The quiet story of May, buried under Google I/O and the Anthropic round, is that Zhipu trained GLM-4.7 on 100,000 Huawei Ascend 910B processors — and shipped it as an open-weights coding model with credible benchmarks. No H100s. No H200s. No Blackwell. No NVIDIA anywhere in the training run.
That’s the chart. The model is the data point underneath it.
What was trained, on what
- Hardware: ~100,000 Huawei Ascend 910B accelerators. Domestic Chinese silicon, fabbed in part on SMIC processes. Not the leading edge of what TSMC ships, but workable.
- Software: Huawei’s MindSpore framework + CANN compute stack. Not PyTorch + CUDA. The entire training pipeline runs on a non-NVIDIA software stack.
- Output: GLM-4.7, an open-weights coding model with credible-to-strong scores on the standard coding benchmarks and aggressive pricing ($0.11/M input is the number I’ve seen quoted in Chinese coverage; haven’t fully verified against the official price card).
The Ascend 910B is not a Blackwell. On raw FLOPs per chip, it’s well behind. The training therefore had to span more chips, with more interconnect overhead, on a less-mature software stack. That Zhipu got a coherent frontier-adjacent model out the other end is a real engineering result independent of how the model itself stacks up to Claude or GPT-5.5.
Why this matters more than the benchmark numbers
The export-control regime that began in 2022 and tightened through 2023–2025 was built on a single bet: that the bleeding-edge AI capability frontier requires NVIDIA’s leading-edge silicon plus CUDA, and that denying both to China would meaningfully delay Chinese frontier capability.
GLM-4.7 is the first widely-discussed data point that the bet may have been timed correctly and outcome-wrong. The frontier got delayed in China — DeepSeek V4 Pro and Kimi K2.6 are six-to-twelve months behind the US frontier on the standard evals — but the training stack that produced GLM-4.7 is, from the export-control perspective, the worst news of the year.
Three implications:
1. The Ascend curve is steeper than the export-control architects assumed. Each generation of domestic silicon shrinks the per-chip gap a little. The compounding effect over three years is non-trivial. If 910B gets you GLM-4.7, 910C (rumoured 2027) likely gets you something at parity with mid-2025 frontier US models. The “we’ll always be a generation ahead” thesis needs revisiting.
2. The software stack lock-in has weakened. CUDA was supposed to be the moat that survived even if the hardware moat slipped. MindSpore + CANN training a real frontier-adjacent model at this scale is the first widely-public evidence that the CUDA moat is a moat made of policy and inertia, not of irreproducible engineering. Inertia gets eroded.
3. The open-weights dynamic compounds the problem. GLM-4.7 ships open. Anyone can download the weights. The training stack is irrelevant to a downstream user — what matters is the model. So even if Ascend stays a generation behind, the outputs of an Ascend-trained run circulate globally on the same terms as Llama or DeepSeek. The export controls were aimed at training-time capability. The deployment-time landscape they end up shaping is much weaker than intended.
The lab-by-lab consequence
I keep coming back to one structural read:
US labs (Anthropic, OpenAI, Google) win the frontier evaluations and the enterprise contracts. The capital and the compute concentrate there. The valuation numbers from this week — Anthropic at $900B, OpenAI at $852B, both with an order of magnitude more revenue than the Chinese labs — confirm that.
Chinese labs (Zhipu, DeepSeek, Moonshot, Qwen) win the open-weights diffusion layer and the developing-world adoption tier. They lose the absolute frontier and win the volume floor. GLM-4.7 on Ascend is the first model that says the volume floor is now untethered from the US compute stack.
Those are not the same game. They don’t compete head-to-head. They compete for the default — what does a developer in Lagos, or São Paulo, or Jakarta, or Bangalore reach for when they need a model. That default is, increasingly, going to be Chinese open weights. And the export controls don’t bite there at all.
What I’m watching next
GLM-5 is the obvious one — Zhipu has been telegraphing a 700B+ MoE successor with a much lower hallucination rate, also trained on Ascend. If that ships and benches close to GPT-5.5, the “trained without NVIDIA” line stops being a curiosity and starts being the default expectation for the Chinese frontier.
The next data point worth waiting for is whether DeepSeek follows Zhipu onto Ascend for V5. DeepSeek has been the most NVIDIA-dependent of the Chinese labs. If even they move, the export controls are, functionally, done.
May’s biggest story might not have been an American one at all.