Why ByteDance leads in AI video despite compute disadvantage — and why it won't open-source CDance

ByteDance's Video AI Lead: Compute Efficiency and Strategic Focus Over Open-Source Play

ByteDance is winning in generative video despite operating with roughly one-fifth the compute of American rivals. The gap exists not because of IP flexibility in China—that alone doesn't overcome finite compute constraints—but because ByteDance has a more direct business flywheel than Anthropic or OpenAI for video products.

OpenAI launched Sora to some adoption, but video generation wasn't an obvious extension of its core business. ByteDance, by contrast, owns TikTok, Douyin, CapCut, advertising, ecommerce, and recommendation systems. Video AI feeds directly into these existing surfaces without requiring a new demand channel. Meta and Google have similar advantages in America, but both have deprioritized video generation—partly because AI-generated video faces cultural resistance in Western markets and partly because their leadership has moved toward agentic workflows, personal agents, and coding models instead.

ByteDance has made a different bet: it's leaning into video dominance rather than spreading compute across coding and AGI research. That focus compounds its advantage.

Training on Hollywood IP without legal exposure

The studio liability question is more subtle than pure "China doesn't care about IP." ByteDance can legally train on movie clips that appear on TikTok because much of that content qualifies as fair use—clips uploaded as commentary, transformation, or criticism don't need studio permission. Long-form film criticism channels post substantial movie excerpts routinely without revenue sharing, and it passes the fair use test because the content is transformational and non-competitive with the original work.

Clips live on ByteDance servers regardless of whether they're officially "deleted" from the platform. Over time, this builds a training corpus that includes Hollywood material without requiring explicit licensing or creating a legal paper trail of infringement.

The output risk is separate. Even if training on movie data is defensible, recreating a specific actor's likeness can still trigger a right-of-publicity claim. Google's music generation model handles this by refusing to generate exact replicas—it will produce a generic pop song rather than a Madonna track. ByteDance appears to take no such guardrails, generating recognizable likenesses. This suggests either a willingness to tolerate litigation or a bet that enforcement from Hollywood will focus on obvious replicas rather than subtler outputs.

Why no open-source despite China's playbook

ByteDance breaks from the standard Chinese AI strategy by refusing to open-source its video models, datasets, or weights. This contradicts the pattern China established with coding models, where open-sourcing became a tool to build an indigenous tech stack independent of American dominance.

Two dynamics explain the difference. First, no company open-sources when it leads—this is universal, not geopolitical. Second, open-source coding models serve a strategic purpose beyond market share: they reduce developer dependency on American tools and strengthen local supply chains. Video models offer no such ecosystem play. ByteDance is vertically integrated from model training to distribution, so closed weights don't cost it anything and preserve its moat.

The longer-term exception: if video models become critical for sim-to-real training in robotics, ByteDance might face pressure to open-source as part of a broader humanoid-robotics push. But that's speculative and farther out. For now, ByteDance's calculus is straightforward—it owns the distribution surface, so it closes the model.

The strategic logic

Challengers in fragmented markets open-source to build ecosystems. Incumbents with distribution lock-in stay closed. ByteDance has the latter advantage in video, which is why it's keeping its model proprietary even as it dominates the space with inferior compute.