While AI debates often center on ethics and regulation, they frequently overlook the material foundation, which is “compute”. The advanced semiconductors, massive data centers, and energy infrastructure required to sustain AI are increasingly concentrated among a select few states and corporations.

This insight argues that AI inequality has shifted from a question of digital divide (access to use) to a question of structural power (capacity to produce). By integrating Susan Strange’s structural power framework with Farrell and Newman’s concept of weaponized interdependence, the study maps an emerging three-tier global hierarchy. It demonstrates how compute-as-infrastructure is redrawing the map of global power and explores the narrowing path for middle powers striving for strategic autonomy in a compute-dependent world.

From digital divide to compute hierarchy

When people talk about inequality in AI, the conversation usually moves in one of three directions. The first one is access, who has internet, who can afford a smartphone, and who can use ChatGPT. The second one is bias, which algorithms misjudge, mistreat, or render invisible. The third one is regulation, whether countries have laws strong enough to govern AI safely. Each of these is a real problem and the leading reports in the field, the International Monetary Fund’s AI Preparedness Index, which ranks countries by their readiness to adopt AI, [1] the United Nations Development Program’s Next Great Divergence, a flagship study warning that AI may widen gaps between countries, [2] and Bhaskar Chakravorti’s writing on what he calls “artificial inequality” [3] devote most of their pages to these three axes. [4]

But there is something missing in this picture, and the gap is widening every year. AI does not run on ideas, regulations, or even data alone. It runs on machines. As Kate Crawford has argued in her work on the material foundations of AI, the technology rests on a planetary infrastructure of minerals, energy, and labor that the public conversation rarely sees. [5] Specifically, it runs on a narrow, expensive, and physically demanding stack of equipment which are the advanced semiconductors that perform the calculations (the most sophisticated of which are made by a single Taiwanese company, TSMC, using lithography machines built by a single Dutch firm, ASML), the hyperscale data centers that house tens of thousands of those chips under one roof, and the electrical grids and water systems that keep them running. Together, these elements are what the industry calls compute. Without compute, there is no AI to regulate, no model to debias, and no service to enter. And compute is concentrated geographically, financially, and politically, in a remarkably small number of hands.

This is the gap this insight sets out to fill. The argument is simple to state, even if its implications are not. AI inequality has shifted from a question of the digital divide, from who can use the technology, to a question of structural power, which is who can produce it. The familiar metrics of usage and entry tell us about the symptoms of the new global order. A country with high smartphone penetration and well-trained engineers may still find itself locked out of the AI frontier because it cannot buy the chips, cannot afford to build the data centers, or cannot guarantee the electricity needed to run them. Its citizens may have AI, but its government does not have leverage over how AI is built, on what terms it can be used, or whether reach can be cut off when geopolitical winds change.

Structural power and weaponized interdependence

The first is structural power, a concept developed by the British scholar Susan Strange in the late 1980s. Strange’s insight was that conventional analyses of world politics paid too much attention to what she called relational power, the ability of one state to make another do something, and too little to the deeper power that shapes the rules of the game itself. Structural power, in her account, is the capacity to determine the framework within which other actors must operate, often without anyone noticing the framework is there. [6] She identified four domains in which this kind of power operates security: (who can protect or threaten whom), production (who controls how goods are made), finance (who creates and distributes credit), and knowledge (who decides what counts as expertise and on what architecture that expertise depends). A state that dominates even one of these domains, Strange argued, can shape the choices of others without ever issuing a direct order. [7]

The argument here is that compute now constitutes a fifth structural domain or, more precisely, that it has become the material substrate on which the other four increasingly depend. National security planning rests on AI-enabled intelligence and autonomous systems. Production runs through algorithmic supply chains. Finance is reorganizing around AI-driven trading, credit scoring, and risk modeling. And knowledge itself, once the most abstract of Strange’s categories, is now generated, mediated, and filtered through large models that only a handful of organizations can train. To command compute, in other words, is increasingly to set the conditions under which security, production, finance, and knowledge unfold. This is why several recent observers have started describing the present moment as a “technopolar” order, one in which a small number of technology firms exercise the kind of structural influence once reserved for great powers. [8]

The second concept that helps clarify the picture is weaponized interdependence, introduced in 2019 by the political scientists Henry Farrell and Abraham Newman. Their starting point is that globalization did not produce a flat, decentralized world; instead, it produced a small number of hubs through which most cross-border activity must pass. The actors that sit at those hubs, such as the financial messaging system SWIFT, the U.S. dollar clearing system, and major cloud providers, gain two distinct capacities. The first is what Farrell and Newman call the panopticon effect, the ability to observe everything flowing through the hub, generating intelligence on every user. The second is the chokepoint effect, the ability to cut specific actors off from the hub entirely, denying them service when political circumstances demand. [9] What looks from a distance like neutral material foundation turns out, on closer inspection, to be a switch that someone, somewhere, has their hand on.

Compute is almost a textbook case of both phenomena at once. The U.S. firm Nvidia, which designs the graphics processing units (GPUs) that train and run nearly every frontier AI model, holds an estimated 88 percent of the worldwide GPU market for AI workloads. [10] The actual manufacturing of those chips is even more concentrated: the Taiwanese firm TSMC produces more than 90 percent of the world’s most advanced semiconductors, and the lithography machines that make production possible at the leading edge come from a single Dutch firm, ASML, which holds an effective monopoly on extreme ultraviolet (EUV) technology. [11] The hyperscale cloud infrastructure that delivers compute to most enterprise and government users is dominated by three American companies: Amazon Web Services, Microsoft Azure, and Google Cloud, which together command well over 60 percent of the global market. [12] The visibility and the leverage these positions confer are exactly what Farrell and Newman describe: every model trained, every query routed, every export license granted or denied flows through a handful of hands.

The three-tier global compute hierarchy

Tier 1: The Architects

At the top of the hierarchy sit two states that, between them, dominate most of the compute stack from chip design to deployed model: the United States and, in a parallel and partially decoupled track, China. The United States hosts the firms that design the most advanced AI chips (Nvidia, AMD), the dominant cloud platforms (Amazon, Microsoft, Google), and the leading frontier-model laboratories (OpenAI, Anthropic, Google DeepMind, Meta). In 2023 alone, the United States captured US$67.2 billion in private AI investment, 8.7 times China’s figure and more than the rest of the world combined. [13] China, blocked by U.S. export controls from availability to the most advanced foreign chips, has responded with a parallel architecture: Huawei’s Ascend chips, the SMIC foundry working at increasingly competitive process nodes, hyperscalers like Alibaba Cloud and Tencent Cloud, and models like Qwen and DeepSeek that demonstrated, in late 2024 and early 2025, that frontier-comparable performance can be achieved at a fraction of Western training costs. [14] The architects are the only actors who can, in principle, establish a complete AI capability without depending on anyone else, though China’s autonomy at the leading edge remains constrained, and probably will remain so for some years. [15]

Tier 2: The Aspiring Builders

Below the chokepoints sits the most critically active tier in the contemporary moment: states with significant capital, real technological ambition, and growing investment portfolios in AI infrastructure, but without the ability to produce the most consequential parts of the stack themselves. India is establishing a sovereign AI compute mission with a US$1.25 billion budget. The European Union (EU), through the EuroHPC Joint Undertaking and its AI Factories program, is investing in continental compute capacity to reduce dependence on American cloud providers. The United Kingdom (UK) has announced sovereign compute investments through its AI Research Resource. Israel maintains a sophisticated AI research ecosystem tied closely to its defense industry. Several Gulf states have, with striking speed, become some of the largest single buyers of AI infrastructure in the world: there, capital is abundant, energy is cheap, and political leadership has framed AI as a defining national priority. Türkiye, working through TÜBİTAK (the Scientific and Technological Research Council) and its defense industry players such as ASELSAN, Baykar, and TUSAŞ, is developing AI capacity adjacent to its existing strengths in autonomous systems and signals processing.

Tier 3: The Compute Peripheries

At the base of the hierarchy sits the great majority of the world’s countries, where compute is not produced, not stockpiled, and often not even reliably available. Much of sub-Saharan Africa, large parts of South and Southeast Asia, most of Central Asia, and considerable portions of Latin America fall into this category. The barriers here are not principally regulatory but material: unreliable electricity, limited connectivity, scarce technical labor, and capital constraints that make hyperscale infrastructure unaffordable. UNDP’s Next Great Divergence report notes that AI usage in many low-income countries remains near 5 percent of the population, against two-thirds or more in high-income economies. [16] The IMF’s AI Preparedness Index shows the same gap in starker form: low-income countries average less than half the readiness score of high-income countries across infrastructure, human capital, innovation, and regulation. [17] These states are negotiating whether they will participate in the AI ecosystem at all. [18]

The mechanisms of compute coercion

A hierarchy is only as real as the tools that hold it in place. If the three-tier map described in the previous section were merely a snapshot of who happens to be ahead today, it would be a curiosity rather than a structural condition. What makes the compute hierarchy consequential is that the actors at its top have developed, and routinely use, a set of instruments that turn their position into leverage. Three mechanisms stand out, and together they explain why Tier 3 states, even wealthy and technically capable ones, find their critical options narrowing rather than expanding. [19]

The first and most visible mechanism is export controls. In October 2022, the U.S. Department of Commerce, through its Bureau of Industry and Security (BIS), issued sweeping new restrictions on the export of advanced AI chips and chipmaking equipment to China. The rules were tightened in October 2023 and again in 2024, closing loopholes that had allowed firms like Nvidia to sell modified, slightly slower chips into the Chinese market. [20] In January 2025, the outgoing Biden administration unveiled what it called the AI Diffusion Rule, a framework that divided the world’s countries into three categories with different levels of permitted reach to American AI chips, placing close allies in a top tier with unrestricted availability, adversaries in a bottom tier with effective denial, and most of the rest of the world in a middle tier with strict licensing requirements and per-country compute caps. [21] The Diffusion Rule was, in important respects, the most explicit acknowledgment yet that the United States now understands compute reach as a tool of foreign policy, comparable to financial sanctions or arms export controls. Although the Trump administration rescinded the rule in May 2025 and replaced it with a country-by-country negotiating model, the underlying logic that chip reach is a sovereign American prerogative to grant or withhold has remained intact across both administrations.

The second mechanism, more recent and less formalized, is compute-bundled diplomacy: the practice of using AI infrastructure availability as a bargaining chip in broader bilateral relationships. The clearest illustration came in May 2025, when President Trump’s regional tour in the Gulf produced a cascade of multi-billion-dollar AI deals involving Nvidia chip allocations, hyperscaler partnerships, and joint sovereign-cloud projects between American firms and several Gulf governments, including the announcement of HUMAIN, Saudi Arabia’s new state-backed AI company. [22] What was striking about these arrangements was their structure. Chip access was bundled with security guarantees, investment commitments, and political alignment in ways that would have been familiar to observers of mid-twentieth-century arms diplomacy. AI infrastructure has become a currency in great-power relationships, and Tier 3 states discover, often quickly, that buying that currency in volume requires paying for it in policy alignment as well as cash.

The third mechanism is more structural than the others and operates without anyone needing to decide on energy and water constraints. Training a frontier AI model consumes enormous quantities of electricity, and running one at scale consumes even more. Worldwide data center electricity demand is projected to rise by roughly 35 percent by 2026, with AI workloads driving most of the increase. [23] Hyperscale data centers also require vast cooling water supplies, often in regions where water is already contested. The states that can absorb this infrastructure burden, which are those with abundant power generation capacity, water resources, or both, are positioned to host the next wave of compute build-out. Those who cannot are, in practical terms, excluded from the production side of AI regardless of their financial resources or political ambitions. Energy abundance has become an unexpected new entry into the calculus of structural power, which is part of why several oil-rich states have moved so quickly to position themselves as hosts of the worldwide AI buildout. Their comparative advantage in cheap energy translates directly into compute capacity, even where chip design and model development remain out of reach.

Strategic autonomy dilemma

There are two issues within the strategic autonomy dilemma.

The first is bandwagoning, aligning closely with one architect, typically the United States, in exchange for guaranteed access to its chips, models, and cloud infrastructure. The benefits of bandwagoning are real. The country receives guaranteed chip allocations, frontier-model access, and the political legitimacy that comes with being treated as a trusted partner by the dominant architect. The cost is equally real. Bandwagoning resolves the access problem by deepening the structural dependence it was meant to address. The bandwagoner gains compute today by surrendering the optionality to develop it differently tomorrow, and by accepting that future access will depend on continued alignment with whatever the architect’s foreign policy priorities turn out to be.

The second strategy is indigenous deployment, investing in domestic capacity across selected layers of the compute stack rather than relying on access from outside. The EU has pursued perhaps the most ambitious version of this through its EuroHPC Joint Undertaking and the recently announced AI Factories program, which aims to provide European researchers and firms with access to publicly funded continental compute capacity. The UK is investing in a sovereign compute resource. Several Gulf states have begun to position themselves as compute hosts, leveraging cheap energy and political stability to attract hyperscale build-outs that, while foreign-owned, sit physically within their territory and create local technical capacity. Türkiye’s approach falls into a related but distinct category through TÜBİTAK and the defense industry ecosystem around firms like ASELSAN, Baykar, and TUSAŞ. The country has developed real strengths in AI-adjacent capabilities—autonomous systems, signals processing, and defense electronics—without yet attempting to compete at the frontier-model or advanced-chip layer.

Policy implications

For middle powers

The first and most important shift is conceptual. States pursuing an AI strategy should abandon the ambition of full-stack autonomy and embrace the more realistic goal of targeted autonomy in selected layers. Frontier model training and leading-edge chip fabrication are, for the foreseeable future, beyond the reach of any Tier 3 actor working alone. [24] What is achievable, and where investment yields disproportionate returns, lies elsewhere in the stack. Sovereign inference cluster facilities that run models rather than train them can be built at a meaningful scale within a Tier 3 state’s budget. Domain-specific models tailored to local languages, legal systems, and policy priorities can compete with general-purpose foreign models in the applications that matter most. Assembly, testing, and packaging capacity in semiconductors, while less glamorous than frontier fabrication, captures real economic value and cultivates the technical workforce needed for any future move upmarket. Data center hosting, where geography and energy permit, converts physical advantages into negotiating leverage with the architects.

For multilateral institutions

The international architecture for AI governance, as currently constituted, is poorly designed for the problem described. The major existing processes—the Bletchley Declaration of November 2023, the G7 Hiroshima AI Process, the United Nations High-Level Advisory Body on AI, and the Seoul and Paris AI Summits that followed—have concentrated their attention on safety risks (existential, misuse, and loss-of-control scenarios) and on the harmonization of regulatory standards. These are reasonable concerns, but they leave the question of who can produce AI in the first place almost entirely untouched. Patricia Gruver-Barr and Gordon LaForge, writing for New America, have argued that the transnational governance conversation needs to recognize compute access itself as a development issue, comparable in structure to vaccine access or climate finance, and they have proposed an “AI Gavi”, a multi-stakeholder coalition modeled on the Gavi vaccine alliance, which has immunized nearly 900 million children in developing countries to assemble the financing, infrastructure, and data ecosystems that Tier 4 states cannot cultivate alone. [25] The proposal deserves serious attention. Even if the specific institutional form is debatable, the underlying logic that compute should be treated, at least in part, as a global public good, not merely as a commercial product or a critical asset, points in the right direction. [26]

Endnotes

1. International Monetary Fund, “AI Preparedness Index Mapping the Readiness of Nations,” IMF Staff Discussion Note SDN/2024/002, 2024.
2. United Nations Development Programme, “The Next Great Divergence Artificial Intelligence and Human Development in Asia and the Pacific,” UNDP Regional Bureau for Asia and the Pacific, 2025.
3. Bhaskar Chakravorti, “How Companies Can Mitigate the Harms of AI-Driven Inequality,” Harvard Business Review, 2025, https//hbsp.harvard.edu/product/H08PLG-PDF-ENG.
4. Pippa Norris, Digital Divide Civic Engagement, Information Poverty, and the Internet Worldwide (Cambridge Cambridge University Press, 2001); Jan van Dijk, The Digital Divide (Cambridge Polity Press, 2020).
5. Kate Crawford, Atlas of AI Power, Politics, and the Planetary Costs of Artificial Intelligence (New Haven Yale University Press, 2021).
6. Susan Strange, States and Markets: An Introduction to International Political Economy (London Pinter Publishers, 1988).
7. Strange, States and Markets for the broader intellectual lineage of this framework, see Benjamin J. Cohen, International Political Economy: An Intellectual History (Princeton University Press, 2008).
8. Ian Bremmer, “The Technopolar Moment How Digital Powers Will Reshape the Global Order,” Foreign Affairs 100, no. 6 (2021) 112–128; Ian Bremmer and Mustafa Suleyman, “The AI Power Paradox Can States Learn to Govern Artificial Intelligence Before It’s Too Late?” Foreign Affairs 102, no. 5 (2023) 26–43.
9. Henry Farrell and Abraham Newman, “Weaponized Interdependence How Global Economic Networks Shape State Coercion,” International Security 44, no. 1 (2019) 42–79 expanded in Henry Farrell and Abraham Newman, Underground Empire How America Weaponized the World Economy (New York: Henry Holt and Company, 2023).
10. Stanford Institute for Human-Centered Artificial Intelligence, “Artificial Intelligence Index Report 2024,” Stanford University, 2024.
11. Chris Miller, Chip War: The Fight for the World’s Most Critical Technology (New York Scribner, 2022).
12. Synergy Research Group, “Cloud Infrastructure Services Market Share Report,” Synergy Research Group Quarterly Reports, 2024.
13. Stanford Institute for Human-Centered Artificial Intelligence, “Artificial Intelligence Index Report 2024.”
14. Gregory C. Allen, “DeepSeek, Huawei, Export Controls, and the Future of the U.S.-China AI Race,” Center for Strategic and International Studies, 2025, https//www.csis.org/analysis/deepseek-huawei-export-controls-and-future-us-china-ai-race.
15. James A. Lewis, Strengthening a Transnational Semiconductor Supply Chain (Washington, DC: Center for Strategic and International Studies, 2022).
16. United Nations Development Programme, “The Next Great Divergence.”
17. International Monetary Fund, “AI Preparedness Index.”
18. Jai Vipra and Sarah Myers West, Computational Power and AI (New York: AI Now Institute, 2023).
19. Lennart Heim and Konstantin Pilz, Compute at Scale: A Broad Investigation into the Data Center Industry (Oxford Centre for the Governance of AI, 2023).
20. Gregory C. Allen, “Choking Off China’s Access to the Future of AI,” Center for Strategic and International Studies, 2022, https//www.csis.org/analysis/choking-chinas-access-future-ai.
21. Center for Strategic and International Studies, “Understanding the Biden Administration’s Updated Export Controls,” CSIS Critical Technologies Program, 2025, https//www.csis.org/analysis/understanding-biden-administrations-updated-export-controls.
22. “Trump Wraps Up Gulf Tour with AI and Energy Deals in UAE,” Reuters, May 16, 2025.
23. International Energy Agency, Electricity 2024 Analysis and Forecast to 2026 (Paris IEA Publications, 2024).
24. Mario Damen, EU Strategic Autonomy 2013–2023 From Concept to Capacity (Brussels European Parliamentary Research Service, European Parliament, 2022), https://www.europarl.europa.eu/thinktank/en/document/EPRS_BRI(2022)733589.
25. Patricia Gruver-Barr and Gordon LaForge, “Minding the AI Power Gap: The Urgency of Equality for Global Governance,” TechPolicy, 2023.
26. Susan Ariel Aaronson, “Data Disquiet Concerns about the Governance of Data for Generative AI,” Center for International Governance Innovation, 2024.