This article is authored by Manuel Kilian, Managing Director, Global Government Technology Centre Berlin.

Governments keep publishing AI strategies while the real game is being played one layer down. Compute — not algorithms, not talent, not ethics frameworks — is where power accumulates. Most states will never own the full stack. Their task is to secure access, build leverage where they can, and stop mistaking software policy for strategic positioning.

To understand the AI race, you have to understand compute. AI capabilities are bounded by compute availability. Training runs, inference capacity and model refresh cycles all depend on access to a scarce, geographically concentrated and increasingly contested physical substrate. Governments that frame their strategic challenge as 'keeping up with AI' are looking at the wrong layer. The real question is: who controls the compute, and on what terms can others access it?

This article attempts to provide a map. When struggling to find the right structure for this argument, turning to a German philosopher proved to be an unexpected help. Immanuel Kant once summarised the field of philosophy with a set of questions: Was kann ich wissen? Was ist der Mensch? Was darf ich hoffen? Was soll ich tun? — What can I know? What is man? What may I hope for? What ought I to do? 

These questions were designed to separate clarity from illusion, to force rigour before action. When applied to statecraft in the age of AI compute, they provide a valuable framework: what can governments actually know about compute power, what roles can they realistically play, what futures should they prepare for, and who benefits from the current trajectory?

One disclaimer: this analysis assumes compute scarcity persists. If algorithmic efficiency gains outpace hardware scaling (and they might), compute may become less strategic over time, not more. But betting on that outcome is a gamble most governments cannot afford to lose.

Now let us decode the board.

What Can I Know? — The Structure of Compute Power

First principle: compute is not one thing. It is a stack.

When governments say “we need compute,” they often imagine a warehouse full of GPUs. But compute is a stack with different layers: advanced chips, fabrication capacity, cloud orchestration, hyperscale data centres, stable cheap energy, high-skill operators, and a tightly concentrated global supply chain. If you control any one of these layers, you gain leverage over the rest. If you control several, you define the frontier. And notice: every layer of this stack ultimately collapses to electricity, as without power, none of it runs.

This is why current ‘AI sovereignty’ debates are usually backwards. They start at the software layer — models, regulation, ethics — and treat compute as an implementation detail. But compute is the substrate. No training run, no scaling law, no frontier model exists without it. AI policy without compute policy is theatre.

Second principle: compute is positional power.

Countries sit in different places on this stack, and that position defines what they can do.

The United States sits at the top: it controls design, capital, cloud, and the institutional ecosystem that turns compute into capabilities. Taiwan sits at a critical bottleneck: fabrication of the most advanced chips. The Netherlands holds another bottleneck: the lithography tools that make those chips possible. The Gulf states hold energy and capital. Most of the rest of the world sits at the end, as renters.

This is not a moral judgement. It is a cartography of power. And if you do not map it honestly, you will design fantasies.

Third principle: compute creates a new hierarchy of sovereignty.

In the industrial era, sovereignty meant controlling territory and production. In the AI era, sovereignty increasingly means controlling training runs, inference capacity, and update cycles. The frontier is not about having AI talent or an AI strategy. The frontier is about having enough compute to repeatedly train and refresh models at scale. That creates compounding advantage: compute today buys capabilities that buy more compute tomorrow.

Sovereignty shifts from “who owns the factory?” to “who can run the model?”

Once we internalise these three things — compute as stack, compute as positional power, compute as new sovereignty hierarchy — the rest of the geopolitics becomes legible.

What is Man (in This Context: the State)? — The Roles States Can Play in the Compute Order

Many governments are approaching compute in the wrong way because they are copying templates that only work for superpowers. This is visible in Europe, but it is a global pattern. A mid-power reads about US hyperscalers or Chinese industrial mobilisation and then tries to replicate it domestically. That is like Luxembourg deciding to build an aircraft carrier because the US Navy exists.

So the first act of compute strategy is humility: place yourself correctly. If there is higher ambition (and for Europe, there should be!) it must be followed by action. But no strategy can ignore its starting point.

Right now, there seem to be four roles states can play in the compute order. These roles are not freely chosen — geography, history, and capital constraints matter — but states do choose whether to act in line with their role or against it.

Stack owners are the United States and China. They control multiple layers of the stack at once. They can design chips, host clouds, mobilise talent and capital, and set rules for others. Stack ownership is not about self-sufficiency; it is about being the system — and about the flywheel that turns compute into capability, capability into capital, and capital back into compute.

Bottleneck holders include Taiwan, the Netherlands, South Korea, and Japan. They do not own the entire stack, but they control a layer without which the rest cannot function. Their sovereignty is asymmetric: not superpowers, but capable of tilting the global equilibrium. Bottlenecks create leverage, and they also make their holders the first target in a crisis. The strategy for a bottleneck holder is to preserve it, deepen it, and trade it for hard security guarantees.

Compute financiers and hosts, such as the UAE, Saudi Arabia, or Singapore, and a few others, do not build the stack but buy into it by converting energy abundance, capital, and diplomatic access into compute presence. Their play is simple: host compute, attract the frontier, and turn electrons into geopolitical relevance.

Compute strategists are most countries. They do not own the stack, do not hold a bottleneck, do not have unlimited capital or energy. Their goal is not ‘independence.’ Their goal is secured access, designed interdependence, and resilience, which translates into real power if architected deliberately.

Once a state accepts its role, it stops wasting time on delusions and starts building agency. And here is the important subtlety: compute power is not binary. States do not ‘have AI’ or ‘not have AI.’ They have degrees of access to training, inference, and specialisation. That is why the right strategy looks like a portfolio, not a moonshot: multiple routes to capacity, multiple suppliers, multiple alliances, multiple fallback modes.

For those who are not stack owners, sovereignty is not about owning everything. It is about not being cut off.

What May I Hope for? — What Futures Are Plausible, and Where the Traps Are

The problem with most AI strategies is that they implicitly assume one future: continuity. More chips, more scaling, more cloud, more growth, no large shocks. That might happen. But if it is the only future governments plan for, they are building strategic helplessness.

Three compute regimes are worth planning for:

• The Continuation Regime assumes the AI bubble holds. Compute keeps scaling. Supply chains remain intact. Cloud access stays broadly open. Energy remains manageable. In this world, compute centralises further into a handful of hyperscalers. Most governments remain renters. AI becomes a utility layer of global life. With smart interdependence, this scenario is manageable for many countries.
  
  
• The Fragmentation Regime assumes compute blocs emerge. Export controls tighten. Supply chains bifurcate. Access becomes explicitly political. Some countries get discounted compute as allies; others get throttled. ‘Neutral’ positions become less viable. Here, compute access becomes an alliance privilege, so treat diplomacy as part of your compute stack.
  
  
• The Shock Regime assumes the bubble pops. A Taiwan crisis, a major energy shock, or cascading supply disruption hits. The frontier freezes for many actors. Compute rationing becomes normal. AI progress turns uneven and intermittent. Resilience is the main differentiator, so governments need to plan for this regime as baseline risk.

Here is the trap: planning only for the first scenario. Hope is not a policy. Portfolios are.

The compute bubble can pop in different ways, and they do not all hurt equally. If the demand bubble bursts — AI hype deflates, enterprise spending drops — compute becomes cheaper and more available. Uncomfortable for investors, but survivable for states. If the supply side breaks — a Taiwan crisis, cascading export controls, an energy shock — the picture inverts. Chips become scarce, prices spike, allocation turns political. Those without redundant access, local fallbacks, or strategic reserves get locked out. Demand collapse is a correction. Supply collapse is a crisis. Governments should hedge accordingly.

And, across all three regimes, one variable matters more than most governments realise: energy. No serious compute strategy can ignore electrons. Training a frontier model uses electricity on an industrial scale. Running inference at the population level uses even more. Data centres are becoming the largest new source of electricity demand in many jurisdictions.

This changes the landscape. Countries with abundant, cheap and stable electricity gain a structural advantage. Those with fragile grids, expensive power or lengthy permitting processes lose out. The Gulf states understand this intuitively — their aggressive GPU purchases are not just about technological status, but also about converting energy into post-oil relevance. 

For most governments, the implication is uncomfortable: energy policy is now AI policy. Nuclear power, renewable energy, grid investment and permitting reform are not separate tracks. They are foundational to computing sovereignty. A state that cannot keep its data centres powered cannot maintain its AI ambitions.

What Ought I to Do? — Manoeuvring the Political Economy of Compute

If governments do not decode incentives, they will become instruments of other people’s business models. So the first question has to be cui bono: who benefits from the compute race?

Hyperscalers clearly gain from current dynamics. The race makes cloud platforms the indispensable middle layer. They harvest rents, data gravity, and geopolitical bargaining power. The more the world depends on them for compute, the more they become strategic assets.

Chip designers and tool monopolies sit on the scarce inputs. If a tiny set of firms continues to control design software, lithography, and advanced packaging, their pricing power becomes geopolitical power. The compute race entrenches a dependency structure so deep that even those who complain about it still reproduce it.

Energy-rich states are converting electrons into strategic relevance. AI compute is a huge demand shock for power. Countries with abundant cheap electricity can trade it for compute presence, and compute presence for geopolitical weight.

States with strong national security apparatuses benefit from scarcity. Compute scarcity becomes a justification machine: export controls, industrial subsidies, surveillance expansion, nationalisation of innovation policy. The compute race expands state capacity. That expansion can be healthy or dangerous, but pretending it is not happening is naïve.

In light of the above, what should governments do? Follow a few key principles.

Write compute strategies, not AI strategies. Most AI strategies are written as if models were the scarce resource. They may not be in the future. Compute is. Governments need plans that spell out where capacity comes from, who controls it, how it scales, and what happens when access tightens. If compute is the bottleneck of capability, it has to be the bottleneck of policy too.

Treat secured access as the first form of sovereignty. For most states, sovereignty will not mean owning the stack end-to-end. It will mean never being cut off from it. That implies long-term capacity guarantees rather than ad-hoc cloud rental, domestic hosting for sensitive workloads where possible, multi-vendor setups with credible switching rights, and legal and contractual control over critical compute dependencies. Autarky is a fantasy. Secured access is strategy.

Run compute like critical infrastructure. Compute is becoming as foundational as energy or telecoms, and it needs the same logic of resilience. Governments should assume interruption and plan accordingly: dependency mapping across chips, clouds, and energy; choke-point identification; redundancy requirements; stress-tests for fragmentation and shock scenarios. Planning only for continuity is how strategic helplessness gets built.

Play a portfolio game on the layers where influence is possible. If high-end GPUs are out of reach, do not try to imitate a superpower. Build leverage where the needle can move: energy abundance and fast permitting to attract clusters, specialised industrial or edge AI tied to national strengths, secure public-sector inference capacity, strategic open models for priority domains, vertical chips, domestic data assets, or regulatory regimes that make a territory compute-attractive. The logic is many bets, fast learning, and scaling winners.

Shape dependence through alliances and political economy. The compute race is not neutral. Hyperscalers, chip monopolies, and frontier labs profit from concentration; energy states convert electrons into leverage; national security establishments expand capacity through scarcity. If governments do not design interdependence, they drift into it. Pool compute with allies. Co-procure and build shared corridors. Set common standards and fallback access. Trade market access for real capacity and governance rights. States that cannot be stack owners can still be part of a backbone coalition, but never a passive renter.

These principles sound obvious once stated. But they are not what most governments are doing. Many are still writing software-first strategies in a compute-first world.

Conclusion: Three Sentences that Governments Shall Internalise

First, AI power is compute power.

Second, compute power is geopolitically allocated, not globally available.

Third, governments that do not place bets on secured compute access lose before the game starts.

That last sentence is the sharpest one. From a game-theory lens, not betting is conceding. And governments that concede on compute will spend the next decade regulating the surface of a world they no longer have the power to shape.