After the H20 generation, Zhipu is using GLM-4.6 to push further into support for Chinese AI chips.
Background
U.S. chip export controls continue to tighten. H20 is already a restricted version, and even that class of hardware could face further limits. Chinese AI companies therefore need to prepare for scenarios in which they must rely entirely on domestic chips.
The Domestic AI Chip Stack
The main options now include Huawei Ascend, Cambricon, Hygon and Biren. Ascend has the most complete ecosystem but still trails NVIDIA on performance. Cambricon has strengths in inference chips, Hygon follows an x86-compatible path, and Biren is building around a GPU architecture at an earlier stage of development.
Zhipu has chosen a multi-track adaptation strategy. Instead of tying GLM-4.6 to one domestic chip vendor, it is trying to make the model run across several Chinese platforms.
Technical Challenges
Adapting to domestic chips is not simply a matter of swapping hardware. Operator libraries differ sharply: NVIDIA's CUDA ecosystem reflects decades of accumulation, while domestic libraries remain less mature. Communication efficiency can become a bottleneck in multi-card parallel training. Precision and stability also vary, because the same training process may behave differently across chip platforms.
Strategic Significance
For Zhipu, domestic chip adaptation is a risky but necessary investment. In the short term, it raises R&D costs and may deliver weaker performance than NVIDIA platforms. In the medium term, if controls tighten further, companies that adapted early will have a first-mover advantage. In the long term, once the domestic chip ecosystem matures, the companies with the deepest accumulated experience inside that ecosystem should benefit most.
Doing the best research on the best hardware and building good-enough products on obtainable hardware are different problems. Zhipu is placing bets on both sides, and that strategy looks reasonable.
Sources: CocoLoop, Zhipu AI official materials, chip industry analysis