The Race America Already Lost: In AI Video, China No Longer Competes With Silicon Valley — Only With Itself

OpenAI's Sora is dead. Runway and Google trail. In generative video, the only real contest left is between two Chinese giants — and their split over open versus closed reveals the whole Chinese playbook in a single field

A Front Where the War Is Already Over

Across fifty-four briefings, watchchina.ai has tracked a race that remains genuinely contested in most domains — chips, frontier models, consumer adoption. But there is one field where the contest is, for practical purposes, finished, and almost no one in the West has noticed: generative AI video. Here, China does not lead. China owns the field.

The evidence is stark. OpenAI's Sora 2 was shut down on April 26, 2026. The remaining Western contenders — Runway Gen-4.5 and Google's Veo 3.1 — topped out at roughly 8 to 15 seconds per native clip, requiring visible, seam-riddled stitching for anything longer. Meanwhile ByteDance's Seedance 2.5 generates 30-second clips natively, in 4K, with synchronized audio, holding character appearance and lighting consistent across the full duration — a capability no Western model matches. And the scale is already staggering: Seedance 2.0 alone reached roughly $2 billion in annualized revenue in about four and a half months, with users generating an estimated 3.3 million hours of video — more than three times every feature film ever made. As one prominent analyst put it plainly, Seedance 2.5 "cements China's absolute dominance in video." This is not a gap closing. This is a race that has already been run.

The Tell: China Is Now Fighting Itself

The clearest proof that China has won a field is when the only serious competition left is between Chinese firms — and that is exactly what happened on July 31, in a moment that watchchina.ai considers a perfect miniature of the entire Chinese AI strategy. On the same day, two Chinese giants launched competing video models with diametrically opposite philosophies. ByteDance rolled out Seedance 2.5 as a closed, API-only product, keeping its weights locked inside its own servers. Hours later, MiniMax released H3 — a competing model with open weights that anyone can download, run locally, and customize. One keeps everything inside a closed API; the other ships the model to developers to run on their own machines.

This split is the whole watchchina.ai thesis rendered in a single field. It is the two-tier strategy this portal forecast in July — open "good enough" models for diffusion, closed premium models for monetization — now playing out not as US-versus-China, but as China-versus-China. ByteDance takes the closed, high-margin path (Seedance 2.0 reportedly runs 70-90% gross margins), monetizing through its Doubao and Jimeng consumer platforms and its vertically integrated Douyin advertising engine. MiniMax takes the open path, seeding the global developer ecosystem exactly as DeepSeek and Qwen did in text. Between them, they cover every strategic base — premium and free, closed and open, consumer and developer — leaving no oxygen for a Western entrant to breathe. China isn't just winning AI video. It is running both sides of the winning strategy simultaneously, the way a chess player who owns the board plays out both colors.

Why This Field Matters More Than It Looks — and the Honest Caveats

watchchina.ai will not overstate the strategic weight of video generation relative to chips or frontier reasoning — but it will not understate it either, because generative video is quietly one of the most consequential AI domains for the actual future of human information. Video is how most of humanity now consumes information, forms beliefs, and experiences culture. Whoever controls the tools that manufacture synthetic video controls a substantial share of the future's persuasion infrastructure — a concern that connects directly to the deepfake and narrative-warfare threats tracked by watchchina.ai's sister portal warrier.ai. When the dominant tools for generating realistic video are Chinese, closed where it matters and open where it spreads, and subject — as both Seedance and MiniMax are — to China's National Intelligence Law, the implications extend well beyond who wins a benchmark.

The honest caveats remain, as always. Western dominance in video was never as entrenched as in text, so the loss is less dramatic than it sounds; ByteDance's models carry unresolved copyright exposure, having drawn cease-and-desist letters from every major Hollywood studio, which genuinely constrains Western enterprise adoption; and US availability of these tools remains limited by exactly those legal disputes. China's video dominance is real but legally encumbered, and those encumbrances are a genuine brake. But the structural reality watchchina.ai has documented all summer holds here in its purest form: in a whole category of frontier AI, the competitive question is no longer whether America can catch China. It is which Chinese company will win — and China has arranged its ecosystem so that whichever one does, China wins. That is what total dominance of a field actually looks like. And it happened in generative video while the West was watching the chips.

Sources: TechTimes, Japan Times, Bloomberg, YUAN Talks, tech360tv, 36Kr