AI with Chinese Characteristics — Decouple Podcast Raw Notes¶

Date: 2026-05-13 Guest: Kyle Chan (High Capacity Substack, Brookings Institute, previously Princeton/RAND) Topics: nuclear-industry, ai-automation, geopolitics-energy

Deep Seek Shock (January 2025)¶

Deep Seek R1 caused Nvidia to lose $500B market cap
Achieved competitive performance with orders of magnitude less compute
Surprising even within China; became a national pride moment
Deep Seek is an outlier: originated from High Flyer hedge fund that bet early on compute
Showed China can perform at world-class level despite US export controls

Chip Race & Export Controls¶

US paradigm: scaling laws — more compute = better models; export controls aim to limit China's compute access
Chinese startups: $6-8B valuations vs OpenAI $500B+; capital-constrained
Chinese response: efficiency maximization — quantization, memory architecture optimization
Irony: Trump wants to ease controls (allow H200 exports), but Beijing may not want them — prioritizing domestic chip industry
China building entire semiconductor supply chain: HBM memory, lithography, photo resists, metrology tools

"AI with Chinese Characteristics" — Different Application Philosophy¶

US: Building "digital god" (AGI/ASI) that will either save or destroy humanity
China: Practical applications — smart cities, smart transport, smart grids, energy logistics
Focus on today: "Can the train system run more efficiently?" not future digital deity
This approach predates the ChatGPT moment

State Capacity & Bureaucratic Efficiency¶

Foundation: performance-based legitimacy
Officials promoted by KPIs: economic growth, local AI development, investment attraction
AI integrated into government services (license renewal, permits)
US contrast: hollowed-out state capacity, AI used more for policing/immigration (Palantir)

AI Regulation & Censorship¶

China unique: all large foundation models must register with Cyberspace Administration
Big tech "disciplining": Jack Ma era taught companies not to challenge the Party
Self-censorship layer: models know politically sensitive answers but are blocked before responding
Tiananmen, Taiwan — immediate API-level blocking

"Opinionated" governance: if state deems something harmful, it gets shut down
Examples: tutoring industry dismantled, youth gaming time limits, HFT crackdown, crypto ban
US contrast: platforms optimize for advertisers, neglecting mental health

Soviet Parallel & Chinese Model¶

Soviet computing paradox: computers couldn't save central planning, only exposed inefficiency
China's difference: incentive-based system, not command economy
Competition even within state-owned enterprises
AI used for corruption detection, digital data analysis (WeChat, mobile payments, train data)

Energy & AI Infrastructure¶

Decades of massive energy buildout: ~6% annual energy consumption growth
2025: 500+ GW new capacity (80% wind/solar) — ~40% of entire US grid
Bloomberg NEF: China to double power capacity in 5 years
AI data centers deliberately sited in Western China (cheap renewables, regional development)
"Western Data Eastern Compute" project: train in west (compute-heavy), infer in east (low latency)
Issues: some data centers underutilized, built by state telcos without expertise

Humanoid Robots¶

Unitree: 15,000+ humanoids produced in 2025, IPO expected in 2026
Full supply chain advantage: EVs, batteries, servo motors, LIDAR, sensors
"Electric tech stack": new technology cluster centered around electricity
Demographics make automation less threatening — shrinking workforce needs automation
Hype cycle: unclear if technology is truly ready for mass deployment

Cultural Differences in AI Perception¶

Western sci-fi: apocalyptic dystopias, technology takeover
Chinese fear: falling behind — individually ("will I be left behind as a worker?") and nationally
US: "technology moving too fast" vs China: "what if we can't keep up?"
Robotaxi protests exist but automation seen as partial solution to demographic decline

Key takeaway: China's AI development is not simply copying the US model with fewer chips — it involves a fundamentally different application philosophy, distinct political-economic incentives, and deliberate state-building of the entire stack (research, chip manufacturing, energy, human capital, regulation).