As nations race to build sovereign AI capabilities, successful strategies are converging on a common formula: control critical bottlenecks. The real differentiator is no longer ambition or spending, but the ability to turn power, infrastructure, regulation and demand into lasting strategic advantage. Vish Nandlall with the analysis.
Sovereignty has become the organizing word of national AI policy. Add up announced commitments across the US, Europe, the Gulf, and Asia and the headline figures approach a trillion dollars, though much of that total is buildout intentions and mobilization targets rather than funded spend, with Stargate’s up-to-$500bn pledge alone accounting for half the arithmetic.
Behind the shared vocabulary sit very different theories of what sovereignty is. Enough national strategies are now in execution that we can stop debating the theories and start scoring the mechanisms.
Let’s start with a definitional problem the word ‘sovereign’ conveniently hides. National AI strategies are pursuing three distinct objectives that get conflated under one banner. The first is supply security: can the nation get compute, power, and models when it needs them, on terms it can live with. The second is legal authority: do the nation’s laws, not a foreign court or a foreign cloud act, govern the data and decisions inside the infrastructure.
The third is value capture: do domestic firms earn durable margins in the stack, or is the nation a well-served customer. Every strategy on the board optimizes one of these and gestures at the others. Hold that test in mind, because it is what separates the strategies that are working from the ones that are merely spending.
Five models in the field
The dealmaker model: France and the Gulf. France ran the most instructive experiment of the cycle. At the 2025 AI Action Summit it announced €109 billion in AI infrastructure commitments, since extended by SoftBank’s plan for up to €75 billion and 5 GW of capacity, anchored by a €45 billion first phase targeting 3.1 GW by 2031. The French treasury contributed almost none of it. What France sold was firm nuclear power at industrial rates below most US sites, pre-secured land including retired EDF plant sites converted to campuses, and the personal attention of a head of state who closed deals himself.
There is a clear tradeoff France has made, considering that much of the incoming capital is foreign, the GPUs are foreign, and the hyperscale logic is imported. France has secured supply relevance and negotiating position, but not yet value capture. The Gulf variant adds sovereign capital and the state as first buyer: the UAE has run a dedicated AI ministry since 2017, built the Falcon open models, and pairs MGX-scale vehicles with guaranteed public demand, while Saudi Arabia’s Humain follows the same template with PIF capital and energy as the anchor asset.
The procurement model: Korea and Japan. Seoul treats GPU acquisition as a state act with dates attached. The government committed to securing 10,000 GPUs within a single year to launch its national AI computing centre early, and the broader public-private program now anchors over a quarter-million accelerators across sovereign clouds and domestic providers.
Japan’s GENIAC program works the demand side, subsidizing compute access for domestic model developers against delivery milestones. Both rest on the same insight: committed offtake and direct procurement move on silicon timescales, while grant programs move on fiscal-year timescales, and the gap between those clocks is measured in hardware generations.
The subsidy model: the European Union, and much of the middle-power pack. The EU’s InvestAI program targets €200 billion in mobilized investment, with €20 billion of public money seeding four to five AI gigafactories of roughly 100,000 advanced processors each. The January 2026 amendment to the EuroHPC regulation created the legal machinery, mandated European-led consortia so the facilities remain sovereign assets, excluded high-risk vendors, and gave the Union an ownership share of computing time to redistribute to startups.
The call for interest drew 76 proposals from 16 member states. But European industrial electricity runs roughly double US rates, which taxes every training run the gigafactories will host, and the regulation’s own language concedes the deeper issue. The consortia must be European-led even though the engine, the GPUs, is imported. That is legal authority over the building, purchased at public expense, with value capture still an open question. India’s variant, subsidizing common compute through empaneled providers under the IndiaAI Mission, is leaner but shares the theory that access, subsidized, eventually becomes capability.
The governance model: Singapore and the United Kingdom. Singapore sells neither power nor capital. It sells certainty and plumbing through sectoral deployment programs, the AI Verify assurance toolkit that lowers the trust friction blocking enterprise deployment, and skills pipelines tied to actual deployments. That is a credible adoption machine, though adoption is value capture’s precondition, not its proof.
The UK’s contribution is the AI Growth Zone, a designation that bundles planning approval, environmental timelines, grid queue priority, and direct power purchase agreements into a single decision. The government’s own assessment names grid connection as the single biggest blocker for the zones, which is the candid version of a truth every strategy on this list eventually meets: the binding constraint is power and interconnection, not capital.
The dependency model: most of the world. The fifth model is the one few governments name out loud. A large share of sovereign AI announcements are sovereign-branded cloud regions purchased from Microsoft, AWS, Google, Oracle, or their local partners. These arrangements can improve legal authority and access, with data residency commitments and local operation under national law.
They usually worsen value capture, because the margin in the region accrues to the platform, and they leave supply security hostage to a vendor roadmap. This is simply a rational procurement choice for a small state.
The bottleneck pattern
The pattern that I see developing is state capacity applied to bottlenecks. France converted nuclear power, land, and permitting into committed gigawatts. The Gulf converts capital, energy, and state demand. Korea and Japan convert procurement clocks and industrial coordination. Singapore converts regulatory certainty. The UK is converting zoning and grid queue reform. The EU converts legal authority and pooled public finance. In every case that is working, the state monetized a bottleneck it actually controls.
The strategies that are struggling share the inverse signature: they subsidize around bottlenecks they do not control, buying access to someone else’s silicon, someone else’s models, and someone else’s operating layer, and labeling the invoice sovereignty.
That yields the harsher formulation the field needs. Residency is not sovereignty. Ownership of the facility is not sovereignty. Even subsidized access is not sovereignty. Sovereignty begins when a nation can decide which workloads run, under whose law, on whose infrastructure, with domestic firms capturing some durable margin.
What is being missed?
I see some common mode faults emerging. Let’s talk first about training sovereignty versus inference sovereignty. Training sovereignty, frontier-scale compute access, is what most of the announced capital is buying, and it is a race with three or four credible entrants. Inference sovereignty is about serving cost, data jurisdiction, latency, model routing, and policy enforcement, and it is where most of the economic value and nearly all of the legal exposure will actually live. Most governments are funding the former. Few have noticed the latter is a different problem.
The infrastructure being deployed is heterogeneous, spanning vendors, generations, memory architectures, and fabric topologies, and the layer that decides which workload runs on which silicon, under which jurisdiction, against which service target, is software. This control plane is not unclaimed territory. Hyperscalers built theirs years ago, NVIDIA is extending downward from the chip, and an open ecosystem from Kubernetes lineage through Ray, vLLM, and SGLang is racing upward from the community.
What remains open is whether that layer develops as proprietary platform estates or as an open, governable standard, the way no nation owns BGP yet every nation routes its own traffic. For middle powers, that is a key square on the board where coalition strategy beats capital, which makes it less a prediction than an option worth buying.
A bottleneck blueprint is emerging: monetize scarcity the state controls, commit demand on dates, concentrate rather than distribute, fix power and permits before announcing targets, and distinguish the sovereignty you are buying from the sovereignty you are branding.
The next sovereign layer is unlikely to be another national GPU farm. It is the control plane that governs heterogeneous compute, routes sensitive workloads, enforces jurisdiction, and turns subsidized infrastructure into usable industrial capacity. The trillion dollars announced so far has mostly been spent one layer below it.
Vish Nandlall is a leading technology strategist focused on the convergence of 6G, AI, and autonomous systems. As Founder and Lead Analyst of Vish Nandlall Consulting, he advises global operators, investors, and policymakers on the economics and architecture of next-generation networks. A former CTO at major telecom and cloud organizations, Vish has shaped industry roadmaps across 5G, edge computing, and AI infrastructure. His current work explores how intelligence, connectivity, and computation will fuse to define the 6G era.
