orchery

BREAKING: The US Approves Nvidia H200… China Doesn’t Care The timing is almost embarrassing. As Washington finally approves Nvidia’s H200 sales to China, Zhipu AI drops a serious multimodal model trained entirely on a 100% Chinese stack. No Nvidia. No US chips. No TensorFlow or PyTorch. GLM-Image was built end-to-end on Huawei Ascend hardware using Huawei’s MindSpore framework. Full pipeline. Domestic. Independent. This matters because US sanctions were meant to slow China and lock in dependency. Instead, they’ve done the opposite. Despite years of export controls and pressure, China has now publicly proven that high-quality, open-source multimodal AI no longer needs US silicon. And this isn’t a toy. GLM-Image delivers industry-leading performance among open-source models for text rendering and excels at Chinese character generation — something Western, English-first models still struggle with. It may trail ByteDance’s proprietary models, but that’s beside the point. This is about who controls the stack, not who tops a benchmark. So when the US finally says, “We’ll allow limited H200 exports,” Beijing’s response is effectively: no thanks. Imports reportedly blocked. Approvals “exceptional.” Translation: China is done building its future around American permissions. This isn’t anti-Nvidia. It’s post-Nvidia. Sanctions forced a full-stack rebuild — chips, frameworks, tooling, talent — and once domestic systems hit “good enough,” dependency breaks for good. Approving the H200 now doesn’t look like control. It looks like turning up late to a door that’s already closing. The question isn’t whether China can do advanced AI without US chips — that’s been answered. The real question is how fast the rest of the ecosystem follows.