An overview of Chinese tech companies' rush to match generative AI tools like DALL-E 2 despite tighter regulations, censorship, US chip sanctions, and more
Chinese tech companies rush to match Stable Diffusion and DALL-E 2, but roadblocks lie ahead — The gigantic technological leap …
Context & Ripple Effects
In January 2023, Chinese tech companies were racing to replicate DALL-E 2 and Stable Diffusion just as Beijing tightened generative-AI rules and Washington's chip sanctions cut off their easiest path to compute. The constraint showed quickly: a year later, insiders told the New York Times that China trailed the US in generative AI by at least twelve months, with Chinese teams leaning on US models like Meta's LLaMA to close the gap at least a year behind.
The subsequent arc is the interesting part. Rather than keep chasing US closed models, Alibaba, Baidu and DeepSeek pivoted to open-sourcing their models to route around US curbs and pull in global refinement; DeepSeek then became infrastructure for Chinese automakers and phone makers building services on top of it; and by 2026 ByteDance and Kuaishou were judged ahead of US rivals in video generation, trained on the short-form libraries only their own apps possess.
First-order effects
- Chinese labs including Alibaba, Baidu and ByteDance had to build image-generation products under simultaneous censorship compliance and restricted access to top-tier US training chips, capping how fast they could match DALL-E 2-class output.
- US chip sanctions made compute procurement, not model research, the binding constraint on every Chinese generative-AI launch in this wave.
Second-order effects
- Locked out of the best hardware, Chinese firms converted openness into strategy — releasing model weights openly to decentralize development and tap global talent, which seeded the domestic DeepSeek ecosystem that automakers and smartphone vendors now build on.
- Sanctions pushed the bottleneck downstream into domestic silicon, culminating in state-level pressure on AI companies to use local chips, with the Vice Premier reportedly framing resistance as disloyalty.
Third-order effects
- If the pattern holds, China ends up with a parallel AI stack — local chips, open-weight models, and app-native data advantages like ByteDance's and Kuaishou's short-video corpora — that competes category-by-category rather than head-on, already visible in its lead in video generation.
- Censorship-compatible, openly licensed models become China's export vehicle: the same properties that satisfy domestic regulators make the weights easy for foreign developers to adopt, turning a political constraint into distribution.
The trend: China's generative-AI playbook has shifted from imitating US closed models under sanction to an indigenized, open-source-weighted stack where data scale and state-aligned industrial policy substitute for restricted hardware.