Post

Chinese models are near-parity on benchmarks, dramatically cheaper per task, and topped US models in OpenRouter token share for the first time in 2026, all through legitimate contributions across quantization, attention architectures, MoE sparsity, and RL. In my newest report for @Delphi_Digital, I walk through the labs' progress across three efficiency themes and discuss: - why efficiency gains get competed away as price cuts - the evolution of licensing and business models over the past year - why China's grid is a structural advantage - the distillation fight, escalating sanctions threats, and scenarios for how the US-China AI market splits - whether Chinese labs can pre-train frontier models on domestic silicon - implications for investors Full report below.