Explainer

What is sovereign AI? A plain definition, and the debate

Sovereign AI means control over the AI you depend on. Rival definitions, the seven layers, degrees of sovereignty, how it is measured and the critics.

Lumen, wearing a hard hat, places the top layer on a tower of glowing slabs on a hill under a protective glass dome, like an architect finishing a building.

Sovereign AI means having control over the AI you rely on, from the data centres and chips up to the models, the apps and the rules. Nobody owns every layer, so in practice it is a question of degree: which layers you control, which you steer and which you depend on. This page explains the definitions, the layers, the scale and the debate; for what each country is building, see our country-by-country sovereign AI tracker.

184Sovereign AI projects in 67 countries plus the EU (Aug 2026)
$500–600BAI spending a year that sovereignty could shape by 2030
45%Tracked infrastructure projects running on NVIDIA GPUs
Source: CNAS, McKinsey, CNAS interactive

Sovereign AI definition: five versions side by side

WhoDateWhat they mean by sovereign AIEmphasis
NVIDIA2024A nation producing AI with its own infrastructure, data, workforce and business networksNational capability
McKinsey report3 Mar 2026A nation's or organisation's ability to develop and control its own AI, for strategic independence and alignment with its own values and lawsControl and independence
McKinsey explainer6 Mar 2026Not owning the technology, but keeping control over the whole AI life cycle, from compute to model logicControl, not ownership
OracleUndatedA government's or organisation's control over AI technologies and the data around themCloud and data
Stanford HAI17 Feb 2026Governments use a hard version (self-sufficiency) or a soft one (strategic autonomy); companies mean operational controlMap of the meanings

Every definition shares one word: control. They differ on who holds it (a nation or any organisation) and how much of the stack it must cover. Vendors tend to define sovereign AI as the thing they sell, such as a cloud region or a database. McKinsey's own partner Melanie Krawina says there is no official definition yet.

Our plain version: sovereign AI is a country's or organisation's control over the AI it depends on, meaning where the compute sits, whose law covers the data, who owns the models and who can switch them off.

Where the term came from

  1. Jensen Huang tells the World Governments Summit in Dubai that every country needs to own the production of its own intelligence (NVIDIA)

  2. Oxford researchers map the world into Compute North, Compute South and Compute Desert (AIES)

  3. Tony Blair Institute sets out seven layers and Control, Steer, Depend (TBI)

  4. Stanford HAI maps the hard and soft meanings of AI sovereignty (Stanford HAI)

  5. McKinsey sizes the market at $500–600 billion a year by 2030 (McKinsey)

  6. Stanford HAI reviews commercial sovereignty offerings (Stanford HAI)

  7. CNAS Sovereign AI Index counts 184 projects (CNAS)

NVIDIA made the idea mainstream. At the Dubai summit Huang argued that a country's AI codifies its culture, history and common sense, so it should own its data and the "AI factories" that turn data into intelligence (NVIDIA). NVIDIA also sells the chips most of these projects run on.

From 2026 the think tanks caught up. They added structure (layers and degrees), sizing (McKinsey) and counting (CNAS). Our tracker follows how the term then moved into earnings reports and budgets.

The sovereign AI stack: seven layers

  1. 1
    Compute infrastructure

    Data centres, chips and the networks that link them

  2. 2
    Energy

    The power that runs the compute, now planned in gigawatts

  3. 3
    Data

    What models learn from and work on, and which law covers it

  4. 4
    Models

    The AI systems themselves, trained at home or adapted from foreign ones

  5. 5
    Applications

    The tools people and agencies actually use

  6. 6
    Talent and skills

    The people who build, run and audit all of the above

  7. 7
    Governance

    The rules, standards and oversight that steer the stack

Source: Tony Blair Institute, 19 January 2026

The Tony Blair Institute splits AI into seven layers. The point of the list is that sovereignty is decided layer by layer. A country can own its data and its governance but rent its chips, which is the normal case. The institute's own conclusion is that in the age of AI no country can be fully self-sufficient.

McKinsey cuts the same problem a different way, into four dimensions of control (McKinsey):

  • Territorial: where data and compute physically sit.
  • Operational: who manages and secures them.
  • Technological: who owns the underlying stack and its intellectual property.
  • Legal: which jurisdiction governs access and compliance.

Put the two together and you get a grid. For each layer, ask the four questions. A data centre on home soil (territorial) run by a foreign company (operational) on foreign chips (technological) under local law (legal) is partly sovereign, and only on some axes.

An exploded stack of six glowing slabs representing power, chips, data, model, apps and people, some solid to show what is owned and some translucent to show what is borrowed, with Lumen pointing at the gaps.
Most sovereign AI stacks mix owned layers with borrowed ones.

Degrees of sovereignty: control, steer, depend

  1. 1
    Control

    The state owns, legally commands or exclusively runs the layer, because it is too critical to outsource

  2. 2
    Steer

    The state shapes outcomes through regulation, procurement, partnerships and standards, without owning the system

  3. 3
    Depend as a maker

    Reliance on foreign providers, but negotiated: co-development or technology transfer

  4. 4
    Depend as a taker

    Foreign systems adopted wholesale, with little ability to modify, audit or govern them

Source: Tony Blair Institute, 19 January 2026

This is the most useful idea in the debate. Sovereignty is not yes or no. It is a choice of posture for each layer. A country might control its health data, steer its cloud providers through procurement rules and depend on imported chips as a maker, with a deal that brings some skills home.

McKinsey's Krawina describes the same thing as a spectrum, where the right level depends on the use (McKinsey). Stanford HAI goes one step further: the goal should be strategic interdependence, meaning chosen and managed dependence, not self-sufficiency (Stanford HAI).

The difference between a maker and a taker is the key test. Both depend on someone else. Only the maker has a seat at the table.

Four ways countries and companies do it

ModelWhat is sovereignWhat is notExample
Full national stackCompute, models and data at homeChips, almost always importedIndia's Sarvam models, trained from scratch (tracker)
Sovereign cloudData and operations in-country, under local lawThe cloud vendor's ownershipAWS European Sovereign Cloud, Brandenburg, Germany
Fine-tuned open weightsA national-language model the country runsThe base model, often foreignMost model projects in the CNAS index
PartnershipNational capacity, built or run under contractThe partner's technologyHUMAIN with AMD; Stargate UAE (tracker)

Most real programmes mix these. The full stack is the rarest because almost nobody makes advanced chips at home.

The sovereign cloud is the fastest route for companies. AWS made its European Sovereign Cloud generally available on 14 January 2026. Its first region, in Brandenburg, runs independently of other AWS regions and is to be operated only by EU residents (AWS). The data stays under EU control; the company is still American.

Fine-tuning is the cheapest route for models. CNAS found that most model projects with a disclosed base are fine-tuned versions of foreign open-weight models, and three-fifths of those bases are American (CNAS). About 40% of model projects are trained from scratch.

How sovereign AI is measured

CNAS Sovereign AI Index: project mix (% of 184 projects)

CNAS Sovereign AI Index: project mix
Infrastructure59% % of 184 projects
Models32% % of 184 projects
Data9% % of 184 projects
Source: CNAS interactive, updated 18 August 2026 (projects in more than one category count in each)

The main yardstick is the CNAS Sovereign AI Index. It tracks government-backed projects: 184 in 67 countries plus the EU as of 20 August 2026, with 41 added in the first half of 2026 (CNAS). Infrastructure dominates, and the share is growing: 80% of new projects in the first half of 2026, up from 62% in the six months before.

80%New projects in H1 2026 that were infrastructure
3 in 5Projects naming at least one foreign partner
4 in 5Of those, involving a US company
Source: CNAS, CNAS interactive

The second yardstick is where compute physically sits. Oxford researchers sorted countries into a Compute North, with cloud GPUs fit for training models; a Compute South, with capacity mainly for running them; and a Compute Desert, with no public cloud AI compute at all (AIES). A 2025 follow-up counted 132 AI-equipped public cloud regions in 33 countries; most countries are in the desert (ITPro).

Neither measure tells you how sovereign a project is. A count of projects rewards announcements; a map of cloud regions ignores who owns them. The honest test is still the layer-by-layer one: control, steer or depend.

The case for and against

The case for

  • Security: critical systems are not switched off from abroad
  • Law: citizens' data stays under domestic rules
  • Language and culture: models that handle local languages and norms
  • Resilience if a supplier cuts access
  • Economics: McKinsey sees about €480B a year of GDP impact in Europe by the end of the decade

The case against

  • Cost and waste: duplicate infrastructure at a premium
  • Lag: home-built models trail the frontier
  • Energy: more data centres, more power demand
  • Building ahead of demand, before talent and rules are ready
  • Dependence moves rather than disappears
Source: McKinsey explainer, Stanford HAI

The case for is about control in a crisis. If a foreign supplier can cut access, change terms or be ordered to hand over data, a country's hospitals, courts and armed forces are exposed. McKinsey adds an economic case: about €480 billion a year of GDP impact in Europe by the end of the decade (McKinsey).

The case against starts with McKinsey's own risk list: cost and waste, latency, innovation risk, energy use and building ahead of demand. Stanford HAI's July 2026 brief is sharper. It found that sovereignty offerings from NVIDIA, Microsoft, Google, AWS and OpenAI do add domestic control, but "reconfigure, rather than eliminate, dependence" on those firms (Stanford HAI).

Investor Nathan Benaich of Air Street Capital argues that much of the label is political branding, since the chips, software and networks stay foreign (Air Street Press). The CNAS data backs the dependence point: NVIDIA supplies GPUs to 45% of tracked infrastructure projects. Supporters reply that partial control still beats none. Both sides have a case; the evidence so far says most "sovereign" AI is managed dependence.

Sovereign AI vs data sovereignty vs open source

TermWhat it controlsCan you have it without sovereign AI?
Data sovereigntyWhere data is stored and processed, and which law appliesYes. Data can sit at home while the AI on top is foreign
Open-weight or open-source AIWho can inspect, run and modify a modelYes. Open weights say nothing about who owns the compute or runs the service
Sovereign AIControl across the whole life cycle: compute, data, models, deploymentIt usually includes the other two, at least in part

These terms get mixed up constantly. McKinsey's Ali Ustun puts the gap simply: "you can actually have data sovereignty, but you may not have sovereign AI" (McKinsey). Data sovereignty is one layer. Sovereign AI is the intelligence built on top of it.

Open weights help but do not settle the question. A national model fine-tuned from a foreign open-weight base can be run, audited and changed at home. That is real control at the model layer. It does nothing for the chips, the power or the cloud contract underneath.

What this means if you are naming a sovereign AI company

Signal in a nameWhat it tells a buyer
Region or peopleBuilt for this market, under its rules
LanguageTrained on and for the local language
ExtensionWhat category you belong to

A sovereign AI company sells trust that it belongs to a place. Names built on a people or region, such as a demonym, say that in one word, and .si now reads as superintelligence (why .si matters). Disclosure: we sell .si names; the ones for European markets are listed below and do not change what we report.

FAQ

What is sovereign AI, according to NVIDIA?

NVIDIA describes sovereign AI as a nation's capability to produce artificial intelligence using its own infrastructure, data, workforce and business networks. The company popularised the idea in February 2024, when CEO Jensen Huang told the World Governments Summit in Dubai that every country needs to own the production of its own intelligence. NVIDIA also sells the GPUs behind 45 percent of the infrastructure projects that CNAS tracks.

What countries have sovereign AI?

Most large economies now have some form of it. As of 20 August 2026 the CNAS Sovereign AI Index counted 184 government-backed projects in 67 countries plus the European Union No country controls every layer, because advanced chips are almost always imported. Our country-by-country sovereign AI tracker at whyworry.si covers 17 countries and the EU in detail.

What is a sovereign AI fund?

The term has two meanings. The first is a state-backed fund that invests in AI, such as Abu Dhabi's MGX, which closed a $49 billion fund in July 2026. The second is a proposed public fund that would give citizens a stake in AI companies, such as the American AI Sovereign Wealth Fund proposed by Senator Bernie Sanders in June 2026.

Is sovereign AI the same as data sovereignty?

No. Data sovereignty covers where data is stored and processed and which country's law applies to it. Sovereign AI is broader: it covers control of the compute, the models, the applications and the rules as well. A country can keep all its data at home and still run foreign AI on top of it, which gives it data sovereignty without sovereign AI.

How is sovereign AI measured?

There is no single score. The CNAS Sovereign AI Index counts government-backed projects by type: 59 percent involve infrastructure, 32 percent models and 9 percent data. Oxford researchers map where AI-capable cloud compute sits and sort countries into Compute North, Compute South and Compute Desert. The Tony Blair Institute's Control, Steer, Depend scale asks how much control a country has at each layer.

Sources

Every figure in this guide links to its source. Last verified 2026-10-11.

  1. mckinsey.com mckinsey.com/industries/technology-media-and-telecommunications/our-insights/sovereign-ai-
  2. mckinsey.com mckinsey.com/featured-insights/mckinsey-explainers/what-is-sovereign-ai
  3. web.archive.org web.archive.org/web/20250126124210/https://blogs.nvidia.com/blog/what-is-sovereign-ai/
  4. blogs.nvidia.com blogs.nvidia.com/?p=69708
  5. oracle.com oracle.com/artificial-intelligence/sovereign-ai/
  6. hai.stanford.edu hai.stanford.edu/news/ai-sovereigntys-definitional-dilemma
  7. institute.global institute.global/insights/tech-and-digitalisation/sovereignty-in-the-age-of-ai-strategic-c
  8. aws.amazon.com aws.amazon.com/blogs/aws/opening-the-aws-european-sovereign-cloud
  9. ojs.aaai.org ojs.aaai.org/index.php/AIES/article/view/31683
  10. itpro.com itpro.com/infrastructure/ai-infrastructure-global-divide
  11. cnas.org cnas.org/publications/cnas-insights/cnas-insights-sovereign-ais-second-wave-is-coming-into
  12. interactives.cnas.org interactives.cnas.org/?p=2715
  13. hai.stanford.edu hai.stanford.edu/policy/the-commercial-landscape-of-ai-sovereignty-offerings
  14. press.airstreet.com press.airstreet.com/p/sovereign-ai-paradox
  15. thenationalnews.com thenationalnews.com/business/markets/2026/07/01/abu-dhabis-ai-investment-firm-mgx-raises-4
  16. fortune.com fortune.com/2026/06/03/bernie-sanders-ai-ownership-sovereign-wealth-fund-electrification/

Names that fit this guide

All names →

We own and sell these names. Prices are public; every offer gets a reply.

Names this good don't wait.

Escrow.comProtected payment
.si registryOfficial transfer
Every offerGets a reply
WorldwideAny buyer, any country
Browse the names