Key Takeaways
- AI diffusion is fastest where governments treat it as a coordination problem (Asia) rather than purely a liability problem (EU).
- Enterprise AI adoption ranges from 61 % in Singapore to 20 % EU‑wide, with a large‑small firm gap of 38 percentage points in the EU.
- Productivity gains from generative AI are measurable: 5‑25 % gains in individual tasks and 1‑2 percentage‑point boosts to annual growth when adoption widens.
- Five distinct Asian models exist: mandate‑led diffusion (China), statute‑plus‑capital (South Korea), subsidy‑as‑infrastructure (India), capital‑intensive catch‑up (Japan), and governance‑by‑measurement (Singapore).
- The EU’s AI Act focuses on risk allocation and compliance, which does not solve the underlying adoption coordination failure and leaves SMEs behind.
- Without demand‑side policies that lower adoption barriers, the EU risks falling further behind in productivity, innovation, and global AI value capture.
Framing the AI Race: Coordination vs. Liability
Artificial intelligence has become a central axis of geopolitical competition, yet most analyses stop at input metrics such as GPU counts or investment totals. The decisive question for scholars, policymakers, and strategists is why different political economies adopt AI in divergent ways and what those choices reveal about assumptions on state capacity, risk tolerance, and the link between technology and national power. In Asia, AI is framed as a coordination problem that the state must solve; in the European Union, it is framed primarily as a liability problem that the state must manage. This framing difference produces structural consequences that compound over time and are now evident in the 2024‑2026 data.
Global Adoption Landscape: What the Data Show
The Stanford AI Index 2026 reports that generative AI reached 53 % global population adoption within three years—faster than the PC or internet. Organizational adoption, per McKinsey’s 2025 State of AI survey, stands at 88 % of firms using AI in at least one function, up from 78 % the prior year. Global corporate AI investment more than doubled in 2025 to $581.7 billion, with private investment alone at $344.7 billion. Yet these aggregates mask stark divergence: population‑level adoption is 61 % in Singapore but only 28.3 % in the United States, while the EU27 enterprise average sits at 20 % (Eurostat 2025), concealing a chasm between large firms and small ones.
Productivity Gains from AI Adoption
Empirical evidence links adoption to tangible productivity improvements. OECD experimental studies reviewed in a July 2025 paper find average gains of 5‑25 % in customer support, software development, and consulting when generative AI is integrated. Macro‑level research from April 2026 shows that the rise in frequent AI users—from 12 % in mid‑2024 to 26 % by late 2025—corresponds to roughly 1.4‑2.8 % higher real output, or about one to two percentage points of annualized growth. These are measured outcomes, not projections, indicating that economies that close the adoption gap earliest will capture compounding gains first.
Three Models of State‑Led AI Adoption
The five Asian economies examined—China, South Korea, Japan, India, and Singapore—share a common premise: AI adoption will not reach required scale or speed without deliberate state intervention to accelerate demand. The form that intervention takes varies, mapping onto each polity’s structural features. The models range from administrative mandates to statutory frameworks paired with capital, from input subsidies to precision measurement, and from capital‑intensive catch‑up to governance by measurement.
China: Mandate‑Led Diffusion
China’s “AI Plus” guideline, issued by the State Council in August 2025, targets 90 % penetration of intelligent terminals and AI agents across six sectors by 2030. Beijing treats AI diffusion like past infrastructure rollouts—electrification or broadband—as a coordination problem where market mechanisms underprice adoption because individual actors cannot capture the full social return. The solution is to legislate adoption rates, not merely subsidize inputs. Early results are striking: China’s AI user base exceeded 515 million in the first half of 2025, growing at 36.5 % in six months after the DeepSeek‑R1 launch, and enterprise adoption stands at roughly 58 % (IBM 2025). China also leads globally in AI patent applications, holding 38.58 % of the world total as of April 2025.
South Korea: Statute‑Plus‑Capital
South Korea’s Framework Act on the Development of Artificial Intelligence, effective January 2026, made it the first nation to consolidate governance, industrial policy, and risk management into a single statute. Coupled with a tripling of the national AI budget to 10.1 trillion won (≈$6.94 billion) and the deployment of 260,000 Blackwell GPUs across major conglomerates and government infrastructure, Korea runs regulation and investment in parallel. The statute serves as a coordination mechanism, while the capital commitment signals credibility to private actors. Consequently, South Korea leads the world in AI patents per capita, reflecting deep industrial integration.
India: Subsidy‑as‑Infrastructure
The IndiaAI Mission, approved March 2024, treats compute as a public good akin to earlier development strategies for electricity. By deploying over 38,000 GPUs—including 1,050 Google Trillium TPUs—and offering H100‑class compute at about ₹65 ($0.72) per hour, the mission removes the compute bottleneck that constrains downstream activity. Enterprise AI adoption has responded, reaching 57 % (IBM 2025). Minister Ashwini Vaishnaw reported in February 2026 that committed AI‑related investment totals $90 billion, with projections to exceed $400 billion across the AI stack within two years. India’s strategy bets that lowering input costs will generate compounding returns.
Japan and Singapore: Catch‑Up and Precision
Japan exemplifies a capital‑intensive catch‑up model, responding to a documented adoption failure. The Ministry of Internal Affairs and Communications’ 2025 White Paper showed generative AI usage among the public at 26.7 % in 2024, far below the U.S. (68.8 %) and China (81.2 %). Core business AI use among firms was a mere 1.9 % (OECD 2024), revealing a deep integration gap. Tokyo’s reply has been to deploy capital—NTT pledges $59 billion through 2027 and SoftBank commits over $40 billion to OpenAI’s Stargate—testing whether raw spending can overcome structural barriers such as risk‑averse corporate culture.
Singapore, by contrast, is a precision case: governance by measurement. The National AI Strategy 2.0 (May 2026) emphasizes granular targets—15,000 AI talents and 10,000 enterprises over three years—and leverages the city‑state’s small scale to create a tight feedback loop. With 61 % population‑level generative AI adoption (the world’s highest alongside the UAE) and non‑SME enterprise adoption at 62.5 % (IMDA 2024), Singapore can measure deployment, observe uptake in near real time, and adjust policies swiftly, functioning as a living laboratory for larger states.
The EU’s Structural Misjudgment: Evidence from the Adoption Data
Set against the Asian approaches, the EU’s position reflects a different theory of AI governance—one centered on risk allocation rather than demand stimulation. Eurostat’s December 2025 release records EU27 enterprise AI adoption at 20 %, up from 13.5 % in 2024—a genuine 6.5‑percentage‑point increase. However, adoption is sharply stratified by firm size: large enterprises (≥250 employees) reached 55 % adoption in 2025, while small enterprises (10‑49 employees) lagged at 17 %. The 38‑percentage‑point gap widened by more than 7 points in 2025, indicating that growth is concentrated among big firms.
The AI Annex III compliance costs for high‑risk AI impose initial expenses of €200,000‑€500,000 for SME providers, phased across 2025‑2027. Despite proportional fees, sandbox access, and simplified documentation, 12 of 27 member states missed the August 2026 deadline to appoint competent national authorities. Over 60 % of EU startups now prioritize low‑risk AI to avoid compliance exposure, shaping innovation trajectories before the Act is fully enforced. Private AI investment in Europe was only $20.9 billion in 2025, versus $285.9 billion in the United States—a 13.7‑to‑one ratio—underscoring the investment shortfall that compounds the adoption problem.
AI Adoption as a Collective Action Problem
The comparative evidence supports viewing AI diffusion as a collective action problem akin to those analyzed in industrial policy literature. Individual firms possess information about AI’s benefits, yet adoption creates positive externalities—data generation, workforce upskilling, process knowledge—that spill across firm boundaries. Because firms cannot fully capture these social returns, market‑rate adoption undershoots the socially optimal level. State intervention is therefore a correction for a market failure, not a distortion.
What varies across cases is the mechanism chosen to address the problem. China uses administrative mandates; Korea combines a statutory framework with direct capital; India employs price subsidies on scarce compute; Singapore relies on precision measurement and targeted program support; Japan attempts to substitute private capital concentration for demand‑side coordination. The EU, however, deploys a compliance‑first architecture that tells firms what they cannot do but does not solve the underlying coordination failure. As the OECD’s experimental evidence shows, productivity gains require meaningful human‑AI collaboration and task‑fit, which presuppose widespread adoption—something a pure risk‑allocation regime cannot deliver.
Implications for Firms and Policymakers
The productivity evidence gives the stakes empirical weight: studies summarized in the Stanford AI Index 2026 report gains of roughly 14‑15 % in customer support, 26 % in software development, and 50 % in marketing output from AI integration. These task‑level benefits translate into economy‑wide growth only when adoption is broad. If China reaches 90 % AI penetration by 2030 while the EU stagnates at 30‑40 %, the resulting productivity differential will be difficult to overcome through regulation alone. Early adopters accumulate process knowledge, generate richer training data, and build organizational capabilities that latecomers cannot simply purchase.
For multinational corporations, the asymmetry creates both risk and arbitrage. Firms that develop AI capabilities in Asia’s adoption‑accelerating environments will enter the EU market with superior skills and models, forcing EU‑based rivals to play catch‑up. The AI Act’s extraterritorial reach means compliance costs fall on the European operations of firms that built their capabilities elsewhere—an ironic inversion of the regulation’s protective intent.
For EU policymakers, the implication is not to discard the AI Act’s risk‑stratification logic, which addresses legitimate safety and fundamental‑rights concerns. Instead, the sequencing must change: demand‑side measures that lower adoption barriers—such as subsidised compute for SMEs, sector‑specific penetration mandates, or coordinated budget commitments—are needed to complement the regulatory framework. The EU’s chip ambitions (targeting 20 % of global production value by 2030) are serious on the supply side, but building that supply without a corresponding enterprise adoption base is akin to constructing a motorway network without linking it to the towns it is meant to serve.
Conclusion
The AI race is less a contest over the most sophisticated regulations or the largest nominal investment totals and more a competition over which political economies solve the adoption coordination problem first. China’s 58 % enterprise adoption, India’s 57 %, Singapore’s 61 % population‑level adoption, and South Korea’s leadership in patent density are measurable outputs of deliberate demand‑side policy. The EU’s 20 % enterprise average, its 38‑percentage‑point large‑small firm gap, and its $20.9 billion in private AI investment (versus $285.9 billion in the U.S.) reflect a different orientation—one that prioritizes liability management over coordination. Both orientations respond to real challenges, but they are not addressing the same challenge with equal urgency. Recognizing the coordination problem as the binding constraint opens a tractable path forward: align regulatory safeguards with proactive measures that stimulate widespread, productive AI adoption. If the EU can correct this sequencing before the window closes, it still has a chance to capture a share of the projected digital value currently slipping through its grasp.

