I keep seeing claims that Zai’s GLM-5.2 model was trained on Huawei chips. There’s no evidence for this. If it were true, it would be huge news. It would mean China had mostly “solved” the problem of U.S. export controls on AI chips. But if it were true, we would’ve heard about it. I would expect Zai, Huawei, or Chinese official sources to loudly claim victory, like they did with tau-scaling. And we haven’t heard anything like that. There’s been confusion about this since at least Zai’s GLM-5 model. Zai has been optimizing their models to run on Chinese domestic chips, including Huawei Ascend. And they did train one image model—GLM-Image— on Huawei chips. But while Chinese models are increasingly optimized to run on Chinese AI chips, including DeepSeek-V4, very few are actually trained on Chinese chips. The confusion over training vs. inference on Chinese chips will likely continue. At some point, we will see sizable models being fully trained on Chinese chips. But in the meantime, watch out for confusing claims like these without clear evidence.