US export controls gate Qwen's frontier compute: policy toggles open or close the gate on Nvidia GPU supply into frontier training and cloud inference The regime whipsaws from an April 2025 H20 halt to a July 2025 reversal and a December 2025 H200 approval. Permitted Nvidia imports pass through the export-control gate into frontier training and Alibaba Cloud inference. Domestic chips trail by about 5x today, widening toward 17x by H2 2027. Tightening or a stalled whipsaw drives both lanes into a compute wall; loosening yields headroom. Apr '25 H20 halt Jul '25 reversal Dec '25 H200 OK Nvidia imports H20 · H200 imports US side China domestic ceiling ≈5x performance gap → ≈17x by H2 '27 Domestic chips Huawei · T-Head US export control whipsaw Frontier training Qwen cadence Cloud inference Alibaba scale Compute wall stalls Headroom training + inference scale Qwen downloads keep rising compute wall caps the flywheel
Whipsaw regime: supply is uncertain and the ≈5x Huawei/T-Head gap holds, so Qwen frontier training and Alibaba Cloud inference sit against the compute wall even as downloads keep rising.