Why Chinese AI Is Suddenly So Good (ft. DeepSeek, SeeDance 2.0) | AB Explained

Trigger event and policy shift

• AlphaGo’s 2016 victory over Lee Sedol — notably “move 37” — is framed as the catalyst that accelerated China’s national AI push.
• Result: China’s Next Generation AI Development Plan set a national target to be the world leader in AI by 2030.

“Move 37” (AlphaGo vs. Lee Sedol, 2016) is often cited as the inflection point that spurred China’s accelerated, state-backed AI strategy.

AI as a stack

The analysis frames AI as a layered stack: hardware, model (software), and data/application. Each layer shapes how capability and competitive advantage are developed.

Hardware layer: chips, datacenters, cooling, interconnects

• GPUs (especially Nvidia designs) are the engine of modern AI training and inference.
  • Notable Nvidia architectures mentioned: Blackwell B200 (208B switches, two-chip package) and the upcoming Rubin platform (expected Jan 2026).
  • A single top-tier GPU costs roughly $30k–$40k; training large models requires tens of thousands of GPUs.
• Manufacturing concentration and export controls:
  • Nvidia designs chips; TSMC (Taiwan) is the dominant foundry (~70% overall, >90% at most advanced nodes).
  • U.S. export controls and dependence on Western manufacturing equipment create a hardware bottleneck for China.
  • Building advanced fabs is extremely complex and costly; even large firms (e.g., Samsung) face struggles.
  • Example response: Elon Musk’s Project Terafab as an attempt to mitigate supply limits.

Model (software) layer: architectures and engineering choices

• Foundation models are largely based on the Transformer architecture (2017) and scale via next-token prediction.
• DeepSeek (Chinese open-source foundation model) highlights software and engineering innovations that reduce hardware dependence:
  • Extreme Mixture-of-Experts (MoE): divides the model into 256 tiny experts; a router activates ~8 experts per request to save compute.
  • Multi-head Latent Attention (MLA): memory-compression technique that shrinks the key-value cache / short-term memory footprint by >90%, drastically reducing GPU memory needs.
  • Low-level hardware optimizations: custom PTX operators under CUDA to squeeze performance from older Nvidia GPUs rather than relying on high-level defaults.
  • Cost efficiency: reported training cost < $6M vs. hundreds of millions for comparable Western models.
  • Open-source distribution: enables inspection, fine-tuning, and rapid parallel innovation by many developers.
• Training strategy:
  • Heavy use of reinforcement learning to cheaply build reasoning capabilities.
  • Followed by some targeted human-labeled data to correct communication issues (e.g., language mixing, formatting).

Data / application (consumer) layer: China’s structural advantage

• Multimodal, high-quality, native consumer data (video, audio, images + rich engagement metadata) is critical for next-gen models that must understand the physical world.
• Western limits:
  • Many Western firms have largely scraped available public multimodal data and are hitting constraints (quality, legal/copyright/privacy), increasing reliance on synthetic data.
• China’s advantage:
  • The super-app ecosystem (WeChat, Douyin, ByteDance platforms) continuously generates massive volumes of native, often uncompressed, well-labeled multimodal content with rich engagement metadata (camera angle, lighting, watch-thresholds, etc.).
  • This vertically integrated data pipeline is a unique training asset for multimodal and grounded models.
• ByteDance product highlights:
  • Seedance / SeeDance 2.0: multimodal video-generation model focused on natural motion synthesis, physical consistency, and audio-visual sync — benefits from native Douyin data.
  • Doubao: ByteDance’s multimodal AI chatbot that leverages ByteDance’s video/audio data pipeline and reportedly achieved broad user adoption.
• Consequence:
  • DeepSeek shows China can compete on reasoning and model efficiency; ByteDance demonstrates advantage in multimodal, consumer-facing AI products.

Geopolitics and strategic implications

• The U.S. retains control over critical hardware supply via sanctions and export controls, but China has partly compensated through software innovations and an integrated consumer-data ecosystem.
• Limitations for China:
  • Restricted access to the latest chips and some international data.
  • Models trained primarily on Chinese-sourced multimodal content may underperform on non-Chinese contexts without broader data access.
• The AI race is shifting:
  • From being purely compute-driven to including breakthroughs in model architecture, efficiency engineering, and access to unique data sources.
• Future frontier:
  • Real-world, non-internet data (e.g., structured interviews, physical sensors) could become increasingly important for further progress.

Product / service mentions (features & value)

• DeepSeek
  • Open-source, highly efficient foundation model.
  • Key features: extreme MoE, MLA compression, PTX-level CUDA optimizations.
  • Value: cheap to train, strong at reasoning; initially required RL and targeted fixes for natural language output.
• Seedance (SeeDance) 2.0
  • ByteDance multimodal video generator.
  • Strengths: physical realism, audio-visual alignment, natural motion synthesis; benefits from Douyin data and metadata.
• Doubao
  • ByteDance chatbot leveraging multimodal data and broad consumer adoption in China.
• Incogni (sponsor)
  • Consumer service that contacts data brokers for removals, offers custom removals and deletion verification; promo code mentioned in the source video.

Practical takeaways / conclusions

• China’s rapid AI progress is driven less by hardware parity and more by:
  • Software ingenuity (efficiency engineering such as MoE and MLA).
  • Unparalleled access to structured, native multimodal consumer data through integrated platforms.
• Engineering optimizations (Mixture-of-Experts, memory compression, low-level GPU optimizations) can dramatically reduce training costs and enable high-quality models on older hardware.
• Ownership of data and integrated app ecosystems (super-apps + content platforms) are strategic assets for multimodal AI development.
• The ongoing race will span hardware, model innovation, and unique datasets; future gains may depend on access to non-internet, real-world data and improved multimodal grounding.

Main speakers / sources referenced

• Stephen Park (narrator, Asian Boss)
• Companies / people: DeepMind (Demis Hassabis), AlphaGo, Lee Sedol, Google (Sergey Brin), Nvidia (Jensen Huang), TSMC, OpenAI, Anthropic, ByteDance, DeepSeek, Seedance / SeeDance 2.0, Doubao, Douyin, SMIC, Huawei, Elon Musk (Project Terafab)
• Key technologies & concepts: Transformer (2017), LLMs, Mixture-of-Experts (MoE), Multi-head Latent Attention (MLA), CUDA / PTX, reinforcement learning, multimodal data, natural motion synthesis.