SemiAnalysis analyzes Huawei’s production, and reports that the export controls are absolutely working to hurt their production of chips, which if we prevent smuggling will not only not scale in 2026 but will actively fall sharply to below 2024 levels, as they have been relying on purchases from Samsung that will soon run dry.
This is false for compute dies, which the article says will actually get manufactured in volume soon. The claim is about HBM specifically, which is necessary for decoding/generation (processing of output tokens), but not for pretraining or processing of input tokens.
So if they develop chips and systems without HBM, then at several times the cost they might be able to obtain a lot of pretraining capability soon. Not very useful directly without a lot of RLVR and inference compute though, and no announcements to the effect that this is being attempted.
This is false for compute dies, which the article says will actually get manufactured in volume soon. The claim is about HBM specifically, which is necessary for decoding/generation (processing of output tokens), but not for pretraining or processing of input tokens.
So if they develop chips and systems without HBM, then at several times the cost they might be able to obtain a lot of pretraining capability soon. Not very useful directly without a lot of RLVR and inference compute though, and no announcements to the effect that this is being attempted.