Temporarily impossible
Everyone is racing, everyone is spending, and yet the integrated supply chain everyone depends on keeps running because nobody can fully replace what the others provide.
That paradox has a name. It is the Prisoner's Dilemma — and it sits at the heart of why no single country can currently build general-purpose AI alone, even though the world collectively has everything required to do it.
GPAI #
Before we get to the dilemma, it is worth being precise about what we are actually talking about.
Most AI today is narrow. It does a specific task well — generating text, classifying images, recommending content, translating language. The economic value is real. The strategic significance is growing. But narrow AI is not what the race is fundamentally about.
The prize everyone is racing toward is general-purpose AI: systems capable of performing most cognitive tasks humans can perform, across domains, without being specifically trained for each one. Systems that can do science, engineering, strategy, and economic planning — not one of those things, but all of them, simultaneously, at superhuman speed.
The reason that prize matters so much is compounding. A country or company that deploys genuinely general-purpose AI gets productivity gains across its entire economy. Those gains fund further development. That development produces better AI. Which produces more gains. The advantage compounds in ways that most economic advantages do not.
This is why the first-mover calculus is so aggressive. The prize is in a competitive market. It is a structural advantage that, if consolidated, may be very hard for anyone else to close.
The collective capability #
Here is the thing that should be obvious but rarely gets stated plainly: the world, right now, has everything it needs to build general-purpose AI.
The chips exist. The data exists. The models are progressing. The energy infrastructure is being built. The talent is distributed across multiple countries. The materials and manufacturing base are in place. Assembling all of it in one place, with one coordinated effort, is not a fundamental technological challenge. The knowledge is out there.
The problem is that "the world" is not a single actor.
The capabilities are distributed across countries that do not trust each other, are competing with each other, and are actively trying to deny each other access to the most critical pieces. The world collectively can build GPAI. No bloc currently can do it entirely from its own controlled resources. That is the gap.
The quality problem #
It is tempting to think this is purely a supply chain problem. Get the chips. Build the fab. Train the model. Problem solved. But the gap between blocs is not just a hardware gap. There is a model quality gap — and it matters more than it looks.
The leading foundation models are trained by a small number of organisations in the US: OpenAI, Anthropic, Google DeepMind, Meta AI. They have accumulated years of training runs, alignment research, evaluation frameworks, and iteration cycles that are not easily replicated by starting fresh. The gap between the frontier and the second tier is about having done the work.
Chinese AI companies — Baidu, Alibaba, Zhipu, ByteDance — have capable models. They are behind the US frontier on most rigorous benchmarks. The restriction on the most advanced NVIDIA chips is one factor. The harder factor to close is the accumulated learning gap: the years of training, feedback, and iteration that the leading US labs have and everyone else is trying to replicate.
This creates a quality gap that sits on top of the hardware gap. Even if China solved its chip access problem tomorrow, it would still face several years of catching up on model quality — assuming the US frontier does not continue advancing during that time, which it will.
The EU has no frontier model provider at scale. It has regulation, data-protection frameworks, and research institutions. What it does not have is anything equivalent to OpenAI, Anthropic, or DeepMind. European companies and governments consuming AI are, almost entirely, consuming American AI.
The Prisoner's Dilemma #
Imagine two blocs — call them A and B — that collectively have the full supply chain for GPAI. A controls chips and models. B controls materials and manufacturing scale. Together they could build GPAI faster and better than either can alone.
The payoff matrix looks roughly like this:
B cooperates B defects
A cooperates Both get GPAI B wins alone,
faster, share A gets nothing
the benefits
A defects A wins alone, Neither gets GPAI quickly.
B gets nothing Both waste resources.
Race continues.
The dominant strategy for both, if they do not trust each other, is to defect. To restrict. To build domestic alternatives. To try to win alone rather than risk being the one who cooperated while the other defected and took everything.
Both sides defecting produces the worst collective outcome. It is slower. It is more expensive. It may produce AI that is less capable than coordinated development would achieve. But neither side will commit to cooperation because the downside of being the cooperating party in a world where the other defects is unacceptable.
This is not a new problem. It is the same logic that has driven arms races throughout history. What makes this version distinctive is the prize: not territorial control, not nuclear deterrence, but cognitive infrastructure that could reshape the entire global economy.
The time window argument #
The US lead in model quality and chip architecture is real, but it is not fixed. The leading labs are advancing, but so is everyone else — just more slowly. China's chip designers are making progress under constraints. Chinese models are improving. Alternative supply chains for rare materials and equipment are being developed in the West. TSMC is building outside Taiwan.
The window in which the current configuration — where the US controls the leading chips and models, and China cannot replicate them — is finite. Nobody knows how wide that window is. Estimates range from a few years to a decade.
The strategic choices being made right now are bets on that window. The US export control strategy is a bet that the window is wide enough to consolidate advantage before China closes the gap. China's investment strategy is a bet that it can close the gap faster than the restrictions can prevent it. Both cannot be right.
What makes this interesting is that the window is not a fixed physical quantity. It can be expanded by pushing the frontier faster — investing more in model research, chip architecture, and energy infrastructure. It can be narrowed by the other side advancing. It can be collapsed by a geopolitical event that disrupts the supply chain. It is a moving target under active competition.
Calculation change #
Three things could shift the current stalemate in meaningful ways.
A breakthrough in chip alternatives. If China develops a manufacturable path to leading-edge chips without EUV — through new architectures, chiplet stacking, or a genuine improvement in DUV-based processes — the hardware gap closes faster than expected. This is the outcome the US export controls are designed to prevent. It is not impossible.
A step-change in model efficiency. If a new training architecture emerges that produces frontier-quality models on significantly less compute, the chip access gap becomes less decisive. Some researchers argue that the field is approaching algorithmic improvements that could partially substitute for raw hardware. If true, restrictions on the best chips become less powerful as restrictions.
A supply chain disruption. A conflict over Taiwan, a natural disaster affecting a key fab, a political decision that severs a critical connection. The scenario that has been discussed in policy circles for years without happening. The supply chain has proven more resilient than many predicted — but its concentration in Taiwan, in particular, remains a systemic risk that has not been resolved.
The uncomfortable arithmetic #
Given all of this, where does each bloc actually stand on the path to general-purpose AI?
The US position is the strongest — but not as dominant as the political rhetoric suggests. It controls the frontier of model development and chip architecture. It depends on allies for fabrication, optics, and materials. Its lead in model quality is real but contested. The CHIPS Act is moving slowly. The talent pipeline remains partially dependent on global immigration.
China's position is constrained but not desperate. It has the industrial scale, the domestic market, the state resources and the engineering talent to be a serious competitor. What it lacks is the leading-edge chip access that would let it compete at the frontier of training runs. That constraint is real today. It may not be real in five years.
Europe's position is the most uncomfortable of the three. It controls the most irreplaceable components — ASML, ZEISS, TRUMPF — but converts none of that leverage into AI capability. European companies and governments consume AI made elsewhere. The EU Chips Act is attempting to address the fabrication gap, but the model development gap is not being addressed at scale.
Nobody is close to GPAI on any realistic assessment. The timelines are uncertain. The capability required is probably several generations of model development beyond current systems.
But the race to get there first is running regardless — with all the costs, distortions, and risks that entails.
The formal model #
The political narrative describes what is happening. A game theory model explains why it is stable — why it persists even when every party knows the outcome is collectively worse than cooperation would produce.
The model has three inputs:
- Multiple blocs, each controlling some subset of the capabilities required for GPAI
- A weakest-link production function — output is constrained by the worst input
- A first-mover prize that is large enough to justify significant costs to win and significant costs to prevent others from winning
Weakest-link production #
Most production processes are forgiving about gaps. A car manufacturer short on leather for seats can substitute a different material. A software company short on senior engineers can hire more juniors. Inputs trade off against each other. You can compensate.
The AI supply chain is not like that. It is a Leontief weakest-link production function — the output is limited by the worst input, not the average. Each bloc's GPAI capability score is:
G = e × min( y_1/t_1, y_2/t_2, ..., y_k/t_k )
Where each y_k is the bloc's effective available capability in dimension k, each t_k is the minimum threshold required in that dimension, and e is an algorithmic efficiency factor — a term that shifts the hardware threshold downward if a major research breakthrough occurs. A bloc reaches the GPAI threshold when G ≥ 1.0.
The minimum operator is the critical feature. Having twice the required compute does not compensate for having no HBM, no packaging capacity, or no functioning model architecture. You cannot train a frontier model without enough compute — full stop. You cannot manufacture the chips without EUV lithography. You cannot build the EUV machine without ZEISS optics and TRUMPF lasers.
Being excellent at nine of ten dimensions does not compensate for being blocked at the tenth. In a standard competitive market, a country with gaps can partially compensate. In a weakest-link market, gaps cannot be compensated — only closed. Every gap is a veto on the entire system.
This also tells you where rational investment goes. Because the production function takes a minimum, the marginal return on improving an already dominant capability is low. The rational investment target is always the binding constraint — whichever dimension has the lowest y_k / t_k ratio. This is why China invests heavily in fabrication and chip design despite having capable models. It is why the CHIPS Act specifically targets fabrication despite US dominance in design. Both are targeting their own weakest links.
The capability trap #
The weakest-link function explains where investment should go. What it does not capture is how hard it is to close a gap once you have one.
A bloc cannot build a missing capability using money alone. Building any advanced capability typically requires a set of upstream capabilities as prerequisites. If those are also missing, investment produces far less than expected.
Building an advanced fab, for example, requires:
lithography equipment
+
process tools (etch, deposition, inspection)
+
EDA software
+
ultra-pure wafers and chemicals
+
precision components
+
accumulated engineering experience
If the bloc lacks any of these, its investment in the fab is constrained not by budget but by the upstream gap. You need advanced capability to build advanced capability. The capability you are trying to create depends on capabilities you do not yet have.
The result is a predictable trap:
The bloc lacks advanced tools
↓
It tries to build domestic tools
↓
But building the tools requires other advanced tools
↓
Domestic substitutes develop more slowly than funded
↓
The quality gap persists
Supply-chain restructuring takes many years even when funding is available. TSMC Arizona is slow because the process knowledge, engineering depth and supply ecosystem that surrounds TSMC Taiwan took decades to accumulate. Those things do not relocate by budget line.
Depreciation adds a second pressure. Capabilities do not sit still once built. Advanced chips become obsolete as the frontier advances. Equipment requires continuous service and upgrades. Engineers leave or retire. Software ecosystems evolve. A facility that matches the frontier today may lag it substantially in three years without continued investment at pace. Having a fab is not the same as having a competitive fab.
The Nash equilibrium #
A Nash equilibrium is a state where no player can improve their outcome by changing their strategy, assuming everyone else holds theirs constant.
Each bloc's payoff over the race has four components:
Payoff = expected sum of discounted [
commercial AI value from current capability
− cost of domestic investment
− cost of restrictions and lost trade
]
+ first-mover prize V [if this bloc reaches GPAI first]
− strategic loss L [if a rival reaches GPAI first]
V is the prize for being first; L is the penalty for being second. Both are large enough that they dominate the day-to-day commercial terms — which is exactly why the race looks the way it does.
In the AI race, the Nash equilibrium is mutual defection: every bloc restricts access to its critical capabilities, invests heavily in domestic alternatives, and accepts the higher cost and slower progress that results.
Consider the US. If China cooperated — shared its materials, opened its manufacturing, allowed talent flow — the US would still be better off maintaining its own restrictions. Cooperating while China cooperates gives a shared outcome; defecting while China cooperates gives the US the best possible outcome. So defecting dominates regardless of what China does.
China faces the same calculation from the other side. Both defect. This is stable. Neither side can improve by unilaterally changing strategy.
The equilibrium is not a good outcome. It is slower, more expensive, and produces less capable AI than coordinated global development would. But it is stable because the individual incentive to defect dominates — and trust is too low to make cooperative commitments credible.
First-mover prize makes it harder #
In most markets, being second is acceptable. You still capture revenue, still build a business, still provide value. The gap between first and second place is real but not catastrophic.
In a winner-take-most market with compounding advantages, being second is qualitatively different from being first. And GPAI, if it delivers on its potential, may be exactly that kind of market.
Consider the simplified stage game where each bloc chooses between Open (cooperate) and Restrict (defect):
| Bloc B: Open | Bloc B: Restrict | |
|---|---|---|
| Bloc A: Open | R, R | S, T |
| Bloc A: Restrict | T, S | P, P |
The prisoner's-dilemma structure holds when T > R > P > S. In a repeated game, cooperation can be sustained only if players sufficiently value the future — specifically when the discount factor β satisfies:
β ≥ (T − R) / (T − P)
A large first-mover prize V raises T, the temptation payoff from defecting while the other remains open. This makes the condition harder to satisfy. If the winner gains permanent control of global AI platforms, intelligence advantages, or the ability to restrict others indefinitely, the temptation to defect becomes extremely high — and cooperation becomes very hard to sustain, even across many repeated interactions where both parties would prefer the cooperative outcome.
A country that achieves GPAI first gets productivity gains across its entire economy. Those gains accelerate further development. The advantage grows. The gap between first and second place widens over time rather than closing.
This is why everyone is racing harder than near-term economics justify. They are not racing for current AI. They are racing for what current AI might become — and for the compounding advantages that accrue to whoever gets there first.
The three-bloc arithmetic #
With more than two blocs the game becomes richer — and the infeasibility result stronger. Consider three abstract blocs roughly corresponding to the US-led alliance, China, and the industrial democracies of Europe and Northeast Asia. Each controls different capabilities. Assume the GPAI threshold is 1.0 in every dimension.
| Capability | Bloc A (US+) | Bloc B (China) | Bloc C (Europe/Asia) |
|---|---|---|---|
| Algorithms and data | 1.2 | 0.8 | 0.6 |
| Chips and software | 1.3 | 0.8 | 0.5 |
| Fabrication and memory | 0.7 | 1.3 | 0.8 |
| Electricity and data centres | 0.9 | 0.7 | 1.3 |
| Materials and equipment | 0.6 | 0.9 | 1.3 |
Applying the weakest-link function to each bloc in isolation:
Bloc A (US+): min(1.2, 1.3, 0.7, 0.9, 0.6) = 0.6 — below threshold
Bloc B (China): min(0.8, 0.8, 1.3, 0.7, 0.9) = 0.7 — below threshold
Bloc C (Eur/Asia): min(0.6, 0.5, 0.8, 1.3, 1.3) = 0.5 — below threshold
No bloc reaches the threshold of 1.0. Now check pairwise coalitions, where each takes the best available capability from either member:
A + B: min(1.2, 1.3, 1.3, 0.9, 0.9) = 0.9 — still short
A + C: min(1.2, 1.3, 0.8, 1.3, 1.3) = 0.8 — still short
B + C: min(0.8, 0.8, 1.3, 1.3, 1.3) = 0.8 — still short
Only full global cooperation clears the threshold:
A + B + C: min(1.2, 1.3, 1.3, 1.3, 1.3) = 1.2 — clears threshold
The world can build GPAI. No bloc can. No pair of blocs can, under current strategic constraints. This is the mathematical statement of the central argument: the technological pieces exist globally, but no single actor can legally and reliably assemble them.
The model also shows where restrictions are most strategically valuable. The optimal denial target is whichever capability sits closest to the rival's binding constraint — the move that keeps the rival's minimum below 1.0. Restricting Bloc B's access to chips matters more than restricting ordinary servers, because chips are closer to Bloc B's weakest dimension.
Explore the equilibrium #
The arithmetic above holds its assumptions fixed. The explorer below makes those assumptions dynamic. Adjust trust, export controls, domestic substitution, algorithmic efficiency and the value of winning first. Countries regroup as the political conditions change; each bloc's score is still determined by its weakest capability.
Loading the equilibrium explorer…
Holding infeasibility #
The three-bloc example illustrates the result numerically. The formal model states it precisely: GPAI is temporarily infeasible when five conditions hold simultaneously.
Condition 1 — Domestic incompleteness. Every bloc has at least one dimension where its domestic capability falls below the GPAI threshold. No bloc is self-sufficient.
Condition 2 — Strategic access denial. Under equilibrium restrictions, no bloc can import enough from others to close all its gaps. The access channels have been constrained below the level needed.
Condition 3 — No viable restricted coalition. Every politically feasible coalition — the groupings that could actually form under current geopolitics — still has at least one dimension below threshold. No realistic alliance holds a complete frontier-quality stack.
Condition 4 — Substitution delay. Even under maximum feasible investment, at least one dimension in every bloc remains below threshold within the relevant time horizon. Gaps cannot be closed fast enough.
Condition 5 — Restriction is individually rational. Each bloc prefers restricting to cooperating unilaterally. The temptation payoff T exceeds the mutual cooperation payoff R, and mutual restriction P is preferred to being the exploited open party S.
When all five hold, no bloc reaches GPAI during the strategic period. The model is therefore falsifiable: if any one condition breaks — a bloc closes its last gap through domestic substitution, a large enough coalition forms, an algorithmic breakthrough removes a hardware dimension from the threshold, or a credible commitment mechanism makes cooperation individually rational — the equilibrium can shift.
The hoarding dynamic #
The Prisoner's Dilemma describes a single decision: cooperate or defect. The AI race is a repeated game, played across many rounds, with escalating stakes.
Repeated defection produces hoarding: restricting exports — actively accumulating and withholding, model weights not published, training data not shared, talent movement restricted, research collaborations cancelled. Each round of restrictions makes the next round of cooperation harder, because trust erodes and the cost of being the first to re-cooperate rises.
The hoarding dynamic is self-reinforcing. As each side accumulates more behind walls, the value of what is inside the walls increases. That increases the incentive to protect it. Which raises the walls higher. Which increases the perceived value further.
The endpoint of this dynamic, if nothing interrupts it, is two largely separate technological ecosystems — one centred on the US and its allies, one centred on China — that share less and less over time and develop along diverging paths, with deep divergence in the layers that matter most for AI capability.
Substitutes compound #
A natural response to restrictions is to build substitutes. The capability trap explains why that takes longer than expected. There is a second problem: even when a substitute works, it does not work as well — and that shortfall compounds upward through the entire stack.
Suppose a domestic accelerator delivers 60% of the frontier chip's performance per unit of energy. That single gap cascades:
Less efficient accelerator
↓
More accelerators needed for equivalent compute
↓
More HBM and networking required
↓
More electrical power required
↓
More cooling required
↓
Larger data centres required
↓
More capital and more construction time
The 40% efficiency gap does not produce a 40% capability shortfall. It produces something larger, because every resource requirement in the chain scales with the original gap. A bloc operating on 60% chips may need two or three times the physical infrastructure to reach equivalent effective compute.
Effective compute is the metric that actually matters:
Effective compute = (chip quantity × performance × utilisation)
÷ (energy cost per unit × network loss)
Nominal self-sufficiency — "we have chips" — is therefore not the same as threshold capability — "we have enough effective compute to cross GPAI." The gap between those two statements can be very large when the chips are substantially below the frontier. This is what the model captures when it scores capabilities: raw capacity matters, but quality-adjusted effective capability is what actually moves the G score.
Equilibrium change #
Equilibria can be broken by changes in the payoff structure. Three are plausible.
The prize shrinks. If GPAI turns out to be less winner-take-most than feared — if second-place models are commercially viable, if the compounding advantage is smaller than expected — the incentive to defect weakens. Cooperation becomes more attractive when the cost of being second is lower. This depends on how AI capability and competition actually develop.
A credible commitment mechanism emerges. The classic solution to the Prisoner's Dilemma is a binding agreement with enforcement. Arms control treaties are the historical model. A formal AI governance framework with real verification and enforcement could, in principle, change the payoff structure. The difficulty is that verification is hard (how do you confirm someone isn't training a secret frontier model?), enforcement is harder (what is the sanction that deters a superpower?), and trust is currently close to zero.
An exogenous shock forces cooperation. A global AI safety incident serious enough to threaten both sides. A technological development that makes unilateral GPAI development genuinely dangerous in a way both blocs recognise. Shared existential risk has historically been the most reliable path to cooperation between adversaries — as with nuclear arms control after near-misses. The risk would need to be large enough and credible enough to overcome the defection incentive.
None of these is imminent. All of them are conceivable.
What the model predicts #
Given the current payoff structure — weakest-link production, first-mover prize, mutual distrust — the model generates seven empirically testable predictions:
- Blocs invest disproportionately in capabilities where their quality gap is greatest, even when those investments have poor short-term commercial returns
- Export restrictions target bottleneck technologies rather than the largest industries — the aim is to keep the rival's weakest-link score below 1.0
- Supply chains become more redundant but less economically efficient
- New domestic facilities initially lag incumbents in quality, yield and cost despite substantial investment
- Alliances deepen around hardware and infrastructure, but access to the most advanced models and chips remains politically conditional
- The industrial race continues rapidly while observable model-quality improvements slow — investment is spent reproducing existing capabilities rather than extending the global frontier
- A major algorithmic breakthrough — one that raises the algorithmic efficiency factor e significantly — could abruptly overturn the industrial equilibrium by reducing the hardware requirements that the weakest-link function measures against
The core equilibrium:
mutual restriction + incomplete bloc capabilities + slow quality convergence = temporary GPAI infeasibility
Or in full:
Each bloc wants GPAI first
↓
Each fears rivals using shared technology
↓
Each restricts its strongest chokepoints
↓
No bloc can access the globally optimal combined stack
↓
Each builds domestic substitutes
↓
Substitutes remain below incumbent quality and scale
↓
The weakest-link threshold is not reached
↓
GPAI remains infeasible for all blocs
The game is locked because cooperation would maximise the probability of GPAI being created — but each participant reasonably fears that cooperation might allow another bloc to create it first. That makes the current system a threshold prisoner's dilemma embedded inside a dynamic technology arms race.
The supply chain, the chokepoints, the irreplaceable companies, the materials, the energy, the export controls, the industrial policy: all of it is the physical substrate of a competition whose outcome will shape the next generation of economic and strategic power.
The race is real. The stakes are high. And the structure of the game makes it very hard to stop.