Silicon Valley VC firm Dimension's August 2026 letter to LPs, based on a week-long tour of Chinese AI labs, argues that US export controls pushed Chinese labs into extreme code-level optimization (e.g. DeepSeek's V3 trained on 2,048 H800 chips with low-level PTX programming), that five Chinese labs are tightly clustered with rankings resetting quarterly, and that a 'Pacific Data Circulation' has emerged in which US frontier models are distilled into Chinese open-source weights that US vertical AI companies then fine-tune and sell. The letter also reports a near two-order-of-magnitude revenue gap between Chinese and US model companies alongside far higher Chinese valuation multiples. For a financial institution, this bears on how export controls affect the addressable market and competitive position of US AI vendors and on whether US app-layer revenue rests partly on Chinese open-source weights.
431 other tracked developments touching the same claims.