AI Isn't a New Weapon. It's a New Way of Counting Power.

This is the first essay in a series, Technology as Strategic Power: Compute, Capital, and the New Grammar of National Advantage.

States have always kept score. The Greeks counted warships; the empires of the nineteenth century counted soldiers and gold; the superpowers of the twentieth counted missiles, steel and the size of their economies. To be a great power was to have more of the things that could be added up, more territory, more people, more factories, more of the raw material from which armies and influence are made.

For a long time the scorecard held. Each new invention, the railway, the airplane,e even the atom bomb, was simply another line in the same familiar ledger. Power was still counted in land, people, factories and money.

Artificial intelligence, this essay argues, is rewriting the ledger itself.

The conventional view is that AI is a powerful new instrument states must factor in, much as they once had to factor in the tank or the jet. The claim here is more demanding: AI changes what counts as power in the first place.

The inputs that now decide who leads are not land and population. They are four things:

· Compute, the raw calculating capacity needed to build and run advanced models.

· The capacity to fabricate the most advanced chips, the processors on which all of it depends.

· Access to the handful of frontier AI models, the systems only a few laboratories can build.

· A thin layer of talent, the few thousand people able to design and train those systems.

These inputs decide the contest, and they do not map onto the old atlas. A state can be vast and populous yet hold almost none of them. A small state, or a single company, can command a decisive share.

Call this way of seeing the world technological realism.

What the idea is (and is not)

It remains grounded in material reality. Power still flows from real, countable things, and their distribution still shapes the system. That much of the older thinking survives.

What has changed is which things matter, and, more importantly, how they come to be held.

Here is the point on which the rest of the series rests. These new sources of power are not lying about to be seized, as oilfields once were. They are constructed, deliberately assembled from capital, chips, energy, models and talent, wired together into a working whole. Power in the AI age is less a stock to be captured than an ecosystem to be built. And because building the full stack is so costly and so demanding, the capacity to do it concentrates in remarkably few hands. Concentration, in other words, is a consequence of construction, not an accident of geology.

One clarification. This is not the fashionable notion that power has dissolved into ideas and image. Quite the reverse. Technological realism holds that power is more physical than ever: bound up in silicon, in vast fabrication plants, in the electricity that feeds them and in a thin pool of talent. It is a harder-nosed view than the one it replaces, not a softer one.

Why the old scorecard fails

The classical method of ranking states, associated with Cold War thinkers such as Kenneth Waltz, insisted that no single ingredient of power could be judged in isolation. A state’s rank came from the whole basket at once: population, territory, resources, economy, military, competence, stability.^[1] John Mearsheimer put it plainly: power is ‘the currency of great-power politics … What money is to economics, power is to international relations’, and it rests on a state’s wealth and the size of its population.^[2]

Technological realism keeps the temper of this. Power is still real, material and countable; the list is what changes. And it breaks the old prohibition on singling out one ingredient. One specific chain (compute, fabrication, frontier models and talent) now governs economic and military advantage simultaneously. To refuse to isolate that chain is to miss the central fact about it: that it is built, and built by few.

The strongest rival account

Serious scholars already study how AI shifts the balance of power. The most persuasive of them, Michael Horowitz, argues that AI resembles electricity or the internal-combustion engine more than any single super-weapon, a general-purpose technology whose payoff depends on how widely it diffuses and how well states adopt it, rather than on who invents it first.^[3]

This is right, at the level of use. But it accepts the old scorecard and asks how AI moves the existing numbers. The argument here comes one step earlier. AI does not merely reshuffle the rankings; it changes what is being ranked. Skill at adoption decides who makes the most of the technology. The capacity to build the stack decides who reaches the frontier at all. That is the deeper contest, and it is settled not by diffusion but by construction.

What power now looks like, four cases

Theory is one thing. What makes it concrete is watching capital move, over the past two years, in precisely the pattern the theory predicts, and watching each mover assemble not a single asset but an interlocking system.

Chips: capital pooling where the stack is hardest to build.
In July 2026 TSMC, the Taiwanese firm that fabricates the world’s most advanced chips for the likes of Nvidia and Apple, said it would add $100bn to its Arizona plants, taking its American commitment to $265bn.^[4] The sum is less striking than what it buys: not a factory but a cluster, fabrication lines, advanced packaging, suppliers and a trained workforce, co-located so that each part makes the others more valuable. A rival cannot simply match the cheque; it would have to reproduce the whole cluster, and the handful of firms able to attempt that is exactly why capital and governments keep converging on the same names.

Energy: rebuilding the physical base beneath the models.
Compute runs on electricity, and the appetite is enormous. Global data-centre demand is on course to double within a couple of years, to roughly the entire consumption of Japan.^[5] The strain is such that Microsoft has signed a 20-year agreement, worth some $16bn, to restart the Three Mile Island nuclear plant in Pennsylvania, the site of America’s most notorious nuclear accident, reserving every watt for its AI data centres.^[6] Read the move carefully: a software company is now in the business of running a nuclear reactor, because the model is only ever as strong as the electricity behind it. When the frontier of AI depends on a decommissioned power station in Pennsylvania, the boundary between a technology firm and a piece of national infrastructure has quietly dissolved.

A compute node priced like a national programme.
In early 2025 OpenAI, Oracle and SoftBank announced ‘Stargate’, a plan to spend up to $500bn over four years on AI data centres in America.^[7] What that half-trillion assembles is telling: chips from Nvidia, capital from SoftBank and the Emirati fund MGX, cloud operations from Oracle, models from OpenAI, several layers of the stack bound into one venture. No single company owns the result; it takes a coalition to build something on this scale, and a coalition of this calibre is not a thing many places on earth could convene.

A Gulf state building an ecosystem from scratch.
The clearest illustration is the United Arab Emirates, because it shows the stack being assembled as a matter of state strategy. Abu Dhabi’s sovereign fund, Mubadala, joined the Emirati AI group G42 to launch an investment vehicle, MGX, in 2024; by mid-2026 it had closed a fund of roughly $49bn aimed wholly at AI and taken stakes in OpenAI, Anthropic and Elon Musk’s xAI at once.^[8] The state then went further, agreeing to invest $1.4trn in America over a decade in return for the right to import up to 500,000 of Nvidia’s most advanced chips a year, and building a five-gigawatt AI campus in Abu Dhabi.^[9] Each Emirati institution supplies a different layer, ADNOC the energy, Mubadala and MGX the capital, G42 the compute, the state the chip access. A country with a small population and no chip industry of its own is converting the one input it holds in abundance, capital, into a complete ecosystem, and thereby into a seat in a club it could never have entered on the old scorecard. Where an earlier age would have sought to seize an oilfield, the Emirates are quietly assembling every link in a chain instead, and buying their way into the top table by building it.

The through-line is the same in every case. What is being built is never one asset but a system, capital wired to chips wired to power wired to models wired to talent, and the systems are so demanding to assemble that only a few actors can attempt them.

And the state has noticed.
This is not merely an analyst’s framing. It is now the operating doctrine of the world’s most powerful defence establishment. The Pentagon’s Office of Strategic Capital, created in 2022 to steer private money into technologies deemed vital to national security, organises its entire investment strategy around the same logic this essay describes.^[10] Its central aim in the near term is, in its own words, ‘capturing chokepoints in economic networks’: the points where a rival’s control of a single refiner, component or licence can shut down another country’s production line at will.^[11] It treats global capital markets as ‘contested spaces for competitive advantage in national security’, and its stated purpose is to help build a domestic and allied industrial ecosystem ‘from seed to GDP’, the deliberate construction of a stack, financed by the state, precisely because the market alone will not concentrate capital fast enough where security demands it.^[12] When the Department of Defense starts behaving like a venture investor mapping chokepoints, the argument that power now runs through constructed technological ecosystems has stopped being a theory and become policy.

Set against the old arithmetic

On the old scorecard power was concentrated, but not extraordinarily so. Global defense spending reached about $2.63trn in 2025, with America at roughly a third and a long tail of states with real militaries behind it.^[13]

The new scorecard is different in kind. The most advanced chips are made by three firms; the machines that make those chips come from one, the Dutch company ASML.^[14] The frontier models come from a handful of laboratories, led by the United States and China. The elite talent gathers in essentially two national ecosystems. This is not ‘one power holds a third and the rest divide the remainder’. It is ‘almost everything runs through a few doors, and most of the world stands outside them’, because most of the world cannot build what lies behind those doors.

Why it concentrates, the mechanics

This concentration will not dissolve as the technology spreads, because it is a property of how the stack is built. Three forces drive it.

Rising marginal cost. Each increment of model capability demands disproportionately more compute. The cost of the largest training runs has grown about 2.4-fold a year, and single runs are projected to pass $1bn by 2027.^[15] As the price of admission compounds, the field of those who can build at the frontier narrows.

Fabrication economics. Advanced chips are among the hardest objects humanity manufactures. A single leading-edge line costs $15bn-25bn; a full campus exceeds $40bn.^[16] Only a few firms can amortise such outlays across enough volume to survive, hence three advanced fabricators, and a single supplier, ASML, of the lithography machines that pattern the chips, at up to $400m apiece.^[17] Each link narrows to a few players, and the bottlenecks compound. No previous resource had a supply chain shaped quite this way.

Increasing returns. More compute yields a better model, which wins more users and revenue, which buys more compute. Oil never made its owner better at finding oil; compute makes its owner better at acquiring compute. The lead widens rather than closes, which is why the advantage, once assembled, is so hard to dislodge.

Why this exceeds any single industry

Here is what makes it a question of power rather than of commerce. Oil drove engines; compute drives cognition, and cognition feeds the economy and the military alike. Susan Strange drew a useful distinction decades ago: the deepest power is not the ability to compel a particular act but the ability to shape the structures within which others must operate, setting the terms so that one’s preferences prevail without open coercion.^[18]

Compute is structural power of exactly this kind, and it fuses several forms of control into one physical stack. Whoever commands the AI frontier shapes what others can know, what they can build, the financial terms on which they build it and the military balance besides, all through control of the same chips, models and electricity. And unlike the diffuse ‘soft power’ of reputation, this power has an address. It resides in particular plants and machines, in specific data centres and the power stations that feed them. It can be located on a map, and so defended or denied.

The obvious objections

A claim this large deserves scrutiny. Three objections are worth taking seriously.

Will it not diffuse, as technologies do? In part, and genuinely so: cheap, capable AI is spreading fast, and the gap between the leading American and Chinese models has all but closed.^[19] But there is a difference between yesterday’s frontier becoming commonplace, which happens, and the frontier itself dispersing: the capacity to build the single most capable system, and the fabrication and talent behind it, is not spreading, and may in fact be tightening. Both can be true together, and are.

Is this not technological determinism, ignoring human choice? No. The largest stock of compute guarantees success no more than the largest army once did; states can squander a commanding lead through poor judgment. The claim is narrower: the distribution of these constructed resources sets the board and defines which moves are available to whom. It does not play the game.

What if a passing moment is being mistaken for an epoch? This is the fairest challenge, and the one I find hardest to dismiss; having spent long enough inside this argument, I know how easily one can mistake the thing one is studying for the axis on which everything turns. So the conditions should be stated plainly. The argument holds while three things remain true: that capability keeps scaling with compute; that the chokepoints (fabrication, frontier laboratories, elite talent) stay concentrated; and that AI remains the pivotal technology of the age. Should any fail, should progress stall, the bottlenecks open, or some other constraint such as energy or raw materials become decisive, the scorecard would change again. These are wagers, not certainties. A thesis that could never be falsified would be worth little.

What it amounts to

The old order counted soldiers, wealth, land and people, and read a state’s rank from the totals. The new order counts compute, fabrication, frontier-model access and scarce talent, and reads the emerging hierarchy from those. They cluster more tightly than any resource before them. They are built rather than found; leaders pull further ahead the more they invest; and, unlike oil, they bear on the whole of national life at once, economic and military together.

The map of the twenty-first century is not a patchwork of roughly equal great powers. It is a few intense hubs, the fabrication plants of Taiwan and South Korea, the laboratories and data centres of the United States, the lone machine-maker in the Netherlands, the fast-rising capacity of China, set amid a wide periphery that depends on what those hubs produce. And the hubs exist because someone chose to build them.

The essays that follow trace the consequences: how the present compares with the age of oil, how money itself became an instrument of statecraft, what a state must assemble to own a hub, and what raw materials the whole edifice ultimately rests upon.

It begins here, with a single proposition: power now runs through the AI stack; that stack is constructed rather than seized; and the capacity to construct it belongs to very few.

The old realists taught states to count their power. The task of this century is to work out what is worth counting now, and to grasp that it must be built before it can be counted.

Next in the series: From Oil to Algorithms, Compute as the New Strategic Resource.

If this way of thinking is useful to you, subscribe to receive the rest of the series as it appears.

Sources

Notes

1. 1Kenneth N. Waltz, Theory of International Politics (Reading, MA: Addison-Wesley, 1979), esp. pp. 97–98, 131 (a state’s rank rests on the whole basket of capabilities, which cannot be judged one ingredient at a time). Chapters 5–8 (PDF).
2. 2John J. Mearsheimer, The Tragedy of Great Power Politics (New York: W. W. Norton, 2001), ch. 1, p. 12, and ch. 3, pp. 55–61 (’Power is the currency of great-power politics’; latent power rests on wealth and population). Chapters 1–2 (PDF).
3. 3Michael C. Horowitz, ‘Artificial Intelligence, International Competition, and the Balance of Power’, Texas National Security Review 1, no. 3 (May 2018), pp. 37–57 (AI as a general-purpose technology whose effect turns on diffusion and adoption). Full text. A related framework for when an asset becomes ‘strategic’ is Jeffrey Ding and Allan Dafoe, ‘The Logic of Strategic Assets: From Oil to AI’, Security Studies 30, no. 2 (2021), pp. 182–212. DOI.
4. 4On TSMC’s additional $100 billion investment in Arizona (July 2026), bringing its total US commitment to $265 billion for advanced 2nm-class fabrication, see U.S. Department of Commerce, ‘Trump Administration Secures an Additional $100 Billion U.S. Semiconductor Investment’, 16 July 2026; and The New York Times, ‘TSMC Adds $100 Billion to Its U.S. Spending Plan’, 16 July 2026.
5. 5On global data-centre electricity demand on course to more than double, from about 460 TWh in 2024 to over 1,000 TWh by 2026 (roughly Japan’s total consumption), see the International Energy Agency’s figures as reported in Informed Clearly, ‘AI’s Energy Hunger: Data Center Boom Breaks Global Power Records’, 6 July 2026; and S&P Global Market Intelligence, ‘AI Power Demand’, 29 June 2026.
6. 6On Microsoft’s 20-year power-purchase agreement (roughly $16 billion) to restart Three Mile Island Unit 1, renamed the Crane Clean Energy Center, 835 MW, targeted for 2027, with 100 per cent of output reserved for Microsoft’s data centres, see DatacenterDynamics, ‘Three Mile Island nuclear power plant to return as Microsoft signs 20-year 835MW AI data center PPA’, 9 July 2026; and SMR Intel, ‘Every Nuclear Data Center Deal’, 6 July 2026.
7. 7On the Stargate project, up to $500 billion over four years from OpenAI, Oracle and SoftBank (with backing from Nvidia and the UAE fund MGX) to build US AI data-centre and power infrastructure, announced January 2025, see DatacenterDynamics, ‘OpenAI announces “The Stargate Project”: $500bn over four years on AI infrastructure’.
8. 8On MGX, launched in March 2024 by Mubadala Investment Company and G42, which closed a roughly $49 billion fund (July 2026) aimed entirely at AI and holds stakes in OpenAI, Anthropic and xAI, see Crashbytes, ‘The Common Shareholder: MGX’s $49B Fund’, 3 July 2026; and Startup Fortune, ‘Abu Dhabi’s MGX has quietly become the most consequential AI investor on the planet’, 23 June 2026.
9. 9On the May 2025 US–UAE agreement, the UAE committing to invest $1.4 trillion in the US economy over a decade in return for the right to import up to 500,000 of Nvidia’s most advanced chips a year, anchored by a five-gigawatt AI campus in Abu Dhabi, see the Washington Examiner op-ed by the UAE’s ambassador, ‘America holds the chips. The UAE intends to run the region’s AI’, 8 July 2026.
10. 10United States Department of Defense, FY2025 Investment Strategy for the Office of Strategic Capital (Washington, DC: Department of Defense, 2025), pp. 1–6 (the Office of Strategic Capital was established in December 2022 to ‘attract and scale private capital’ into technologies critical to national security).
11. 11DoD, FY2025 Investment Strategy, p. 6 (the near-term arena, 0–3 years, is ‘Capturing Chokepoints in Economic Networks’; a competitor’s control of a single refiner, component or export licence can be used to ‘shut down specific production lines’ at a reliant company).
12. 12DoD, FY2025 Investment Strategy, pp. 5–8 (describing ‘global capital markets [that] have become contested spaces for competitive advantage in national security’, and OSC’s mandate to scale allied production and critical-technology firms from ‘seed to GDP’).
13. 13On global defence spending of about $2.63 trillion in 2025, with the United States at roughly a third, see the International Institute for Strategic Studies, The Military Balance 2026, as reported by the UK Defence Journal, 24 February 2026. Totals vary by methodology; the US share of roughly a third holds across measures.
14. 14On the three leading-edge chipmakers (TSMC, Samsung, Intel) at the 2nm class and ASML as the sole maker of the EUV lithography machines needed to produce them, see Tom’s Hardware, ‘Leading-edge foundry roadmaps for TSMC, Intel and Samsung’, 14 May 2026; and CNBC, ‘Why Nvidia’s AI boom couldn’t happen without Dutch chip equipment maker ASML’, 29 January 2026.
15. 15On the cost of the largest AI training runs growing roughly 2.4× per year since 2016, with single runs projected to exceed $1 billion by 2027, see Ben Cottier and others, ‘The rising costs of training frontier AI models’, arXiv:2405.21015 (2024). Paper.
16. 16On a single leading-edge chip fabrication line costing roughly $15–25 billion (campuses exceeding $40 billion), see SemiconductorX, ‘Leading-Edge Logic Fabs’; and Tom’s Hardware, ‘Firm predicts it will cost $28 billion to build a 2nm fab’, 22 December 2023.
17. 17On ASML as the sole supplier of EUV lithography (current machines ~$200 million; next-generation High-NA ~$400 million each), see CNBC, ‘Why Nvidia’s AI boom couldn’t happen without ASML’, 29 January 2026; and Tech Times, ‘Even TSMC Says ASML’s Newest Machine Is Too Expensive’, 11 June 2026.
18. 18Susan Strange, States and Markets (London: Pinter, 1988), pp. 24–26 (the distinction between relational power and ‘structural power’, the power to shape the structures within which others must operate). Full text at the Internet Archive.
19. 19On the narrowing gap between the top US and Chinese AI models (from double digits in 2023 to near parity by early 2026), see Stanford HAI’s AI Index 2026, summarised by the United Nations University, ‘What the 2026 Stanford AI Index Report Tells Us’, 1 July 2026.