Blog · Review Essay · Last reviewed June 25, 2026

AI Superpowers and the Implementation State

Kai-Fu Lee's AI Superpowers is best read as a 2018 field report from the moment artificial intelligence stopped looking mainly like a research race and started looking like a deployment race. Its central question is not whether machines can think in a general human sense. It is what happens when companies and states learn to turn data, labor, consumer habits, capital, and policy into machine-learning advantage.

For this review, an implementation state is not a country that simply "uses AI." It is an institutional stack that converts models into operational authority: data capture, compute access, procurement, standards, labor discipline, service design, appeals, logs, and public narratives about national progress.

The Book

AI Superpowers: China, Silicon Valley, and the New World Order appeared in 2018 from Houghton Mifflin Harcourt. WorldCat records the original print book as an English-language 2018 Houghton Mifflin Harcourt title published in Boston, and HarperCollins' current page lists a September 14, 2021 paperback. The publisher presents the book as an argument about U.S.-China AI competition, job disruption, and human-AI coexistence.

Lee is unusually well positioned for that argument. Stanford's 2025 event biography lists him as Chairman and CEO of Sinovation Ventures, founder of 01.AI, former executive at Apple, Microsoft, and Google, and former President of Google China. The book is therefore not a neutral outside survey, but it is also not generic futurism. It is written by someone who saw AI research culture, Silicon Valley platform culture, and Chinese implementation culture from the inside.

The book's durable contribution is its refusal to treat AI power as a property of algorithms alone. Lee keeps pointing to the surrounding machinery: large markets, mobile payment habits, delivery platforms, cameras, entrepreneurs, engineers, policy support, venture capital, and the willingness of institutions to push machine learning into daily life.

Current Context

As of this June 25, 2026 review, the book's scoreboard has changed but its central move still matters. Stanford HAI's 2026 AI Index says the U.S.-China performance gap among leading models has effectively closed, with U.S. and Chinese models trading the lead since early 2025. It also says the United States still produces more notable models and high-impact patents, while China leads in publication volume, citations, patent grants, and industrial robot installations.

The same report makes clear that "AI superpower" is now a stack, not a trophy. The United States leads private AI investment, hosts 5,427 data centers, and produces many of the most visible frontier systems. China remains central in research output, patents, industrial robotics, and implementation capacity. AI sovereignty has become a national-policy theme, but the infrastructure underneath it is unevenly distributed across compute, energy, talent, capital, supply chains, and public institutions.

The implementation race is also more material than it looked in 2018. OECD's AI infrastructure work describes a supply chain that includes chips, data centers, cloud computing, power, cooling, and networks. The International Energy Agency projects global data-center electricity consumption roughly doubling to about 945 TWh by 2030 in its base case, with accelerated servers driven mainly by AI adoption. Lee's "implementation" frame now has to include the physical and regulatory capacity to power, cool, secure, allocate, and audit the systems being implemented.

The Age of Implementation

The core thesis is that deep learning had moved AI from an "age of discovery" toward an age of implementation. In Lee's frame, the decisive advantage no longer belonged only to the lab that invented a new technique. It belonged to the society that could apply existing techniques across many domains, collect feedback, and improve quickly.

This is why the book belongs beside Prediction Machines, Platform Capitalism, and The Rise of the Network Society. AI appears as an institutional metabolism. Platforms generate data. Data improves models. Better models increase convenience or control. The improved service attracts more use, and the loop returns as evidence that the system knows the world.

The governance problem is hidden inside that metabolism. Implementation is not only adoption speed. It is the ability to define the workflow, collect the data, route the exception, buy the compute, approve the vendor, train the worker, measure the outcome, and decide whether affected people can refuse, correct, or appeal the system. A society can be good at implementation while being bad at accountability.

Fortune's 2018 review captured this shift by emphasizing Lee's claim that China had advantages in market scale, data, mobile behavior, and practical entrepreneurship. Time later noted that Lee himself revised the balance in 2020, saying the United States had begun catching up in everyday AI adoption. That revision matters. The strongest part of the book is not a fixed scoreboard between two nations. It is the analytic move from invention to deployment.

The State as AI Platform

AI Superpowers is also a book about technological politics. Lee writes about companies, founders, engineers, and consumers, but the background is the state. China's 2017 New Generation Artificial Intelligence Development Plan set national goals through 2030 and described AI as a strategic technology that would reshape society, life, and global competition. Stanford DigiChina's translation makes clear that the plan was not just research policy. It was industrial policy, education policy, infrastructure policy, and institutional coordination.

That is the book's most useful warning for AI governance: there is no purely technical AI race. Models sit inside procurement systems, chip supply chains, cloud regions, labor markets, military planning, export controls, standards bodies, universities, and public narratives about national destiny.

Export controls make that visible. BIS's January 2026 revision moved certain applications for Nvidia H200, AMD MI325X, and similar chips exported to China or Macau to case-by-case review under specified security conditions. BIS's May 31, 2026 guidance then clarified continuing license requirements for advanced computing items destined for entities headquartered in Country Group D:5 or Macau, even when those entities are outside those destinations. GAO's May 2026 decision on the AI Diffusion Rule non-enforcement announcement is a reminder that AI policy is not only strategy; it is rulemaking, legal status, enforcement posture, and procedural accountability.

The result is not one clean winner. It is a distributed rivalry across compute, talent, papers, patents, robots, capital, energy, standards, law, and implementation capacity. Lee's title still works if "superpower" means control over the environment that makes AI useful, not possession of one superior model.

Labor After the Demo

Lee is blunt about job displacement. He argues that AI will affect both blue-collar and white-collar work, especially routine tasks that can be optimized, predicted, classified, routed, or automated. His preferred answer is not simple universal basic income. He turns instead toward human care, service, compassion, and work built around human-to-human relation.

That part of the book is morally attractive and politically incomplete. Care work is not outside the economy. It is gendered, racialized, underpaid, credentialed unevenly, and often managed by the same institutions that adopt automation. To say that humans should do more human work is only a beginning. The harder question is who pays for it, who controls it, and whether care becomes a refuge from automation or another interface for automation to manage.

The book becomes sharper when read against recent labor data. The 2026 AI Index reports that AI's labor effects are appearing unevenly, concentrated in hiring pipelines and younger workers in exposed occupations; it reports that employment for software developers ages 22 to 25 fell nearly 20% from 2024 and that one-third of organizations surveyed expected AI-related workforce reductions in the coming year. It also reports that China accounted for 54% of industrial robot installations globally in 2024. Lee's old warning about displacement has not been settled by the arrival of generative AI. It has become more concrete.

A labor-safe implementation state would therefore measure more than productivity. It would track entry-level hiring, deskilling, surveillance, wage effects, contractor displacement, training access, worker appeal, and whether people whose labor trains, labels, deploys, monitors, repairs, or absorbs AI systems receive bargaining power rather than rhetoric about inevitable progress.

Governance and Safety

Lee's implementation frame becomes useful when it produces controls. NIST's AI Risk Management Framework gives one portable vocabulary: govern, map, measure, and manage AI risk across the lifecycle. For an implementation state, those verbs should apply not only to model behavior but to the surrounding system: data source, compute dependency, procurement file, labor plan, security boundary, human-oversight role, appeal path, incident log, and shutdown authority.

The EU AI Act shows another kind of update to Lee's world. It treats high-risk AI systems, general-purpose AI models, transparency, technical documentation, logging, human oversight, accuracy, robustness, cybersecurity, and compute-scale signals as regulatory objects. That does not settle U.S.-China competition, and it does not prove any system is safe. It does show that implementation is no longer merely a business skill. It is a legally and institutionally reviewable practice.

A serious review of national AI strategy should keep four ledgers. The capacity ledger records compute, chips, data centers, energy, cooling, talent, and public-access capacity. The data ledger records provenance, consent, licensing, retention, representativeness, and deletion rights. The deployment ledger records use cases, model versions, vendors, prompts, tools, human review, appeals, and incidents. The labor ledger records who is displaced, monitored, deskilled, upskilled, newly hired, or made invisible by automation.

Those ledgers are safety infrastructure. Without them, "AI leadership" becomes a scoreboard that rewards speed while hiding extraction, concentration, brittle supply chains, local energy burdens, weak appeals, and institutional lock-in.

What Aged Unevenly

The book is pre-ChatGPT, pre-frontier LLM platform war, and pre-DeepSeek-R1. Its mental model is still strongly shaped by supervised learning, computer vision, speech, recommendation, finance, logistics, and offline-to-online consumer platforms. That makes some passages feel dated in 2026, especially where general-purpose language interfaces, coding agents, synthetic media, and model supply chains now dominate public attention.

Lee also underplays the degree to which "implementation" can become coercion. A society that moves quickly can discover useful applications. It can also normalize surveillance, weak consent, administrative opacity, and competitive pressure that leaves people no meaningful way to refuse. Speed is not only an innovation advantage. It is also a governance risk.

The two-country frame also needs pressure. U.S. and Chinese capacity matter enormously, but a duopoly story can hide the role of Taiwan's chip fabrication, Dutch lithography, Japanese materials, Korean memory, European regulation, Gulf and European sovereign-AI investment, Indian and Southeast Asian service markets, open-weight communities, and public-sector procurement. AI power is concentrated, but it is not contained by one bilateral rivalry.

Still, the dated parts are useful. They show what the AI conversation looked like just before large language models reorganized it. The book's blind spots help mark the transition from AI as pattern-recognition infrastructure to AI as general interface, agent, writing machine, companion, coder, tutor, bureaucratic assistant, and geopolitical symbol.

What This Changes

AI Superpowers is a book about recursive reality at national scale.

The loop is not mysterious. A state or platform defines what counts as progress. It builds systems that measure citizens, users, workers, transactions, speech, movement, productivity, risk, and consumption. Those measurements train models and justify more deployment. People then adapt their lives to the systems that classify them. The adapted behavior becomes the next round of data.

That is why the book's most important unit is not the model. It is the implementation environment. AI power emerges where institutions can make reality machine-readable, act on the reading, and then treat the changed reality as confirmation.

The practical lesson is to review AI systems by the loop they create. Who supplies the data? Who benefits from the prediction? Who absorbs the error? Who can appeal? Who is excluded from the scoreboard? Who has to become legible to survive? Lee's book does not answer all of those questions, but it makes them unavoidable by showing that AI competition is never only a contest of intelligence. It is a contest over the conditions under which intelligence is allowed to count.

Source Discipline

This review separates book facts, author biography, the book's 2018 argument, current empirical snapshots, and legal or standards claims. WorldCat and HarperCollins establish bibliographic and current-publisher details. Stanford's event biography establishes current public biography for Lee. DigiChina supports claims about China's 2017 plan. Stanford HAI's 2026 AI Index supports current comparative claims about models, publications, patents, data centers, investment, robotics, adoption, and labor indicators. NIST, EUR-Lex, BIS, GAO, OECD, and IEA sources support governance, export-control, infrastructure, and energy context.

Those sources do not all make the same kind of claim. An AI Index statistic is a dated measurement, not a forecast of national destiny. A government plan is an official ambition, not proof of delivery. A BIS rule, guidance document, or GAO decision has legal and procedural meaning, but it does not by itself show whether a control is effective. A publisher page summarizes a book; it is not independent evidence for the book's geopolitical thesis.

This page does not claim that any present AI system is conscious, divine, or AGI. It treats AI as institutional machinery: models, data, compute, standards, procurement, labor, and public authority arranged into systems that act on people.

Sources

Book links are paid affiliate links. As an Amazon Associate I earn from qualifying purchases.

Buy on Amazon Browse Books

Return to Blog · Return to Books