The United States, China, and the multipolar sovereignty contest
The proposition that one country "leads AI" is too coarse, and a bilateral U.S.-China frame is incomplete. Models, chips, fabrication, packaging, power, capital, public compute, deployment, standards, security, and trusted access can produce different leaders and coalitions.
The original theory presents U.S. leadership in open source, commercial offerings, hardware, cost effectiveness, security, and trust as imperative to future growth. The imperative is reasonable; the descriptive claim is too broad. U.S. strengths coexist with Chinese capability convergence, European public-compute and regulatory strategy, Indian national infrastructure and data initiatives, Gulf sovereign-capital and full-stack programs, and deep dependence on allied fabrication, memory, packaging, and equipment.
A seven-dimension geopolitical baseline
A country cannot be assessed by one benchmark, one chip count or one spending total. JPMorgan Chase organizes U.S.-China competition across policy, hardware, models and software, energy and resources, finance, socioeconomics, and military and security. Version 1.2.1 uses the same disaggregated dimensions, then extends the comparison to Europe, India, Gulf states and allied manufacturing economies. The dimensions are not summed into a league-table score (JPMorgan Chase Center for Geopolitics 2026).
| Dimension | U.S. baseline | China baseline | Multipolar implication |
|---|---|---|---|
| Policy | Export, investment, safety, procurement and industrial-policy tools. | State support, domestic substitution, standards and market access. | EU regulation/public compute, India mission procurement, Gulf capital and bilateral access shape third markets. |
| Hardware | Design, EDA, equipment, NVIDIA and allied fabrication strengths. | Huawei/Ascend substitution, manufacturing depth and restricted leading-edge access. | Taiwan, Korea, Japan and the Netherlands remain indispensable; Arizona and regional projects diversify selected stages. |
| Models/software | Frontier closed models, cloud and enterprise distribution. | Rapid capability convergence and strong open-weight diffusion. | European and Indian language/domain systems and Gulf-hosted capacity create bargaining options. |
| Energy/resources | Large installed capacity but higher power costs and interconnection friction. | Faster generation expansion and lower reported data-center-region power costs, with utilization concerns. | Gulf energy/capital and regional grids can attract workloads; physical delivery still depends on imported systems. |
| Finance | JPMorgan reports $490B U.S. AI spend and 75% of global AI VC in 2025. | Report estimates $98B spend and 5% of AI VC, supplemented by state support. | Sovereign capital and public compute can alter access without matching U.S. venture depth. |
| Socioeconomics | Private-sector diffusion, labor transition, community and ratepayer constraints. | Industrial diffusion, engineering scale, regional overbuild and utilization risk. | India emphasizes language and public access; EU emphasizes institutional capacity; Gulf states use infrastructure as diversification. |
| Military/security | Allied controls, cyber capability, defense procurement and trusted access. | Military-civil integration, cyber, domestic stack and security governance. | Partners choose interoperability, autonomy and access conditions rather than one binary bloc. |
Table 8.1A. A common seven-dimension baseline without a composite national score.
Capability convergence
Stanford's 2026 AI Index reports two different convergence patterns. As of March 2026, the leading U.S. model was 2.7 percent ahead of the leading Chinese model on the report's Arena-based comparison, a gap that had fluctuated in the single digits. By contrast, the leading closed model was 3.3 percent ahead of the leading open model, up from 0.5 percent in August 2024, and six of the top ten Arena models were closed. The correct conclusion is therefore not that every gap is closing: U.S.-China capability had narrowed, while the top closed-open gap had reopened over that interval (Stanford HAI 2026a).
| Layer | U.S. position | Chinese position | Strategic variable to watch |
|---|---|---|---|
| Frontier models | Large concentration of highly capable proprietary labs and global consumer distribution | Near-parity on many benchmarks; rapid iteration and strong open-weight releases | Reliability and cost on real workflows, not benchmark peaks |
| Open-weight ecosystem | Strong research, tooling, startups, and cloud distribution | Aggressive model publication, price competition, multilingual and industrial focus | Adoption outside home markets; license freedom; total operating cost |
| Chip design and platforms | Leadership in accelerator design, CUDA-class software, EDA and much semiconductor equipment; high NVIDIA platform concentration. | Huawei Ascend full-stack substitution, large domestic demand, and sustained ecosystem investment under controls. | Matched goodput per watt, memory and interconnect, software portability, developer adoption, and total system cost. |
| Fabrication, memory and packaging | Allied supply centered on TSMC/Taiwan, Korean HBM, Japanese materials, European tools, and expanding Arizona capacity. | Large mature-node base and domestic substitution; constrained access to parts of the leading-edge global tool and HBM stack. | Qualified yield and capacity, HBM and packaging lead times, share outside Taiwan, recovery time, and equipment access. |
| Cloud and capital | Deep equity and debt markets; global hyperscalers | Large platforms, state-guided finance, and domestic scale | Cost of capital, utilization, financing durability, access in third markets |
| Power and industrial build-out | Large resource base but slow interconnection and permitting in many regions | Rapid construction capability and integrated industrial planning, with regional constraints | Time to energized MW, reliability, fuel access, transmission expansion |
| Enterprise deployment | Global software and services distribution; strong multinational customer access | Manufacturing and domestic-platform integration; state and industrial use cases | Verified ROI, exportability, and ecosystem lock-in |
| Trust, security, and alliances | Allied relationships, standards influence, and enterprise-security brands | Cost and availability appeal; different governance model | Data governance, supply-chain assurance, political acceptance |
| Sovereign and regional capacity | Allied cloud, capital, research, and standards network | Large domestic platforms, state coordination, and infrastructure scale | EU public compute, IndiaAI access, Gulf full-stack build-out, and third-market adoption |
Table 8.1. Leadership is a portfolio of positions across the stack.
The durable U.S. advantages
The United States combines frontier-lab density, advanced chip design, electronic-design automation, semiconductor equipment, global cloud platforms, venture and public capital, university research, and enterprise software distribution. The Semiconductor Industry Association describes continued U.S. leadership in chip design, EDA, and equipment, while public incentives are projected to restore a meaningful share of advanced logic fabrication by the early 2030s (SIA 2025). These advantages reinforce one another: capital funds compute, compute attracts talent, talent creates products, and distribution generates data and revenue.
The dollar financial system is another advantage. Frontier firms can raise private equity, public equity, syndicated credit, project finance, and infrastructure capital at a scale unavailable in most markets. Yet abundant capital is useful only when paired with disciplined allocation. A country can win the financing race and still earn poor returns if it builds low-utilization capacity or socializes project risk.
The durable Chinese advantages
China's advantages include manufacturing depth, a vast engineering workforce, large domestic demand, a strong publication and patent base, and the ability to integrate AI into industrial systems at scale. Stanford reports Chinese leadership in AI publication volume and citations as well as substantial patent activity, while the narrowing model gap demonstrates that export constraints have not ended capability development (Stanford HAI 2026b). Open-weight releases can also turn cost competition into geopolitical influence by making Chinese models attractive to countries and firms unable to afford premium proprietary services.
China may also benefit where infrastructure coordination and equipment manufacturing reduce the time from plan to energized capacity. This advantage should not be romanticized: local power constraints, capital misallocation, semiconductor bottlenecks, data controls, and trust concerns can slow deployment. But a U.S. strategy that assumes permanent model or open-weight superiority would be fragile.
Export controls: moat, delay, and stimulus
Advanced-semiconductor export controls are intended to restrict access to the compute required for military and frontier applications. The U.S. Department of Commerce continued to revise licensing treatment for high-end accelerators in 2026, illustrating that the control regime is dynamic rather than a one-time boundary (BIS 2024; BIS 2026). Controls can delay capability and raise cost. They also encourage stockpiling, domestic substitution, architectural efficiency, and alternative supply relationships.
The strategic measure is therefore not whether controls stop all progress. It is whether they preserve a meaningful advantage at acceptable cost to U.S. firms and allies while sustaining trust in the broader technology ecosystem. Overly broad restrictions can fragment markets and accelerate competing standards; overly narrow controls can transfer strategic compute with little friction. The policy problem is one of continuous calibration.
Cost effectiveness and security must be evaluated together
A low token price is not necessarily low economic cost. A system that requires more retries, human review, or integration may be more expensive per successful task. Likewise, a high-performing model that cannot meet security, residency, continuity, or audit requirements has limited value in critical sectors. The competitive frontier is secure total cost per successful task: the fully loaded cost of producing a verified result under the required controls.
NATIONAL COMPETITIVENESS METRIC Secure total cost per successful useful task = compute + energy + integration + supervision + security + compliance + expected error loss, divided by verified outcomes. Add deployment speed and resilience as separate constraints. |
|---|
Trust is an economic asset - and can be lost
The United States benefits when foreign enterprises believe that its technology is reliable, secure, contractually predictable, and governed by institutions that permit redress. Trust reduces the cost of adoption and expands the addressable market. It can be weakened by surveillance concerns, abrupt export restrictions, opaque model behavior, unreliable services, weak privacy protection, or the perception that strategic access can be withdrawn for political reasons. Chinese providers face their own trust constraints related to state access, governance, and geopolitical alignment.
Trust should not be treated as branding. It is produced by verifiable controls, transparent incident handling, independent testing, interoperability, legal commitments, supply-chain assurance, and continuity planning. A trusted ecosystem can command a premium even when a competing model is cheaper. A complacent ecosystem can lose that premium quickly.
Quantitative capacity and capital context
JPMorgan Chase reports installed data-center capacity of 53.7 GW in the United States, 31.9 GW in China, 11.9 GW in the European Union, 6.6 GW across Japan and Korea, 3.6 GW in India and 1.1 GW in the Middle East. Installed capacity is not equivalent to AI-ready, commissioned or utilized capacity. The report also cites 2025 AI spending of $490 billion by U.S. firms versus $98 billion by Chinese firms, U.S. receipt of 75 percent of global AI venture capital versus 5 percent for China, and electricity costs of roughly $0.07-$0.09/kWh in major Chinese data-center regions. Its statement that as many as 80 percent of Chinese data centers may be idle is an external estimate, not an official utilization census (JPMorgan Chase Center for Geopolitics 2026).
The same source supplies the counterweights. It reports a fourth-quarter 2025 national average of about $0.18/kWh in the United States versus $0.08/kWh in China, while noting wide U.S. regional variation: about $0.15-$0.22 in Silicon Valley and $0.06-$0.08 in Dallas-Fort Worth. It also reports DeepSeek's claim that V4 trails state-of-the-art frontier models by roughly three to six months. These figures complicate both simple cost-leadership and full-capability-parity claims, and they remain source-qualified estimates or company claims rather than matched independent measurements (JPMorgan Chase Center for Geopolitics 2026).
| Region | Installed data-center capacity reported by JPMorgan | Strategic reading | Measurement caveat |
|---|---|---|---|
| United States | 53.7 GW | Largest installed base, deep capital and cloud ecosystem. | Not all capacity is AI-ready, frontier-capable, energized for new loads or well utilized. |
| China | 31.9 GW | Large base, lower reported power cost and rapid generation expansion. | Advanced accelerators and reported utilization constrain effective frontier capacity. |
| European Union | 11.9 GW | Public-compute, regulatory and industrial-policy strategy. | Depends heavily on imported accelerators and allied manufacturing. |
| Japan and Korea | 6.6 GW | Critical memory, materials, manufacturing and regional capacity. | Aggregate obscures different national roles and grid conditions. |
| India | 3.6 GW | National compute access, language models and public procurement. | Mission capacity and installed commercial capacity must be separated. |
| Middle East | 1.1 GW | Energy and sovereign capital support rapid planned build-out. | Announced gigawatts and bilateral projects are not yet equivalent to commissioned capacity. |
Table 8.1B. Installed capacity is a baseline input, not a measure of productive frontier compute.
A multipolar sovereignty contest
Many countries will not replicate either the U.S. or Chinese stack, but they are not passive customers. Their strategies combine imported accelerators and models with domestic compute access, data authority, language assets, regulation, procurement and capital. These are active decoupling and bargaining strategies, not merely demand for U.S. exports.
European Union: The EU combines regulation with public and industrial compute. By April 2026 the Commission reported 19 AI Factories and 13 regional antennas, alongside the InvestAI facility and an AI IPCEI candidate. This is not merely demand for U.S. exports: it is an attempt to preserve legal control, research access, industrial applications and bargaining power while remaining dependent on allied accelerators, fabrication and power (European Commission 2023; European Commission 2025; European Commission 2026a; European Commission 2026b).
Gulf states: Stargate UAE specifies a 1 GW Abu Dhabi cluster with 200 MW expected online in 2026 and reciprocal investment in U.S. infrastructure; Saudi Arabia's HUMAIN is building across data centers, cloud, models and applications. These states use energy, sovereign capital, procurement and cross-bloc partnerships to build independent leverage. Their success depends on commissioning, operator capability, hardware access and third-market services rather than announcements alone (OpenAI 2025b; PIF 2025).
India: The IndiaAI Mission operates a national compute portal with subsidized access to multiple accelerator types, AIKosh data and model infrastructure, and a domestic foundation-model program involving Sarvam AI. The strategy combines language capability, public procurement, shared compute and domestic deployment. It is an independent sovereignty model whose performance should be judged by sustained production workloads, secure cost, access and portability rather than GPU announcements alone (IndiaAI 2026a; IndiaAI 2026b; Sarvam AI 2026).
Allied manufacturing states: Taiwan, Korea, Japan, the Netherlands and other partners specialize in fabrication, HBM, advanced packaging, equipment, materials, servers and optics. TSMC remains the central leading-edge foundry and packaging node, while NVIDIA's own platform ramp depends on a dense Taiwan systems ecosystem. Arizona, Japan and European projects diversify selected stages but do not yet replicate Taiwan's complete cadence and supplier density. Chapter 4 provides the canonical hardware treatment. Supply-chain resilience and trusted access are sovereign variables in their own right, not background assumptions.
Smaller and nonaligned states: Regional infrastructure, open-weight models, procurement standards, data-residency rules and provider diversification can preserve bargaining power without full autarky. The relevant test is whether these arrangements support critical workloads through a supplier or geopolitical shock, not whether they reproduce a frontier stack end to end.
Two geopolitical stress cases
Chinese fabrication acceleration: If domestic advanced-node, memory, packaging, equipment, and software capabilities mature faster than expected, the U.S. chip-design and export-control advantage can narrow even without benchmark leadership changing immediately.
Allied supply disruption: If fabrication, memory, packaging, materials, or equipment flows from Taiwan, Korea, Japan, or Europe are interrupted, U.S. cloud capital and model distribution cannot substitute quickly for missing physical production. "U.S. leadership" is therefore partly an alliance and supply-chain property.
Two competing U.S. leadership strategies
A frontier-champion strategy concentrates capital, closed frontier capability, compute, and national-security controls in a small number of firms. Its advantages are coordinated scale, accountability at identifiable providers, and the ability to fund expensive research. Its risks are lock-in, concentrated failure, rent extraction, and industrial dependence on a few balance sheets and governance structures.
An open-ecosystem strategy expands compute access, shared datasets and evaluation assets, capable open weights, application-layer competition, and sector-specific deployment. The July 24 coalition statement presents this as leadership through diffusion into ordinary economic workflows rather than ownership of one leading model. Its advantages are specialization, contestability, and a broader developer base; its risks are irreversible capability release, operational fragmentation, and continued concentration at the chip, cloud, and control-plane layers (American Innovators Network et al. 2026).
The strategies are not mutually exclusive. A resilient U.S. portfolio can support frontier closed models and a capable open-weight ecosystem while applying different safeguards to different capabilities and deployment contexts. The adjudicating measures are secure task cost, independent adoption, switching power, safety performance, third-market diffusion, and the distribution of margins and control across layers.
The U.S. policy objective
The defensible objective is not to guarantee that every leading company is American or that every model is closed. It is to maintain an ecosystem in which frontier research, capable open weights, shared evaluation and compute assets, competitive applications, infrastructure, talent, finance, and trusted governance reinforce one another. Policy should distinguish legitimate model improvement from unlawful extraction through targeted, evidence-based rules rather than assuming that either openness or closure is inherently safe. Leadership should be measured by third-market adoption, secure task cost, infrastructure deployment, supply-chain resilience, switching power, and the ability of ordinary firms and public institutions to use AI productively - not by one benchmark or one champion.