A 20,000-word analysis of Chinese AI as of March 2026. Eight labs. Ten frameworks. Twelve steelmanned perspectives. The Stanford AI Index shows the benchmark gap between American and Chinese frontier models compressed from 9.26% to 1.70% in twelve months. On MMLU, the gap is 0.3 points. On ARC-AGI-2, it is 30-70 points. Both measurements are real. What they mean depends on what intelligence is. DeepSeek's V3.2-Speciale achieves gold-medal-equivalent scores on historical IMO, IOI, ICPC, and CMO problem sets from a team of approximately 200 researchers. Zhipu's GLM-5 trains on 100,000 Huawei Ascend chips — zero Nvidia. Anthropic accuses three Chinese labs of industrial-scale distillation. The law has no answer. 61% of OpenRouter token volume flows through Chinese models. 63% of Hugging Face fine-tunes build on Chinese architectures. The U.S. holds 75% of global AI compute. China holds 15%. Three trajectories: convergence, sustained asymmetry, fragmentation. The episode does not resolve which one is correct.

This episode was produced using the Proxima.Earth methodology — an open-source, multi-model AI pipeline for geopolitical synthesis. No human is in the loop after subject selection. The methodology is the editorial control. Full methodology, prompts, and production transparency: proxima.earth/methodology

Corrections, source disputes, or methodology feedback: [email protected]