The sovereign-AI trilemma

There is a useful framework, by now well-rehearsed at Harvard's Carr Center, that describes AI regulation as a global trilemma between innovation, accountability, and inclusion: pick two, the third gives ^[1]. The framework is right as far as it goes. But viewed from a small open economy in 2026, a different trilemma is more immediate and more consequential.

Call it the sovereign-AI trilemma. Any state pursuing "sovereign AI" today must choose between three goods that increasingly cannot be maximised together: compute control, allied alignment, and citizen affordability. Pick two; the third gives.

The three corners

Compute control is the ability of a state to determine what runs on the AI infrastructure deployed inside its borders — which models, which guardrails, which data flows, which workloads are prioritised in a crisis. It is what most ministries mean when they say "sovereign AI".

Allied alignment is the willingness of frontier-model providers, hyperscale cloud operators, and chip vendors to deploy their best technology inside a given state. In 2026 this alignment is no longer informal. The Microsoft–G42 Intergovernmental Assurance Agreement, signed in April 2024 between the United States, the United Arab Emirates, Microsoft, and G42, makes that explicit: $1.5 billion in equity, alongside billions more in cumulative investment, with binding commitments on cybersecurity, physical security, export controls, technology transfer, data protection, responsible AI, and KYC ^[2]. The IGAA is, in effect, a treaty between two states and two companies — and it is fast becoming the template.

Citizen affordability is the price at which the resulting AI services are available to ordinary residents — public-sector users, SMEs, students, hospitals. Sovereignty and alignment both cost money. Someone pays.

Pick any two corners and the third gives.

The choices states are actually making

The United Arab Emirates has chosen compute control and allied alignment, and is paying for citizen affordability through state-led capital deployment. The G42 stack, Microsoft equity, and the IGAA give the UAE both a sovereign data perimeter and access to frontier technology, with the resulting cost absorbed by sovereign budgets ^[2]. This is a viable model only for states with sovereign-fund-scale capital.

Saudi Arabia has declared 2026 the Year of AI and is pursuing the same two corners — sovereignty and alignment — through a different mix of state-aligned platforms, academic build-out, and capital deployment ^[3]. Affordability is again absorbed by the sovereign.

Norway has implicitly chosen alignment and affordability, and is accepting a softer form of compute control. KI-fabrikken at Sigma2, opened in November 2025, is sovereign in the sense that it is operated by a Norwegian institution on Norwegian soil around the Olivia supercomputer; it is aligned in that it is part of the European LUMI AI Factory network and funded jointly by the Research Council of Norway and the EU ^[4]. Workloads are accessible to Norwegian researchers, businesses, and the public sector at thresholds that lower the cost of entry, which preserves affordability. The trade-off is that Norway does not, and arguably should not, attempt to control frontier-model training itself.

Pakistan has chosen affordability and is working to add alignment. The National AI Policy and Indus AI Week's $1 billion commitment to AI by 2030 are pitched as a public-good build-out, with affordability for citizens — students, public sector, SMEs — as a stated goal ^[5][6]. Alignment with frontier providers is being negotiated in parallel; sovereignty in the strong sense is not yet in scope.

The United States, viewed through a small-state lens, has chosen compute control and alignment, and is letting affordability for non-aligned populations be the lever. The 2025 export-control regime around frontier compute, layered with the IGAA template, is not subtle.

The European Union has tried to legislate around the trilemma rather than choose. The AI Act, the AI Office, and the LUMI/EuroHPC network are an attempt to keep all three corners in play. As of mid-2026 it is too early to say whether that succeeds.

Why the Nordic choice is the right one — but only with honesty

For a small, open, high-trust economy like Norway, choosing alignment plus affordability is correct. The cost of attempting full compute control of frontier models — building and operating a frontier-class training stack inside Norwegian borders — is not justifiable on a public balance sheet, and the Norwegian public would not accept the opportunity cost. Alignment via European partners (LUMI, the AI Factories network) and via cooperative compute access from US providers is a defensible second-best.

But this choice has to be stated honestly, and that is where Nordic AI policy debates have been weakest. "Sovereign AI" rhetoric in Oslo, Stockholm, and Helsinki tends to imply a third corner that is not actually being delivered. KI-fabrikken is a serious piece of national infrastructure; it is not a frontier training stack and was never going to be. Saying so plainly is not a concession — it is the prerequisite for spending the next five years on the things the Nordics can lead: literacy infrastructure, procurement craft, public-sector AI quality assurance, and cross-regional translation of policy.

Pretending the Nordics can be the UAE on a Nordic balance sheet, or that European AI sovereignty can be solved by data-centre announcements, is the most expensive form of policy theatre on offer in 2026.

What it means for the four-region playbook

For institutions working across the Nordics, the United States, the Middle East, and South Asia, the trilemma changes how a partnership conversation should open. Three questions clarify everything.

First, which two corners is your jurisdiction actually optimising? Be specific.

Second, who pays for the third? Sovereign budgets, hyperscale-cloud subsidy, donor capital, household cost, or unmet demand?

Third, what is your honest five-year position on the giving corner? Not the official press release — the working assumption.

If the answers are crisp, the trilemma is manageable and partnerships across regions become productive. If the answers are evasive, it is usually a sign that someone has not yet done the policy work the next five years require.