The global AI “race” is still narrated like the space race. This unhealthy obsession with faster models, bigger chips, and first to the frontier is politically convenient, but strategically incomplete. The contest should not only be about what the United States can build first but also about what the world will use, trust, and depend on. The winning stack will be the one that diffuses through supply chains, standards bodies, cloud contracts, developer ecosystems, and public institutions.

Diffusion is as strategically important as frontier breakthroughs. If Washington wants to compete, it must treat allied adoption, institutional alignment, and development finance as core instruments of AI strategy, not side projects. Markets will not fully solve this on their own, especially in emerging economies where capital costs, risk premia, and state capacity constrain adoption.[1] The United States has more tools than export denial.^[*] It should use them.

Problems with Current AI Debates

Most U.S. debates about AI are built on hidden assumptions about how AI develops, how it spreads, and how power is exercised through it.

These assumptions often go unspoken, yet they quietly determine how national security interests are defined and which policies are prioritized. This section identifies and clarifies three of the most important underlying assumptions. Different starting beliefs produce radically different strategies.

1. Pace: Many policy arguments rest on implicit beliefs about how quickly AI capabilities will advance.

Some assume imminent discontinuities — that is, sudden breakthroughs that radically change military power, economic productivity, or intelligence capabilities within a short period of time. In this view, AI development resembles the Manhattan Project or the early nuclear age, under which whoever reaches the frontier first gains an overwhelming advantage.

Others assume AI progress will be gradual and incremental. In this view, improvements accumulate slowly, diffusion happens over time, and adaptation matters more than a single breakthrough moment.

A policy that overcommits to either story risks misallocating scarce resources. Overestimating speed can justify panic and overcentralization. Underestimating speed can leave institutions unprepared.

2. Copying: Frontier breakthroughs may be hard to replicate. Or they may be easy to imitate once techniques diffuse. The most likely world is mixed. Some capabilities will be commoditized. Some will remain bottlenecked by compute, data, and integration.

Here, compute refers to the specialized hardware and large-scale data center infrastructure required to train and run advanced AI models — particularly AI accelerators (GPUs and custom chips), server clusters, and the energy systems that power them. In short, compute is the industrial backbone of modern AI.

The central question is whether frontier capabilities are easy to copy once demonstrated.

Empirically, the evidence already points to a dual-track reality. Frontier development is becoming dramatically more capital-intensive. Training costs for leading models have been rising roughly 2–3x per year, and projections suggest the largest training runs could exceed $1 billion by the late 2020s, limiting frontier development to a small set of firms and states.[2]

This concentration is reinforced by hardware bottlenecks. For example, a single company currently controls roughly 80% of the global AI accelerator market, meaning access to cutting-edge compute remains structurally centralized.[3]  Globally, access to advanced chips is highly unequal. Vili Lehdonvirta, professor at Oxford University’s Internet Institute, concluded that about 30 countries possess the majority of advanced AI compute infrastructure, with many parts of the world effectively excluded from frontier development.[4] This means that countries that don’t have control over AI infrastructure are left with no legislative agency and are more malleable to an AI world shaped by others.

3. China’s Strategy: Washington often treats Beijing as an opponent trying to “win” the frontier. But China’s approach also focuses on exporting infrastructure and governance logics through digital systems, telecommunications, cloud, and surveillance-adjacent platforms. Matthew S. Erie and Thomas Streinz, legal scholars specializing in global data governance and digital regulation, describe this dynamic as the “Beijing Effect.” Rather than a simple, one-size-fits-all model of digital authoritarianism, they argue that China’s influence spreads through a combination of push and pull factors. In practice, many emerging economies actively seek Chinese-built digital infrastructure and increasingly emulate Beijing’s approach to data governance in pursuit of data sovereignty, industrial development, and technological self-sufficiency.[5]

In practice, this means adopting stronger state control over data localization, restricting cross-border data flows, embedding public security access provisions into telecom systems, and linking digital infrastructure to domestic industrial policy goals. Data sovereignty, in this context, is not merely about privacy. It is about ensuring that data generated within a country is stored, processed, and leveraged domestically. And often under frameworks that privilege state oversight and national champions. For example, China’s Digital Silk Road explicitly supports recipient countries’ telecoms networks, AI capabilities, cloud computing, e-commerce, mobile payments, as well as surveillance and smart city tools.[6] This is the Beijing Effect in practice where infrastructure shapes regulatory habits.

Why these Assumptions Change Strategy

These assumptions fundamentally alter what “AI strategy” means and for whom. If progress is fast and copying is easy:

• Doubling down on compute becomes a temporary advantage at best.
• Private firms compete primarily on speed and product deployment.
• Smaller states can rapidly adopt frontier capabilities through APIs and open models.
• The key strategic variable becomes integration and institutional agility, not chip stockpiles.

If progress is slower and bottlenecks persist:

• Control over compute, advanced chips, and large-scale data centers becomes crucial.
• Industrial policy and ally scaling determine long-term advantage.
• Smaller and resource constrained states become structurally dependent on those who control infrastructure.
• Standards-setting and governance export become powerful levers of influence.[7]

In both worlds, diffusion is central, but it operates differently. In a fast-copy world, diffusion is rapid and competitive. Advantage comes from who integrates AI most effectively into enterprises, defense systems, and public administration. In a bottlenecked world, diffusion is selective and infrastructure-driven. Advantage comes from who controls the channels through which AI systems are accessed and governed.

Either way, diffusion cannot be perceived as an afterthought in strategic brokership. It is the mechanism through which technical capability becomes geopolitical leverage. Frontier breakthroughs matter, but only insofar as they reshape the architecture of adoption.

The Overlooked Variable: Democratic Sustenance as Strategic Objective

A second blind spot sits beneath the technical debate: the AI contest will shape whether democratic systems remain resilient both at home and abroad. Democratic resilience is not automatic. It is produced through institutions, legitimacy, and the everyday performance of state capacity.

Historically, U.S. power has never been only military. The U.S. has invested in developing institutional and economic strength. The United States built influence by anchoring international institutions, underwriting development, and offering a growth model that was legible to partners. This is the logic of postwar reconstruction, Bretton Woods institutions, and decades of global development and security assistance.[8]

The U.S. has long wielded influence through international institutions and development aid. Recently, foreign assistance has been considered politically controversial in Washington, but is a fiscally small part of the federal budget. Pew calculates U.S. foreign aid spending in FY2023 at $71.9 billion, about 1.2% of total federal outlays.[9]

When it comes to tech alliance and AI diplomacy, China clearly engages in more and broader AI cooperation diplomacy. Here, AI diplomacy is defined as the pursuit of formal cooperative agreements on AI investment, research, development, and adoption.[10] China has 32% more AI cooperation announcements than the U.S. does.

After DeepSeek (a popular Chinese-based LLM) R1’s launch in January 2025, China’s global market share jumped from ~3% to ~13%. RAND, in its recent analysis of global use patterns of large language models, found that the largest gains in terms of market share by DeepSeek occurred in developing countries and nations with closer ties to China (Russia, the Middle East, Africa, South America), while U.S.-aligned countries and North Atlantic Treaty Organization members experienced smaller increases.[11]

This matters for AI diffusion. If the U.S. wants to compete with an alternative model of techno governance, or better known as the “Beijing effect”, it must invest in the channels that create trust: infrastructure, training, institutions, and predictable partnership. Investing in AI development abroad would export legitimacy and dominance in the techonomic realm by making its partners richer, more capable, and less dependent on coercive systems.

U.S. soft power ultimately shapes AI diffusion by determining which countries choose American-built systems, standards, and governance frameworks and which turn elsewhere, thereby converting institutional trust into long-term technological alignment.

Diffusion Matters as Much as Frontier Breakthroughs

Leadership in AI will not be determined by who builds the most impressive model first but by who diffuses AI most effectively across their economy, institutions, and military.

This argument is developed most clearly by Jeffrey Ding, a political scientist and author of Technology and the Rise of Great Powers (2024). Ding argues that the U.S. is mistakenly treating AI competition as a short innovation sprint toward artificial general intelligence. Instead, he contends that AI resembles earlier general-purpose technologies such as electricity or computing: their geopolitical impact depended less on who invented them first and more on which country diffused them most broadly into productive processes.[12] Electricity did not transform American productivity when the dynamo was invented; it did so decades later, after factories reorganized, workers were retrained, and complementary infrastructure matured. The same logic applies to AI. Benchmark performance is not the same thing as institutional transformation.

That is why diffusion matters as much as frontier breakthroughs. The systems that spread globally will embed regulatory assumptions, procurement norms, data governance practices, and institutional habits. AI will not diffuse as a neutral tool. It will diffuse as infrastructure, which shapes governance, whether or not that is its stated purpose.

The strategic competition, therefore, is not simply U.S. versus China in raw capability. It is a competition over which institutional ecosystems become the default. It is about whose standards, procurement models, cloud architectures, and training pipelines become embedded across third markets.

As already established, China has been explicit about shaping global AI governance and widening AI partnerships through proposals for international cooperation mechanisms, capacity-building initiatives, and infrastructure-backed deployment. Its approach bundles hardware, cloud services, training, financing, and state-backed implementation into integrated packages.

The United States, by contrast, has increasingly treated diffusion through the lens of restriction. The Biden administration’s January 2025 Framework for Artificial Intelligence Diffusion creates a tiered global control architecture governing advanced chips, controlled model weights, and the conditions under which compute can be deployed. It structures access premised on who qualifies, how much capacity is permitted, and under what compliance obligations.

Export controls are strategically justified, especially where direct military applications are concerned. But Ding’s historical warning is relevant here: no great power has successfully monopolized a general-purpose technology. During the Industrial Revolution, Britain attempted to restrict the emigration of skilled steam engineers to preserve its advantage. The policy largely failed. Skilled workers left, diffusion continued, and industrial capacity spread. The lesson was not that Britain should have tried harder to block diffusion; it was that industrial leadership ultimately depended on broader system capacity, not secrecy alone.

A strategy centered primarily on denial risks repeating that pattern. Restriction may delay adversaries at the frontier. It does not build the complementary institutions required to win the diffusion marathon.

Here is the deeper problem. Diffusion is not simply “more adoption.” It is a geopolitical selection. If countries cannot access trusted U.S.-linked compute, models, financing, and training pipelines, they will procure them elsewhere. If they cannot finance data centers and cloud modernization through U.S.-aligned channels, they will accept subsidized alternatives. If they lack technical workforce development partnerships, they will import turnkey systems that come bundled with governance defaults.

In many markets, “choice” is structured less by ideology than by financing terms, standards alignment, and institutional capacity. That is why infrastructure bundling is powerful. It reduces friction.

The democratic implication is not that the U.S. model is inherently virtuous. It is that democratic resilience depends on institutional performance. If democratic states cannot deliver digital infrastructure, productivity growth, and technological modernization, they weaken their own legitimacy. Diffusion shapes that outcome because it determines whose systems become embedded in tax administration, public services, logistics networks, media ecosystems, and critical infrastructure.

AI diffusion sets the operating system of modern governance.

How the U.S. can Operationalize Diffusion

If diffusion is strategic, then Washington needs a diffusion toolkit and not just a frontier toolkit.

1) Build a “trusted alternative stack” for partners

Partners need an AI offer they can actually adopt: secure cloud pathways, audited model supply chains, interoperable standards, and talent training. This is not charity. It is market creation plus alignment.

Concrete moves:

• Shared compute arrangements with allies. Not symbolic MOUs. Real capacity, pooled procurement, and interoperable security standards.
• Reference architectures for government adoption: identity, payments, public services, procurement systems, and AI assurance. The democratic advantage is competence plus accountability.

2) Treat development finance as a core AI instrument

This is the missing piece in most U.S. strategy documents. If AI is infrastructure, then financing decides diffusion.

The U.S. International Development Finance Corporation (DFC) was built as a high standard alternative to China’s infrastructure statecraft. By FY2025, DFC reported combined exposure of about $43.4 billion.[13] Afreen Akhtar, visiting scholar with the American Statecraft Program at Carnegie, describes DFC as having built a portfolio on the order of tens of billions across more than one hundred countries, explicitly linked to strategic competition. She argues that while DFC represents a meaningful institutional step toward economic statecraft, it remains constrained by structural, regulatory, and political limitations that prevent it from operating at the speed, scale, and strategic coherence required to compete with China’s state directed financing ecosystem. All criticisms aside, AI should be treated as – if not the most important – then certainly one of the priority sectors with which the DFC should align its strategy, and the institution’s efforts to support AI diffusion should be expanded.

Use DFC for AI diffusion:

• Finance data centers, cloud migration, secure connectivity, and energy reliability for AI adoption.
• Bundle financing with governance conditions that protect transparency, privacy, and auditability.
• Focus on “place markets,” directing funding to locations where commercial cloud alone will underinvest due to risk and thin margins.

Markets will not automatically finance the democratic stack at scale in high risk environments. If Washington refuses to finance AI diffusion abroad, others will fill the void and whoever finances AI diffusion will determine its future shape.

3) Align export controls with a positive offer

Export controls should not function as a moral lecture to partners. They should serve as a security perimeter around a shared ecosystem.

A practical posture:

• Tighten controls on adversarial military-relevant pathways.
• Simultaneously expand trusted access lanes for partners that meet security and governance requirements.
• Pair controls with deployment support: training, integration, and safety tooling.

If Washington only tells partners what they cannot do, it will lose the narrative and the market. If it helps partners deploy safely, it builds dependence of the healthy kind: on standards, interoperability, and trusted supply.

4) Make durability a strategic objective

Allies and partners plan in decades. AI infrastructure is long lived. Policy volatility is therefore a competitive disadvantage.

To build strong international relationships, Washington should create institutional continuity:

• Multi year diffusion funding authorities.
• Cross administration guardrails for export control predictability.
• A stable allied governance forum for standards and assurance, so diffusion is rule based, not personality based.

Conclusion

The U.S. will not “win” AI through technological breakthroughs alone. Breakthroughs are necessary, but insufficient. Dominance in AI flows not from invention, but from diffusion. The decisive question is not who reaches the frontier first. It is whose systems become embedded in supply chains, public institutions, military organizations, and global markets. Diffusion shapes productivity, institutional capacity, and ultimately political legitimacy.

Export controls can slow adversaries at the edge. They cannot substitute for the harder work of building adoption capacity at home and abroad. If Washington treats diffusion as secondary, it risks conceding the terrain where governance models are quietly locked in.

The country that builds the most advanced model may win headlines but the country whose systems are most widely adopted and trusted will shape the global order.

 

 

 

[*] Export Controls are federal laws that govern how technology, technical data, technical assistance, and items or materials (from software to satellites and more) are physically or electronically exported, shipped, transmitted, transferred, or shared from the U.S. to foreign countries, persons, or entities.

[1] Jake Sullivan and Tal Feldman, “Geopolitics in the Age of Artificial Intelligence: Strategy and Power in an Uncertain AI Future,” Foreign Affairs, January 27, 2026, https://www.foreignaffairs.com/united-states/geopolitics-age-artificial-intelligence.

[2] Ben Cottier, Robi Rahman, Loredana Fattorini, Nestor Maslej, and David Owen, “How Much Does It Cost to Train Frontier AI Models?” Epoch AI, June 3, 2024 (updated January 13, 2025), https://epoch.ai/blog/how-much-does-it-cost-to-train-frontier-ai-models

[3] Bao Tran, Patent Attorney, “The AI Chip Market Explosion: Key Stats on Nvidia, AMD, and Intel’s AI Dominance,” PatentPC, March 3, 2026, https://patentpc.com/blog/the-ai-chip-market-explosion-key-stats-on-nvidia-amd-and-intels-ai-dominance

[4] Billy Perrigo, “Exclusive: New Research Finds Stark Global Divide in Ownership of Powerful AI Chips,” Time, August 28, 2024, https://time.com/7015330/ai-chips-us-china-ownership-research/.

[5] Matthew S. Erie and Thomas Streinz, “The Beijing Effect: China’s Digital Silk Road as Transnational Data Governance,” New York University Journal of International Law and Politics 54, no. 1 (Fall 2021): 1–91, https://www.nyujilp.org/wp-content/uploads/2022/02/NYUJILP_Vol54.1_Erie_Streinz_1-91.pdf.

[6] Joshua Kurlantzick, “Assessing China’s Digital Silk Road: A Transformative Approach to Technology Financing or a Danger to Freedoms?” Council on Foreign Relations, December 18, 2020, https://www.cfr.org/articles/china-digital-silk-road

[7] Jeffrey Ding, “Running the Right Artificial Intelligence Race: A National Strategy for AI Diffusion,” RAND Corporation, Expert Insights, September 30, 2025, https://www.rand.org/pubs/perspectives/PEA4165-1.html.

[8] Council on Foreign Relations, “A Brief History of U.S. Foreign Aid,” CFR Education, last updated March 31, 2023, https://education.cfr.org/learn/reading/brief-history-us-foreign-aid

[9] Drew DeSilver, “What the Data Says About U.S. Foreign Aid,” Pew Research Center, February 6, 2025, https://www.pewresearch.org/short-reads/2025/02/06/what-the-data-says-about-us-foreign-aid/.

[10] Austin Horng-En Wang and Kyle Siler-Evans, “U.S.-China Competition for Artificial Intelligence Markets: Analyzing Global Use Patterns of Large Language Models,” RAND Corporation, January 14, 2026, https://www.rand.org/pubs/research_reports/RRA4355-1.html.

[11] Austin Horng-En Wang and Kyle Siler-Evans, U.S.-China Competition for Artificial Intelligence Markets”

[12] Austin Horng-En Wang and Kyle Siler-Evans, U.S.-China Competition for Artificial Intelligence Markets:

[13] U.S. International Development Finance Corporation, Annual Management Report: Fiscal Year 2025 (Washington, DC: U.S. International Development Finance Corporation, 2025), https://www.dfc.gov/sites/default/files/media/documents/DFC%20Annual%20Management%20Report%20-%20FY%202025.pdf