They also officially disclosed, for the first time, that they have achieved running inference on over 100k Chinese accelerators. They cite the GLM-5.3 powered internal infra agent to have halved the time needed to optimize the infra.
Although their English translated report said "a cluster of...", we believe this is a translation ambiguity, as the Chinese version only said "100k-level", without mentioning that they are in one cluster. In a recent note to our clients, we broke down where we think GLM-5.3 could be served, and how. (2/3)
https://t.co/H1RSZwHlSA