Global AI Competition & Innovation Lab

Competition of AI

The Domestic Angle

The domestic angle treats artificial intelligence as a question of American market structure, not only of foreign industrial policy. It brings three bodies of law to the same stack. Antitrust asks whether control of compute, cloud, data, and foundation models is hardening into durable power — through exclusivity, switching costs, and vertical ties that rivals cannot match. Intellectual property asks when training on protected works is a fair use that spreads capability, and when it is substitution that starves the authors on whom the next generation of models depends. Consumer protection asks whether the interfaces through which households and firms meet these systems are truthful, contestable, and free of designs that convert attention into lock-in.

The stakes are already measurable. Epoch AI estimates that five U.S. hyperscalers — Amazon, Google, Meta, Microsoft, and Oracle — own more than 70 percent of global AI compute, with Google alone holding roughly a quarter. The OECD reports that the five largest U.S. technology firms spent $400 billion on capital expenditure in 2025 and are forecast to spend $660 billion in 2026, a scale that can diffuse innovation if access stays open and can foreclose it if the stack closes. Downstream, payments for access to foundation models exceeded $37 billion in 2025, tripling in a year, while nine in ten firms in Y Combinator’s summer 2025 cohort were AI-native and nearly all depended on three model providers. A domestic policy equal to that structure protects consumers from deception, keeps fair use as a doctrine that rewards transformative use rather than a blanket exemption or a blanket ban, and treats concentration as a risk to the diffusion of innovation — not as a trophy of scale.

Competition of AI

The Domestic Angle

The domestic angle treats artificial intelligence as a question of American market structure, not only of foreign industrial policy. It brings three bodies of law to the same stack. Antitrust asks whether control of compute, cloud, data, and foundation models is hardening into durable power — through exclusivity, switching costs, and vertical ties that rivals cannot match. Intellectual property asks when training on protected works is a fair use that spreads capability, and when it is substitution that starves the authors on whom the next generation of models depends. Consumer protection asks whether the interfaces through which households and firms meet these systems are truthful, contestable, and free of designs that convert attention into lock-in.

The stakes are already measurable. Epoch AI estimates that five U.S. hyperscalers — Amazon, Google, Meta, Microsoft, and Oracle — own more than 70 percent of global AI compute, with Google alone holding roughly a quarter. The OECD reports that the five largest U.S. technology firms spent $400 billion on capital expenditure in 2025 and are forecast to spend $660 billion in 2026, a scale that can diffuse innovation if access stays open and can foreclose it if the stack closes. Downstream, payments for access to foundation models exceeded $37 billion in 2025, tripling in a year, while nine in ten firms in Y Combinator’s summer 2025 cohort were AI-native and nearly all depended on three model providers. A domestic policy equal to that structure protects consumers from deception, keeps fair use as a doctrine that rewards transformative use rather than a blanket exemption or a blanket ban, and treats concentration as a risk to the diffusion of innovation — not as a trophy of scale.

Competitiveness of AI

The Global Angle

The global angle treats artificial intelligence as a traded system, not a domestic market with a foreign footnote. It covers the law and policy of cross-border commerce in chips, models, data, and cloud capacity; the supply chains that run from lithography and memory through foundries and packaging to power and undersea links; and the rivalry among the United States, China, and other powers over who sets the terms of access. Industrial policy sits inside that frame — subsidies, export controls, investment screening, and standards — alongside the market-access question of whether those tools secure critical assets or close markets that innovation needs to stay open.

The geography of capability is already stark. Epoch AI’s inventory of GPU clusters puts the United States at about three-quarters of recorded cluster performance, China at about 15 percent, and the European Union under 5 percent. Upstream, the stack narrows to a single firm: ASML is the only maker of the extreme-ultraviolet machines required for leading-edge chips, each tool drawing on more than 100,000 parts and roughly 2,000 suppliers. Governments are already spending against that concentration. The U.S. Commerce Department has awarded $30.9 billion under the CHIPS Act toward lifting America’s share of leading-edge logic manufacturing from zero in 2022 to about 20 percent by 2030, while OECD figures show Chinese support for its chip firms reaching close to 10 percent of sales in the early 2020s. A competitiveness policy equal to that structure secures the assets rivalry can cut off, keeps market access open enough for innovation to diffuse, and treats standards as a commons to be protected — not as an instrument for shutting rivals out.

Competitiveness of AI

The Global Angle

The global angle treats artificial intelligence as a traded system, not a domestic market with a foreign footnote. It covers the law and policy of cross-border commerce in chips, models, data, and cloud capacity; the supply chains that run from lithography and memory through foundries and packaging to power and undersea links; and the rivalry among the United States, China, and other powers over who sets the terms of access. Industrial policy sits inside that frame — subsidies, export controls, investment screening, and standards — alongside the market-access question of whether those tools secure critical assets or close markets that innovation needs to stay open.

The geography of capability is already stark. Epoch AI’s inventory of GPU clusters puts the United States at about three-quarters of recorded cluster performance, China at about 15 percent, and the European Union under 5 percent. Upstream, the stack narrows to a single firm: ASML is the only maker of the extreme-ultraviolet machines required for leading-edge chips, each tool drawing on more than 100,000 parts and roughly 2,000 suppliers. Governments are already spending against that concentration. The U.S. Commerce Department has awarded $30.9 billion under the CHIPS Act toward lifting America’s share of leading-edge logic manufacturing from zero in 2022 to about 20 percent by 2030, while OECD figures show Chinese support for its chip firms reaching close to 10 percent of sales in the early 2020s. A competitiveness policy equal to that structure secures the assets rivalry can cut off, keeps market access open enough for innovation to diffuse, and treats standards as a commons to be protected — not as an instrument for shutting rivals out.

Creative Destruction of AI

The Regulatory Angle

The regulatory angle treats artificial intelligence as a problem of institutional design, not of prohibition. It covers pro-innovation regulation — rules that police fraud, safety, and deception without freezing the entry on which creative destruction depends — and the unintended consequences of statutes written faster than the technology they name. It also covers fragmentation: federal, state, and foreign regimes that define the same system differently, so a firm scaling across borders inherits several compliance stacks rather than one intelligible duty.

The volume of rules is already a fact of market structure. In 2025 every U.S. state introduced AI-related bills, 1,208 in all, of which 145 became law; by March 2026 a further 1,561 had been introduced in 45 states. Globally, one tracker counts 1,142 AI instruments across 110 jurisdictions, 668 of them already in force, with the United States alone accounting for 271 once state law is included. Europe shows the same strain in the calendar: transparency duties under the EU AI Act applied from August 2026, while the standalone high-risk obligations were pushed to December 2027. A regulatory policy equal to creative destruction prefers innovation principles to the precautionary principle, and simplification to accumulation — clear duties an entrepreneur can read, meet, and scale under, rather than a patchwork that taxes entry more heavily than it disciplines harm.

Creative Destruction of AI

The Regulatory Angle

The regulatory angle treats artificial intelligence as a problem of institutional design, not of prohibition. It covers pro-innovation regulation — rules that police fraud, safety, and deception without freezing the entry on which creative destruction depends — and the unintended consequences of statutes written faster than the technology they name. It also covers fragmentation: federal, state, and foreign regimes that define the same system differently, so a firm scaling across borders inherits several compliance stacks rather than one intelligible duty.

The volume of rules is already a fact of market structure. In 2025 every U.S. state introduced AI-related bills, 1,208 in all, of which 145 became law; by March 2026 a further 1,561 had been introduced in 45 states. Globally, one tracker counts 1,142 AI instruments across 110 jurisdictions, 668 of them already in force, with the United States alone accounting for 271 once state law is included. Europe shows the same strain in the calendar: transparency duties under the EU AI Act applied from August 2026, while the standalone high-risk obligations were pushed to December 2027. A regulatory policy equal to creative destruction prefers innovation principles to the precautionary principle, and simplification to accumulation — clear duties an entrepreneur can read, meet, and scale under, rather than a patchwork that taxes entry more heavily than it disciplines harm.