Techno-Polarity in the Twenty-First Century

For three years the contest for artificial-intelligence supremacy has been reported as a leaderboard: whichever laboratory ships the most capable frontier model is presumed to be winning. That framing is now actively misleading. By March 2026 the performance gap between the best American and best Chinese models had narrowed to under three percentage points — even as American private capital outspent Chinese private capital by more than twenty-to-one.¹ If money bought the lead, the lead would not be closing. The leaderboard has stopped measuring what matters.

This study operationalises AI leadership as a multi-dimensional national capability spanning ten indicators — compute, semiconductors, frontier models, private-sector strength, capital, talent, data and deployment, military integration, state capacity, and standards influence — and scores fourteen actors against them in a transparent AI National Power Index. The Index finds a genuine bipolar top tier (the United States at 88, China at 73 on a 100-point scale) sitting above a tightly compressed field of middle powers, none of which independently exceeds the low fifties. But the headline ranking is the least interesting result. The decisive finding is where American advantage actually lives — and how narrow, physical and contestable that foundation is.

American power in AI rests overwhelmingly on control of the compute supply chain — data-centre capacity, Nvidia's accelerator ecosystem, and a lithography chokepoint held by ASML and TSMC that neither Washington nor Beijing owns. This is a formidable but brittle advantage: it is geographically concentrated, commercially rather than sovereignly held, and it is being eroded from three directions at once — Chinese domestic substitution through Huawei's Ascend line,⁷ the diffusion of near-frontier open models that make capability cheap to copy, and an 89% collapse since 2017 in the inflow of AI researchers to the United States.¹ China, conversely, leads on the inputs that compound over time: talent pipeline, patents, publications, robotics, energy and state coordination.

The central recommendation follows directly. Actors should stop competing to top the model leaderboard — a lead that is real but decreasingly convertible into durable power — and compete instead to govern the chokepoints and shape the standards through which any model must pass to become deployable capability. For the leading power that means hardening the compute advantage while it lasts and converting it into rule-setting leverage; for middle powers it means specialising into indispensable niches rather than chasing frontier parity they cannot afford. Critically, the current equilibrium is a textbook Prisoner's Dilemma: every actor's dominant move is to accelerate, yet mutual acceleration degrades the safety and stability all of them value. Whether institutions can alter those payoffs — through verification, compute-based transparency and conditional reciprocity — is the question on which the character of the next international order turns.

2.1Situation — the world as strategists have understood it

Since the release of large language models into public use, the prevailing mental model among policymakers has been reassuringly familiar. Artificial intelligence is treated as a strategic technology in the lineage of the nuclear weapon, the space launcher and the microprocessor: a capability where a clear leader emerges, holds a measurable edge, and translates that edge into economic and military advantage over rivals who scramble to catch up. On this model the United States is the incumbent leader, China the fast follower, and the task of statecraft is to preserve the American lead through investment at home and export controls abroad.

The comfort of this picture lies in its legibility. It tells strategists what to measure (model capability), who is ahead (the country with the best models), and what to do (spend more, restrict more).

2.2Complication — the leaderboard has stopped measuring power

Three developments, each independently documented, together dismantle the incumbent-leader model.

One — capability is converging while spending diverges

Stanford's 2026 AI Index records that the model-quality gap between the leading American and Chinese systems collapsed from roughly 17–32 percentage points in mid-2023 to 2.7 points by March 2026 — despite the United States investing some US$285.9bn in private AI capital in 2025 against China's US$12.4bn.¹ Two trends are visible in that single comparison, and they point in opposite directions. One of them is sustainable for the leader; the other is not.

Two — the lead is decreasingly convertible

A three-month technical lead in frontier models — the current American margin, by several assessments — matters enormously for the laboratories competing for it and remarkably little for the militaries and economies meant to benefit. Military AI adoption runs on multi-year procurement cycles; a lead measured in weeks does not survive contact with an acquisition timeline. When near-frontier capability is available as open-weight models that any competent actor can fine-tune, the strategic question shifts from who can build the best model to who can deploy capability at scale, cheaply, and under their own control. On that question the leaderboard is silent.

Three — the human foundation is eroding

The number of AI researchers migrating to the United States has fallen 89% since 2017, with 80% of that decline in the last year alone. The pipeline that built American pre-eminence is drying while China leads the world in AI publications at 23% of global output, in patent grants at nearly 70%, and in industrial-robot installations.¹ The leader's advantage is increasingly an inheritance rather than an income.

3.1The advantage is compute, and compute is geography

Roughly 90% of tracked global AI supercomputing capacity sits in just two countries, with the United States holding an estimated 74.5% and China 14.1%.² The asymmetry in built infrastructure is starker still: the United States hosts more than 4,000 data centres — more than the next ten countries combined — and its dedicated AI data-centre capacity is forecast to more than double from roughly 8GW in 2026 toward 21GW by 2031, over twice China's projected 10GW.³ Even were export controls fully relaxed, independent estimates put the American advantage in compute produced in 2026 at between 21× and 49× China's, depending on the performance metric.

The strategic implication is precise: because compute is a physical asset concentrated in known locations and supplied through identifiable firms, it is the single most governable dimension of AI power. Unlike talent or algorithms, it can be counted, taxed, restricted and — in principle — verified. This is why the chokepoints matter so much more than the models.

3.2A single Dutch firm and a single Taiwanese foundry hold the master keys

Every advanced AI accelerator on Earth is fabricated on machines made by one company. ASML of the Netherlands supplies essentially all extreme-ultraviolet (EUV) lithography, the only viable path to chips below 7nm. It produces roughly 40 such machines a year, each costing up to US$400m, and no credible competitor exists after Nikon abandoned the effort a decade ago.⁴ Those machines feed overwhelmingly into TSMC of Taiwan, whose share of the global foundry market rose to roughly 73% and which controls over half of the advanced CoWoS packaging on which AI chips depend.⁵

This reframes export controls entirely. They are not simply an American instrument; they are a multilateral one whose force depends on Dutch, Taiwanese, Japanese and South Korean alignment.⁶ It also locates the single greatest systemic risk in the entire landscape: any disruption to Taiwan would not hand advantage to one side — it would remove the foundation beneath both.

3.3Divided strengths, not a simple hierarchy

Decomposing capability by dimension reveals that the US–China contest is not a race along one track but a divergence across ten. The United States leads decisively on compute, frontier models, private-sector capability and capital. China leads on data and deployment scale, government coordination and state capacity, and runs close on talent and military integration. These strengths are not interchangeable: American capital cannot quickly manufacture Chinese state coordination, and Chinese talent depth cannot quickly buy American compute.

3.4The investment paradox and the talent counter-current

The 23-to-1 private-investment gap — US$285.9bn against US$12.4bn in 2025 — coexists with a 2.7% capability gap. Part of the explanation is that China's true spending is understated, because government-guidance funds sit outside the private-investment tally. But part is genuinely structural: capital is subject to steeply diminishing returns once compute, data and talent are adequate, and China has been converting far less capital into comparable capability. Meanwhile the input that does not obey diminishing returns — human expertise — is moving against the incumbent. The 89% collapse in researcher inflows to the United States since 2017, alongside South Korea's world-leading AI patents per capita and China's dominance of publications and patents, describes a slow erosion of the very foundation the spending is meant to defend.¹

3.5Military advantage runs on mass and integration, not model quality

The United States leads in AI model sophistication and in the integration of autonomy under human-judgement doctrine, through DoD Directive 3000.09 and the Replicator initiative, with a FY2026 autonomy line item of roughly US$13.4bn.⁸ China leads in the industrial capacity that intelligentised warfare actually consumes: in January 2026 a PLA institution demonstrated one operator supervising 200 autonomous drones, and Chinese manufacturers are estimated to produce over 4,000 combat-capable drones a day. Because military AI adoption runs on multi-year cycles, a three-month model lead confers little battlefield advantage; the contest is over which force best harnesses AI at scale, where China's manufacturing dominance is a structural asset the leaderboard does not capture. Russia, foundationally behind, demonstrates a third path — cheap, attritable, battle-tested autonomy refined in Ukraine — proving that catastrophic capability does not require frontier leadership.

3.6Middle powers and firms are structural actors, not spectators

The landscape contains indispensable specialists whose leverage exceeds their overall ranking: Taiwan in fabrication, the Netherlands in lithography, the European Union in regulation and standard-setting — its Cloud and AI Development Act of June 2026 turns digital sovereignty into procurement law⁹ — the Gulf states in capital and energy, with Saudi Arabia's framework exceeding US$40bn and UAE–France and UAE–Italy sovereign compute projects underway, India as a talent and deployment power that hosted the first Global-South AI summit in February 2026,¹⁰ and Israel, South Korea and Japan as specialised technology powers. Firms — Nvidia, ASML, TSMC, the hyperscalers — command infrastructure and set de-facto standards in ways that give them independent geopolitical weight. Any model of AI power that stops at two states misreads the board.

The Index scores each actor from 0–100 on ten weighted dimensions, then aggregates to a single composite. Weights reflect each dimension's contribution to durable, convertible power: compute and semiconductors are weighted most heavily because they are the binding constraints and the hardest to substitute; standards influence is weighted lowest because it shapes the environment rather than generating capability directly.

Two features of the ranking carry the argument. First, the bipolar gap is real but the ordering within the middle is close: from the European Union at 50.6 down to Taiwan at 41.5, nine actors sit within just over nine points, meaning small shifts in weighting or evidence reorder them — a feature, not a defect, reflecting genuine multipolar compression below the top two.

Second, and more important, the ranking hides the leverage. Taiwan ranks eleventh overall yet holds, in fabrication, a more decisive chokepoint than anything in the American arsenal; the European Union ranks third largely on a single dimension — standards — where it scores 90. A composite score measures breadth of capability; it does not measure indispensability. Both matter, and conflating them is precisely the error the leaderboard encourages.

5.1The strategic core — why the system tends toward acceleration

Beneath every scenario lies one interaction. Two leading actors each choose between Accelerate — race capability, minimise restraint — and Restrain — invest in safety, accept transparency. Mutual restraint yields the highest joint payoff, a stable and safer trajectory both value. But whatever the other does, each is individually better off accelerating: if the rival restrains, accelerating buys advantage; if the rival accelerates, restraining means falling behind. The dominant strategy is to accelerate, and the equilibrium is mutual acceleration — a worse outcome for both than mutual restraint.

The value of the formalism is that it tells us exactly what an intervention must do to change the outcome: it must alter the payoffs so that restraint becomes individually rational. Four mechanisms can do so.

6.1Compete for the chokepoints, not the leaderboard

Resolves §3.1–3.2 and the Complication. Because the frontier lead is real but decreasingly convertible while compute and lithography are decisive and governable, strategic effort should shift from subsidising model capability toward securing, and setting the rules for, the compute supply chain. For the leading power: harden the compute advantage while it lasts, and convert it into standard-setting leverage before substitution erodes it. For middle powers: identify and defend an indispensable niche rather than chase frontier parity.

6.2Treat AI power as a coalition, and invest in the alliance that owns the bottlenecks

Resolves §3.2. Since the master keys sit in Veldhoven and Hsinchu, not Washington or Beijing, the leading power's advantage is a coalition property. The priority is durable alignment with the Netherlands, Taiwan, Japan and South Korea on export governance — and a serious, funded contingency for the systemic risk a Taiwan disruption poses to both superpowers. Middle powers holding chokepoints should recognise and price their leverage rather than surrender it cheaply.

6.3Build the verification and transparency regime that changes the payoffs

Resolves the §5.1 dilemma. The acceleration equilibrium is a Prisoner's Dilemma only while restraint is unverifiable. Because compute is countable, AI is unusually amenable to verification: compute-based transparency, third-party audits and reciprocal disclosure can convert the dilemma into a Stag Hunt in which cooperation is stable. This is the highest-leverage institutional investment available, and the one middle powers and international bodies are best placed to lead — precisely because they are not the incumbents.

6.4Compete on deployment and standards, where advantage compounds

Resolves §3.3 and §3.5–3.6. Because strengths are non-substitutable and military and economic advantage flow from harnessing and deploying AI at scale rather than from model quality, actors should prioritise deployment capacity, adoption and standard-setting. The European Union's regulatory reach, India's deployment scale and Global-South convening power, and the Gulf's capital-and-energy position are all more durable bases of influence than a frontier-model programme they cannot sustain.

6.5Defend the human foundation

Resolves §3.4. The 89% collapse in researcher inflows is a slow-moving strategic wound for the incumbent. Talent, unlike capital, does not obey diminishing returns. Immigration, research funding and open-science policy are national-security instruments in this contest, and should be resourced as such.

7.1Risk mitigation — pre-empting what goes wrong

Method

This study uses a mixed-methods architecture: quantitative indicators — compute shares, investment, research output, capacity forecasts — drawn from public datasets; comparative and historical analysis; scenario planning; and game-theoretic modelling. The AI National Power Index aggregates ten analyst-scored dimensions on a 0–100 scale under transparent weights summing to 100%. Scores are evidence-anchored judgements, not direct measurements; the weights encode a view that compute and semiconductors are the least-substitutable sources of durable power.

Limitations