

GLM‑5.3‑Flash is notable on two fronts: it is an open‑weight, frontier‑class multimodal model with aggressive pricing, and it is being served entirely on Chinese AI chips at very large scale. That combination is a concrete demonstration that China can now deliver high‑end inference without relying on Nvidia, at least for sparse, long‑context models aimed at code, documents and agents rather than the absolute training frontier.

For the race to AGI, this chips‑plus‑model story matters because it undercuts two assumptions that have favored U.S. labs. First, that cutting‑edge model quality requires access to Nvidia’s latest GPUs; GLM‑5.3‑Flash’s performance and adoption via OpenRouter suggest otherwise for many workloads. Second, that closed, proprietary models will control the most economically important segments; instead, we are seeing open‑weight Chinese models competing head‑on with Claude Opus–class systems on capability while dramatically undercutting on price.

Strategically, this advances China’s push for compute sovereignty and gives enterprises everywhere more leverage in negotiations with U.S. providers. It also shows how quickly the open ecosystem can absorb and normalize sophisticated sparse architectures. If Chinese vendors can keep iterating on both chips and models, they will remain a serious counterweight in both cost and capability, even if they never quite match Western labs at the absolute frontier.