The argument for semiconductor export controls rests on a compound assumption: that the best chips are American-designed and Taiwan-manufactured; that restricting their export imposes a sustained compute gap on Chinese AI development; and that the gap, given the capital and coordination required to close it, compounds over training cycles into a durable advantage.
On February 11, 2026, Zhipu AI filed a detailed training report alongside the release of GLM-5. The model has 744 billion parameters, 44 billion active per inference under a Mixture-of-Experts architecture. It was trained on 28.5 trillion tokens. On the major benchmarks — MMLU, HumanEval, GPQA — it performs within single digits of GPT-5.2 and Claude Opus 4.5. The training cluster was 100,000 Huawei Ascend 910B processors, designed by Huawei's HiSilicon subsidiary and manufactured by SMIC — China's largest chipmaker — at a seven-nanometer process, using the MindSpore framework. Not a single NVIDIA chip.
The Ascend 910B delivers approximately 320 TFLOPS of FP16 performance, placing it between the A100 at 312 TFLOPS and the H100 at 989 TFLOPS dense FP16. In raw per-chip performance, the Ascend 910B is closer to the export-controlled A100 than to the export-controlled H100 that Western labs trained the current frontier generation on. Zhipu compensated by assembling a cluster large enough to close the gap in aggregate compute — and by doing the engineering work that a cluster of that scale on that hardware required: custom communication kernels, memory hierarchy optimizations, fault-tolerance systems for a deployment that loses nodes regularly.
The interesting thing is not that the cluster worked. It is what had to be built to make it work.
Zhipu could not rely on the software stack that NVIDIA has spent a decade building around CUDA. MindSpore has existed since 2019 but had not been tested at this scale. The training team rebuilt significant portions of the distributed training infrastructure from scratch or through adaptation. The result is that Zhipu now operates with a complete, domestically sourced AI stack: chips, framework, training infrastructure, and model weights. The export controls did not prevent this. They funded it — in the sense that the competitive pressure to close the capability gap internally justified the investment that building the parallel stack required.
The policy community is examining the geopolitical implications. The operational ones for builders elsewhere are less frequently examined.
GLM-5 was open-sourced under an MIT license. The weights are publicly available. For a team building in a market where NVIDIA hardware is expensive to import, difficult to source, or unavailable at the scale a serious deployment requires — which describes significant portions of Central Asia, Southeast Asia, and sub-Saharan Africa — the existence of a frontier-class open-weight model that runs on different hardware is a practical fact worth knowing. The Ascend 910B is sold through Huawei's commercial channels without the same export restrictions that apply to H100s and the Blackwell family. A data center in Ulaanbaatar or Almaty that can provision Ascend hardware can now run a frontier-class model on it. That option did not exist eighteen months ago.
There is a broader story here about what export controls produce when the target is determined and sufficiently resourced. The ceiling that was supposed to hold — the manufacturing constraint at sub-7nm — has not held at the volume required to train a frontier model, even if yields are lower than TSMC's equivalent process. The compute gap that was supposed to compound has instead produced a parallel stack. The lesson is not that export controls are useless. It is that they buy time, and that the time bought has to be used productively on the other side of the restriction to matter.
The semiconductor story has been written for three years as a story of restriction. GLM-5 is the counter-argument, written in 744 billion parameters and published under an MIT license.
The short of it.
On February 11, 2026, Zhipu AI released GLM-5 — a 744-billion-parameter frontier model trained on 100,000 Huawei Ascend 910B chips without any NVIDIA hardware — performing within single digits of GPT-5.2 and Claude Opus 4.5 on major benchmarks. US export controls produced a procurement detour rather than a capability gap: SMIC manufactured enough 7nm chips to provision a frontier training cluster, and Zhipu built the entire software stack to run it. The model is MIT-licensed and publicly available. For builders in markets where NVIDIA hardware is restricted or expensive, a frontier-class open-weight model that runs on commercially available non-NVIDIA silicon now exists.