Sovereign AI: Why Countries Are Building National AI Infrastructure

Sovereign AI is a country's effort to control the AI it depends on, from data centres and chips to models, data and rules. Governments from the UAE to the EU are spending billions on it. But almost no country can build the whole stack alone, so in practice it means choosing which dependencies to accept.

This guide explains what sovereign AI means and where the term came from, why governments want it, what the UAE, Saudi Arabia, Europe and others are building, how US export rules shape it, what it costs in money and power, the main criticisms, and what it means for the UAE and Dubai.

Key takeaways

• It's everywhere, but the money is concentrated. CNAS tracks 184 government-backed sovereign AI projects; the UAE and Japan alone account for nearly two-thirds of disclosed investment.
• It mostly runs on American technology. More than three-fifths of tracked projects disclose a foreign partner, four-fifths of them American, and NVIDIA supplies GPUs to 45% of the infrastructure projects.
• Full sovereignty is off the table. A 2026 Brookings and CEPS study calls full-stack AI sovereignty structurally infeasible for almost any country; the aim is managed interdependence.
• The UAE is the biggest single bet. A 5 GW UAE–US AI campus, the 1 GW Stargate UAE cluster and licence-free access to US chips since July 2026 — conditional, and time-limited for G42.
• Power is the real constraint. The International Energy Agency expects data-centre electricity use to more than double to about 945 TWh by 2030, with around 20% of planned projects at risk of grid delays.
• Announced isn't operating. Europe's AI gigafactories are still in procurement, and Stargate UAE's first phase had no confirmed go-live by late September 2026.

What is sovereign AI?

Sovereign AI is a government's attempt to make sure its country can build, run, govern and keep using AI on its own terms. NVIDIA popularised the term in 2024, defining it as a nation's capability to produce AI using its own infrastructure, data, workforce and business networks. Its chief executive, Jensen Huang, made the pitch at the World Governments Summit in Dubai in February 2024: "Every country needs to own the production of their own intelligence."

The idea has since matured. Think tanks now describe sovereign AI less as self-sufficiency and more as control over choices. Brookings and CEPS call it a spectrum of strategies to strengthen a country's ability to make independent decisions about critical AI infrastructure, rather than literal autarky. The Tony Blair Institute frames it as a question of agency and choice, and Carnegie as agency, interoperability and openness. Even NVIDIA has moved on: in mid-2026 it retitled its Dubai summit post "Every Country Needs AI" and replaced its "What Is Sovereign AI?" explainer with a page about national AI strategies.

Definition

Sovereign AI — a country's capacity to decide how AI is built, deployed and governed within its borders, by controlling or securing reliable access to the layers AI depends on: computing power, energy, data, models, talent, regulation and capital. It usually combines national investment with foreign technology under agreed conditions.

No country controls every layer, so each one is a separate decision:

Layer	What control means	Where dependence usually sits
Compute	National data centres and supercomputers	Chips and systems, mostly from US vendors
Energy	Enough power, grid connections and cooling	Grid capacity and connection queues
Data	National datasets and rules on where data is processed	Cloud platforms and cross-border transfer rules
Models	National or adapted models, often in local languages	Open-weight base models from abroad
Talent	Researchers, engineers and operators	Global competition for skills
Regulation	Rules on safety, data and access	Other countries' export controls
Capital	Money to build and run all of the above	Private investors and state funds

Why do countries want sovereign AI?

CNAS, which tracks sovereign AI projects worldwide, groups the motives into four: security, the economy, culture and autonomy. Governments want sensitive data and defence capabilities under national control; they want AI to bring investment, jobs and productivity at home; they want models that understand their languages and values, which is why the UAE's Technology Innovation Institute builds Arabic Falcon models; and they want less exposure to decisions made by foreign companies or governments.

That last motive grew sharper as chips became an instrument of foreign policy. When access to the most advanced chips depends on another country's export licences, owning data centres doesn't guarantee being able to fill them.

Which countries are building sovereign AI?

Almost every large economy has a programme, but disclosed investment is lopsided. CNAS found the Middle East and East Asia account for more than 80% of tracked, disclosed sovereign AI investment, with the UAE and Japan alone at nearly two-thirds and the ten largest spenders at about 90%. Infrastructure projects make up 59% of the 184 it tracks, models 32% and data the rest.

Country	Headline commitment	Status in September 2026
UAE	5 GW UAE–US AI campus; 1 GW Stargate UAE; Microsoft's $15.2 billion (2023–2029)	Stargate UAE's first 200 MW was due in Q3 2026, with no go-live announced; a 200 MW Microsoft–G42 expansion is due to start coming online by the end of 2026
Saudi Arabia	HUMAIN, owned by the Public Investment Fund: AI factories of up to 500 MW with NVIDIA, and an AMD–Cisco joint venture targeting up to 1 GW by 2030	First AMD systems live since August 2026; next 250 MW from 2027
Qatar	A $20 billion AI infrastructure partnership between Brookfield and Qai, a QIA company	Announced December 2025
European Union	InvestAI, to mobilise €200 billion; up to seven AI gigafactories backed by up to €10 billion of EU and national money	Gigafactory call closes 12 November 2026, with awards expected early 2027
France	More than €109 billion of infrastructure announcements, including up to €50 billion from a Franco-Emirati consortium for a 1 GW AI campus	Mostly private pledges; Mistral raised €3 billion in September 2026 and plans up to 1 GW of capacity by 2030
United Kingdom	Up to £2 billion for public compute; a Sovereign AI Unit with up to £500 million	Isambard-AI, the UK's most powerful public system with 5,448 NVIDIA GH200 superchips, is running in Bristol
India	IndiaAI Mission with an outlay of ₹10,372 crore	More than 45,000 shared GPUs by June 2026
Japan	More than ¥10 trillion of public support over seven years for AI and semiconductors	July 2026 AI Basic Plan calls for open "AI Sovereignty" with strategic autonomy
Canada	About C$890 million for a sovereign supercomputer	In procurement

Public money is usually a fraction of the headline. The EU's gigafactories rest on up to €10 billion of public funding against the €200 billion it hopes to mobilise, and the UK's five AI Growth Zones attracted £28.2 billion of largely private investment against £2 billion of public compute spending.

How do US export rules shape sovereign AI?

For most countries, the most advanced AI chips are American, so US export policy sets the pace. It has changed direction twice in two years.

• January 2025: caps. The AI Diffusion Rule put countries into tiers with limits on how much computing power they could import.
• May 2025: rescinded. The Commerce Department withdrew the rule two days before its compliance requirements were due to take effect.
• July 2025: exports as policy. An executive order set out to promote exports of "full-stack" American AI packages covering chips and cloud, data pipelines, models, security and applications.
• November 2025: deals. The US approved G42 in the UAE and HUMAIN in Saudi Arabia to buy the equivalent of up to 35,000 NVIDIA GB300 chips each, subject to security and reporting requirements.
• July 2026: favoured status. A final rule moved the UAE into a more favourable export group. UAE government agencies, G42 and its cloud arm Core42 can now receive advanced chips without licences, alongside approved US companies including Amazon, Apple, Google, Meta, Microsoft, OpenAI, Oracle and xAI. G42 and Core42's status expires on 6 April 2027 unless they become US companies or win further authorisation.

The deals came with obligations. Under the May 2025 US–UAE AI agreement, the UAE committed to invest in, build or finance US data centres at least as large and powerful as those in the UAE, and to align its national security rules with the US. The UAE's official news agency says compute on the 5 GW campus is reserved for US hyperscalers and approved cloud providers. In January 2026 the UAE became the ninth signatory of Pax Silica, a US-led economic security coalition around compute, chips, minerals and energy, which had 24 signatories by June 2026.

Common misconception

"Sovereign AI means independence from the United States." Almost never. CNAS found more than three-fifths of sovereign AI projects disclose a foreign partner, four-fifths of them American, and warns that building national data centres only shifts exposure from one layer of the US technology stack to another. The share of projects with no disclosed foreign partner has risen, from 31% to 37% since April 2026, but it's still a minority.

What does sovereign AI cost?

The honest answer is billions of dollars, most of it private, and gigawatts of electricity.

What the data shows

The scale in numbers

• Chips: approvals for up to 35,000 GB300-equivalents each for G42 and HUMAIN; HUMAIN's first NVIDIA phase at 18,000 GB300 chips; India's shared pool at more than 45,000 GPUs; the UK's Isambard-AI at 5,448 superchips.
• Microsoft in the UAE: $15.2 billion from 2023 to 2029; licences for the equivalent of 21,500 A100 GPUs already in the country, and a further 60,400 A100-equivalents approved in September 2025 as newer GB300 chips.
• Capital: MGX's first fund closed at $49 billion in July 2026; Khazna Data Centers secured a $2.62 billion bank facility in 2025.
• Power: data centres used about 415 TWh, around 1.5% of global electricity, in 2024; the IEA expects about 945 TWh by 2030, slightly more than Japan uses today.

Energy is the binding constraint. The IEA estimates that, unless grid risks are addressed, around 20% of planned data-centre projects could face delays, and new transmission lines take four to eight years to build in advanced economies. That's why the UAE campus is planned around a mix of nuclear, solar and gas power, and why sites with grid access are the scarcest input.

What are the risks and criticisms of sovereign AI?

The critiques come from the same think tanks that shaped the concept.

• Waste and protectionism. Brookings and CEPS warn that sovereign AI can become a vehicle for protectionism, fragmented markets and duplicative or stranded public investment.
• Dependence moves, it doesn't vanish. Most "sovereign" projects run on foreign chips, and often on foreign base models: Meta's Llama was the base for 14 of the 25 model projects in CNAS's index that disclose one.
• Control is often shared. Brookings describes the UAE's arrangement with Microsoft as hybrid sovereignty: the UAE owns the physical infrastructure through G42 while Microsoft holds operational control, cloud services and compliance obligations under an intergovernmental assurance agreement.
• Intent outruns delivery. CNAS finds a substantial gap between intent and action in sovereign AI programmes, and Europe's flagship gigafactories won't be awarded before 2027.
• Access can be withdrawn. Chip approvals carry security conditions and, for G42, an expiry date.
• Rights. Brookings warns that sovereign AI systems can become tools for digital authoritarianism.

What this means

Sovereignty in AI is now a portfolio question, not a flag. The countries getting the most from it choose a few layers to own — usually data, power and some models — and negotiate firm terms for the rest.

What does sovereign AI mean for the UAE and Dubai?

The UAE is attempting the most ambitious version of managed interdependence: building at gigawatt scale at home, with US chips and partners under negotiated conditions, and exporting its model of sovereignty abroad.

• At home. Stargate UAE, a 1 GW cluster built by G42 and operated by OpenAI and Oracle with NVIDIA, Cisco and SoftBank, sits inside the planned 5 GW UAE–US AI campus in Abu Dhabi. OpenAI said the partnership would make the UAE the first country to enable ChatGPT nationwide. Mubadala's chief executive said in December 2025 that the first 200 MW was due in the third quarter of 2026, with more than 5,000 workers on site.
• Models. The Technology Innovation Institute publishes open Falcon models, including Falcon-H1-Arabic, giving the country its own Arabic-capable models alongside global ones.
• Abroad. G42 is offering "Digital Embassies", arguing that sovereignty can travel with a workload through government-to-government legal agreements, and announced an 8-exaflop supercomputer for India under Indian governance in February 2026. A Franco-Emirati consortium also plans up to €50 billion of investment in a 1 GW AI campus in France.
• Capital. State-linked investors are financing the asset class: MGX's $49 billion fund, Khazna's $2.62 billion facility and Microsoft's local spending.

For Dubai's economy, the practical effects show up in real estate and power. The unit of development is now the campus measured in megawatts, and land with grid access, power supply and US licensing status decide which projects get built and who can lease them. The UAE's data protection law also sets rules for sending personal data abroad, which supports demand for in-country capacity from regulated sectors even when frontier models are trained elsewhere. The investment side is covered in AI infrastructure investment and data center real estate, and the governance side in the AI governance guide.

What should you watch next?

• Stargate UAE's first 200 MW: an official go-live announcement.
• Microsoft and G42's 200 MW expansion: due to start coming online by the end of 2026.
• EU AI gigafactories: bids close 12 November 2026; awards expected in early 2027.
• 6 April 2027: the deadline for G42 and Core42 to keep their licence-free status.
• Saudi Arabia: HUMAIN's next 250 MW from 2027 and progress towards 1 GW by 2030.
• Power: grid connections and new generation, which will decide timelines everywhere.

Final takeaway

Sovereign AI started as a slogan about owning your own intelligence and has become a set of harder, narrower choices: which layers to build, which to buy, and on what terms. The money is real but concentrated, the technology is mostly American, and power is the constraint no budget can wish away. The UAE has placed the region's biggest bet on a model of negotiated sovereignty, with world-scale campuses running US chips under agreed conditions. Whether that looks visionary or exposed will depend on what gets switched on, and what stays on paper, over the next two years.