Latest Assessment by US VCs After China Visit: AI Capabilities Need to Cross the Pacific Twice to Be Sold to US Clients

Beating AI Insight News Brief: U.S. venture capital firm Dimension spent a week in China, visiting AI labs, investors, and founders, then wrote a 5-page internal letter to its limited partners (LPs). The firm previously conducted research in Shanghai last year; this trip covered Beijing and Shanghai, with the goal of recalibrating its assessment of China’s AI sector. A key finding highlighted is that China’s open-weight models have entered the production pipelines of U.S. AI companies. For example, Cursor’s Composer 2 was further trained on Kimi K2.5, while legal AI firm Harvey’s Tenet model underwent post-training on Kimi K3. Dimension summarized this cross-Pacific value chain as a "two-way trans-Pacific flow": U.S. labs first train cutting-edge models, Chinese labs absorb these capabilities and release open-weight models, then U.S. application companies further train on the Chinese models to build products sold to U.S. enterprises. However, despite surging usage, revenue has not accrued proportionally to Chinese model firms. Dimension summed this up in one sentence: "China is getting Western workloads, not Western revenue." Open-weight models can be deployed across multiple platforms, with U.S. inference firms like Fireworks, Baseten, and Modal capturing the bulk of service fees. Chinese models drive down costs, ultimately benefiting U.S. cloud providers, inference platforms, and AI application companies as well. Dimension’s final assessment is that it is no longer feasible to simply split China and U.S. AI into two separate systems. Hardware like chips is being decoupled: the U.S. faces constraints from power and grid infrastructure, while China lacks advanced chips. Yet models, data, inference services, and software frameworks continue to flow across borders. Debating "which side is winning the China-U.S. AI race" in simplistic terms no longer reflects the real industrial chain.