AI Superpowers: China, Silicon Valley, and the New World Order
Former Google China head Kai-Fu Lee breaks down why AI implementation — not research — decides who wins the US-China AI race, and what that means for the 800 million jobs in the crosshairs.
Top Claims — Verdict Check
China will match or surpass the US in AI not through research breakthroughs but through relentless implementation at scale
🟢 Real“The age of AI discovery is over. The age of AI implementation has begun — and China has structural advantages in speed of deployment, data volume, and government alignment. [representative paraphrase]”
AI will displace 40% of jobs worldwide within 15 years
🟡 Partially True“Routine cognitive and physical jobs — truck drivers, radiologists, loan officers — will be displaced within 15 years on a scale that dwarfs the Industrial Revolution. [representative paraphrase]”
The US leads in AI research, but China leads in AI application and data
🟢 Real“Silicon Valley has the best researchers and papers. China has the biggest datasets, the fastest product cycles, and a government that doesn't slow you down. [representative paraphrase]”
Love and compassion are uniquely human and should define the jobs AI cannot replace
🟡 Partially True“AI cannot love. AI cannot care. The jobs that require genuine human connection — caregiving, teaching, social work — are where humans should redirect. [representative paraphrase]”
A new world order will emerge with the US and China as dual AI superpowers, leaving everyone else behind
🔴 Hype“The AI world will be bipolar — the US and China will dominate, and every other country will need to choose which ecosystem to plug into. [representative paraphrase]”
What's Real
The implementation thesis aged extremely well. In 2018, this was a contrarian framing — most Western observers were focused on who had the best papers at NeurIPS. By 2024, China's AI deployment numbers speak for themselves: WeChat's billion-user AI integration, Alibaba's logistics AI handling Singles' Day volumes that dwarf Black Friday, BYD's autonomous driving stack built on a domestic data moat of 100+ million connected vehicles. China's advantage in data volume is structural — 1.4 billion people generating mobile-first data in a regulatory environment that until recently imposed few collection constraints. The government coordination claim is documented: China's 2017 New Generation AI Development Plan set explicit national targets and directed billions in subsidies to AI companies. Baidu, SenseTime, and iFlytek all received designated 'national champion' status with corresponding state support. Lee's call that implementation speed would matter more than research novelty predicted the exact dynamic playing out between GPT-4 and DeepSeek — the latter's R1 model matched frontier performance at a fraction of the cost through engineering execution, not novel architecture.
What's Hype
The '40% of jobs displaced in 15 years' prediction, made in 2018, has not materialized at the stated pace by 2026. Eight years in, automation has transformed specific workflows (customer service chatbots, document processing, basic code generation) but has not eliminated entire job categories at anything close to 40%. The timeline conflates 'could theoretically automate' with 'will actually be deployed' — the gap between technical capability and organizational adoption is consistently 3-5x longer than technologists predict. The bipolar US-China framing also missed the open-source disruption: Meta's Llama releases, Mistral in France, and the broader proliferation of capable models means smaller nations and companies aren't locked into a binary choice. ASEAN nations like Singapore and Malaysia are building AI strategies that draw from both ecosystems — not picking sides. The 'love and compassion' framing, while emotionally resonant for a TED audience, sidesteps the economic reality: caregiving and teaching are chronically underfunded precisely because they're undervalued by markets, and redirecting displaced workers into these roles without restructuring their economics is not a solution.
What They Missed
The talk was given before the transformer revolution fully played out. GPT-2 was still months away. The entire paradigm of foundation models, few-shot learning, and the emergent capabilities of scale — the mechanism by which AI actually became dangerous to knowledge work — is absent. Lee's job displacement model focused on routine tasks (driving, factory work, data entry) but the actual displacement since 2023 has hit knowledge workers first: copywriters, junior coders, graphic designers, customer support. The ASEAN perspective is completely absent, which matters for NerdSmith's audience. Malaysia, Vietnam, Indonesia, and the Philippines sit in the implementation layer of the global AI supply chain — they're the ones doing data labeling, BPO, and software services that are directly in the automation crosshairs. Lee talks about US and China; the 600 million people in between get no playbook. The role of compute cost — not just data volume — as the bottleneck for AI deployment in developing economies is also missing.
The One Thing
AI leadership is won through implementation speed and data access, not research papers — and that insight applies at every scale, from nations down to your team.
So What?
- The implementation-over-research frame applies directly to your business: the company that deploys AI into workflows fastest wins, regardless of who published the paper
- ASEAN businesses sitting between the US and China AI ecosystems should build with both — use OpenAI for English-language tasks, explore DeepSeek and Qwen for cost-sensitive and multilingual workloads
- The job displacement timeline is slower than predicted but the direction is correct — audit which roles in your organization are doing routine cognitive work and build transition plans now, not when the headlines hit
Action Items
- 1Read China's 2017 New Generation AI Development Plan (English translation available on Stanford HAI) — it's a 30-minute read that explains the structural coordination Lee describes and gives you a framework for understanding state-directed AI strategy.
- 2Audit your own team's AI implementation speed: how many days from 'we should use AI for this' to 'it's in production'? If the answer is more than 30 days for a simple workflow, your deployment friction is your actual bottleneck — not model capability.
- 3Map your company's routine cognitive tasks (report generation, data entry, customer triage, scheduling) — these are the roles Lee's displacement thesis targets first. Build a 12-month transition plan for each one, even if you don't execute immediately.
Tools Mentioned
DeepSeek
Chinese AI lab whose R1 model demonstrated that implementation engineering can match frontier research at lower cost
Qwen
Alibaba's open-weights model family — strong multilingual performance relevant to ASEAN builders
Workflow Idea
Build a 'deployment velocity' tracker for your team. Every time someone identifies an AI use case, log it with three dates: identified, prototyped, deployed. After three months, calculate your average cycle time. Lee's thesis says implementation speed is the moat — this metric tells you exactly how fast your moat is growing. Most teams discover their bottleneck isn't model selection or prompt engineering; it's internal approval processes and integration friction. Fix the process, and the AI advantage follows.
Context & Connections
Agrees With
- Sam Altman on compute as a strategic resource
- Demis Hassabis on AI's potential to transform industries at scale
Contradicts
- Those who frame AI leadership as purely a research publication race
- The bipolar US-China-only framing — open-source models have created a third path
Further Reading
- AI Superpowers by Kai-Fu Lee (2018) — the full thesis in book form
- Stanford HAI AI Index Report 2024 — US vs China metrics on papers, patents, and deployment