Why Chinese AI Is Suddenly So Good (ft. DeepSeek, Seedance 2.0)

My notes

Summary

Chinese AI caught up to US labs not by matching hardware (they’re locked out of TSMC’s top-tier Nvidia GPUs) but by rewriting the software layer, extreme Mixture-of-Experts fragmentation, memory compression, and hand-tuned low-level GPU code. On the data side, ByteDance’s ownership of Douyin gives Chinese multimodal models (Seedance 2.0, Doubao) a structural advantage in uncompressed, metadata-rich video that US firms can’t scrape legally.

Key Insight

• AI stack has 3 layers: hardware (GPUs, TSMC/Nvidia bottleneck), model (Transformer-based LLMs), data (multimodal training fuel). China was cut off at layer 1, innovated at layers 2-3.
• Nvidia Blackwell B200 GPU packs 208 billion transistors, costs $30k-$40k, and you need tens of thousands wired together to train a frontier model. Nvidia = $4.5T market cap as of March 2026.
• TSMC controls ~70% of global chip manufacturing and >90% of advanced AI chips. Taiwanese firm but depends on US software/patents, US export rules block them from selling advanced chips to China.
• DeepSeek’s two breakthroughs at model layer:
  • Extreme MoE: 256 tiny experts, only 8 activated per query (vs OpenAI’s dozens of large experts). Rest of the brain stays asleep.
  • Multi-head Latent Attention (MLA): compresses short-term memory (KV cache) by >90%, dramatically cutting GPU memory cost.
• DeepSeek bypassed Nvidia’s CUDA abstraction by writing PTX (Parallel Thread Execution) assembly-level code, “manual transmission” vs CUDA’s “automatic.” Squeezed more performance from older, pre-ban GPUs.
• Result: DeepSeek trained a world-class model for ~$6M vs OpenAI’s hundreds of millions. Then open-sourced it, turning the model into a platform.
• Pure reinforcement learning failure mode: when DeepSeek rewarded only correct math answers, the model developed “language mixing”, a bizarre English/Chinese hybrid. Had to layer in small amounts of human-labeled data to restore readability.
• US multimodal data wall: American labs have already scraped what’s legally scrapable. Synthetic data and lawsuits are now the ceiling.
• ByteDance’s data moat: owns Douyin (750M DAU), uncompressed native video plus engagement metadata (exact swipe-away timestamp, lighting, angle). Seedance 2.0’s physics consistency (water splashes, reflections, audio sync) beats Sora because of this data structure, not just model design.
• Doubao (ByteDance) surpassed DeepSeek as China’s #1 AI chatbot because it’s multimodal end-to-end (images, video, voice) while DeepSeek is text-only.
• China’s blind spot: Douyin data is Chinese-culture-biased. Their models will hit the same data wall when trying to generalize to Western physics/cityscapes.