Bloomberg: Chinese labs hold 9 of top 10 text-to-video slots

The most eye-catching line in Catherine Thorbecke's latest Bloomberg Opinion column is a ranking, not a forecast: outside of Alphabet's Google, nine of the top ten text-to-video systems on Artificial Analysis' leaderboard are Chinese. Kuaishou's Kling, ByteDance's Seedance, MiniMax, and Vidu from Beijing-based Shengshu Technologies are all trading places at the top of a public benchmark that Western labs used to dominate.

The timing matters. MiniMax shipped its H3 model on July 31, releasing open weights that generate up to 15-second clips at 2K with native stereo sound, priced at under a third of mainstream rival products and aimed at advertising, e-commerce, product design and games. ByteDance's rival Seedance 2.5 launched the same day but stayed inside a closed API. Two dueling philosophies, both fast, both cheap, both Chinese.

Thorbecke's argument is that this is not really a Hollywood story, even though the copyright disputes and cinematic use cases get the attention. Multimodal video systems that hold characters consistent across cuts, follow direction and pair scenes with native audio start to look like early world model engines, the same substrate that would eventually feed robots, autonomous vehicles and simulators. If that framing holds, a leaderboard lead in video is a proxy for a lead in physical AI training data.

The piece is opinion, and it does not spell out which benchmark axes on Artificial Analysis put the Chinese systems on top, so how the ordering holds under different prompt sets is left open. Nor does the column pin numbers on revenue or subscribers behind the leaderboard leaders, so top-of-leaderboard and top-of-market are not the same claim.

Still, the stakes Thorbecke sketches are the ones a Western strategy team should be pricing now, not next year. Our tracker has logged hundreds of China AI stories in the last ninety days, and the recent drumbeat around ByteDance, including dueling video launches inside the same week and a reported 10-trillion-parameter pretraining run, suggests the video lead is a leading edge, not a niche. Studios can argue about Hollywood; robotics buyers should be asking who ends up training their simulation stack.