Sovereign AI: the 2026 country tracker

Sovereign AI is a country building artificial intelligence it controls: data centres on its own soil, models in its own language and its own money behind both. By mid-2026 the Center for a New American Security counted 184 sovereign AI projects in 67 countries (CNAS). This tracker covers 17 countries and the EU, with every figure linked to its source. Each entry was last checked on 6 October 2026.

Disclosure: we sell .si domain names, including a name for most markets in this tracker. They are marked "Name for this market" and listed at the end. They do not change what we report.

The tracker at a glance

What sovereign AI means

NVIDIA, whose chips run most of these projects, popularised the term in 2024. It defined sovereign AI as "a nation's capabilities to produce artificial intelligence" using its own infrastructure, data, workforce and business networks (NVIDIA, archived). The World Economic Forum describes states pursuing their own interests by investing in domestic AI capabilities and expertise (WEF).

In practice, national programmes combine four parts:

• Compute: national supercomputers and "AI factories", counted in GPUs or, for the biggest, in gigawatts of power.
• Models: language models trained on, or adapted to, the national language.
• Data and cloud: sensitive data kept on servers under national law.
• Money: state budgets, sovereign wealth funds and tax breaks that pull in private capital.

Few countries own all four. Most "national" models in this tracker start from a foreign base model: Ukraine's from Google's Gemma, Saudi Arabia's newest from China's MiniMax, Singapore's latest from NVIDIA's Nemotron and Portugal's from the EU's EuroLLM. Sovereignty usually means control of the data and the deployment, not the whole stack.

Why sovereign AI is in the news

Three things in 2026 moved the term from policy papers into headlines:

1. US politics. Google Trends shows US interest in "sovereign ai" peaking on 18 June 2026, as Senator Bernie Sanders' proposed American AI Sovereign Wealth Fund made headlines; Fortune reported it would be funded by a one-time 50% tax on large AI companies, paid in stock (Fortune). On 2 July the Financial Times reported that OpenAI had offered the US government a 5% stake (Bloomberg).
2. Earnings. Palantir told investors on 3 August that "demand for AI sovereignty has now been unleashed", as quarterly revenue rose 93% to $1.935 billion (SEC filing). On 26 August NVIDIA said its sovereign AI revenue grew 35% from the previous quarter and more than tripled from a year earlier (NVIDIA earnings call).
3. Money moving. Abu Dhabi's MGX closed a $49 billion fund in July, and the EU opened its call for up to seven AI gigafactories the same month (details below).

A note on search data: keyword tools built on Google Ads data show US searches for "sovereign ai" rising from 3,600 in June to 201,000 in August 2026. Google Trends does not show that August surge; its US interest fell through the month. Treat the August figure as unconfirmed.

The Gulf

Saudi Arabia

HUMAIN, launched by the Crown Prince in May 2025 and owned by the Public Investment Fund, is the kingdom's national AI company (PIF). In 2026 it moved from announcements to hardware. AMD MI355X systems went live for HUMAIN customers in August, and a joint venture of AMD, Cisco and HUMAIN targets 1 GW by 2030 (AMD). In September it released humain-m3, a 428-billion-parameter Arabic model built with China's MiniMax (Bloomberg), and about 14 GW of power has been allocated to its AI factory sites (Asharq Al-Awsat). With the state reining in spending, HUMAIN is now raising outside money: a $2.5 billion data-centre fund and a planned $10 billion venture fund, against targets of 1.9 GW of capacity by 2030 and more than 6 GW by 2034 (Fortune).

Name for this market: saudis.si

United Arab Emirates

The UAE runs its strategy through a few connected players. MGX, backed by Mubadala and G42, closed its first fund at $49 billion in July 2026, above its $45 billion target (The National). In February G42 signed a $1 billion deal for three sovereign AI data centres in Vietnam (The National), a sign the UAE now exports sovereign AI as well as building it. On models, the Technology Innovation Institute released the Falcon H1R 7B reasoning model and Falcon-H1 Arabic in January (TII), and MBZUAI and G42 released K2 Think V2, a 70-billion-parameter model (PR Newswire). Stargate UAE, the first "OpenAI for Countries" deal, plans a 1 GW cluster with a first 200 MW in 2026 (OpenAI). We found no confirmation yet that the first phase is live.

Name for this market: emiratis.si

Europe

European Union

The EU's plan rests on shared compute. Its call for AI gigafactories opened on 30 July 2026 and closes on 12 November: up to seven sites, up to €10 billion of public money and at least €20 billion of private money (European Commission). It sits inside InvestAI, the Commission's target to mobilise €200 billion for AI, including €20 billion for gigafactories (European Commission). In June the Commission proposed a Cloud and AI Development Act as part of a Tech Sovereignty Package (European Commission). Open models lag behind: OpenEuroLLM, launched with €37.4 million (European Commission), aimed for an 8-billion-parameter model by summer 2026 (OpenEuroLLM) but had published only research checkpoints by 6 October.

Name for this market: europeans.si

Germany

Germany has Europe's most visible sovereign model company. Aleph Alpha released Kolibri, a 78.1-billion-parameter open-weight model under the Apache 2.0 licence, on 3 October (Aleph Alpha), weeks after agreeing to merge with Canada's Cohere in a deal reported at about $20 billion, with Schwarz Group investing €500 million (SiliconANGLE). Deutsche Telekom and NVIDIA opened the Industrial AI Cloud in Munich in February with about 10,000 Blackwell GPUs (Telekom). The €500 million JUPITER supercomputer in Jülich was inaugurated in September 2025 (Forschungszentrum Jülich), and in August 2026 the federal government raised its support for a German AI gigafactory to €1 billion (IT-Fachportal).

Names for this market: germans.si and germanic.si

The Nordics and Baltics

The Nordic countries pool compute more than any other region. In August 2026 EuroHPC signed a €387.8 million contract for LUMI-AI, a supercomputer shared by Finland, Denmark, Norway, Estonia, Czechia and Poland (EuroHPC). A Nordic-Baltic AI centre, New Nordics AI, launched in 2025 with DKK 30 million over three years (Nordic Innovation).

Name for the region: nordics.si

Finland

LUMI-AI will be built in Kajaani on AMD MI430X chips, with about ten times the AI capacity of the current LUMI, from the second half of 2027 (LUMI). Finland's model work runs through Silo AI, which AMD bought for $665 million in 2024 (AMD); in August 2026 it released Poro 2 models for long context and maths reasoning (AMD). The government also backs a Nokia-led bid for an EU AI gigafactory (Finnish Government).

Name for this market: finns.si

Denmark

Gefion, Denmark's first AI supercomputer, went live in October 2024 with 1,528 H100 GPUs, funded with DKK 600 million from the Novo Nordisk Foundation and DKK 100 million from the state fund EIFO (Novo Nordisk Foundation). In August 2026 its operator, the Danish Centre for AI Innovation, launched a sovereign AI layer with Corti and began building a B300 cluster as a national inference platform (Corti). Danish Foundation Models, backed by a DKK 30.7 million government grant (University of Copenhagen), released Munin 1.0 in June and Mimir v1, a 1-billion-parameter model trained from scratch, in August (Danish Foundation Models).

Name for this market: dansk.si

Sweden

Sweden's national AI strategy, published in February 2026, aims for a global top-10 ranking (AI Sweden), and the 2026 budget puts SEK 479 million into AI and data (Riksdagen). In April EuroHPC signed a €29.76 million contract for the Mimer AI Factory in Linköping (EuroHPC). Industry runs its own cluster: Sferical AI, backed by AstraZeneca, Ericsson, Saab, SEB and Wallenberg, operates 1,152 GB300 GPUs (Defence Industry Europe). On models, the Wallenberg-funded WASP programme launched a new Swedish language model initiative with up to SEK 30 million (WASP), following the open GPT-SW3 family (WASP).

Norway

Stargate Norway in Narvik changed hands in 2026. Nscale will supply Microsoft with more than 30,000 NVIDIA Rubin GPUs there (Nscale), and OpenAI is not signed up as a buyer (Datacenter Forum). In August the National Library released its Borealis language models, trained on licensed Norwegian newspaper content (Mediebedriftene).

Name for this market: fjords.si

Estonia

Estonia's approach starts in schools. AI Leap, announced with OpenAI and Anthropic in 2025 (Estonian Ministry of Education), reached 154 of 156 upper-secondary schools in its first year (AI Leap), and its learning app went live in March 2026 with a €4 million budget (AI Leap). The government's Eesti.ai programme, launched in January with entrepreneurs (Government of Estonia), has €10.98 million for 2026 (Government of Estonia). The University of Tartu published EstLLM, Estonian models built on 8-billion-parameter base models (arXiv).

Name for this market: estonian.si

Southern and Eastern Europe

Portugal

AMALIA, Portugal's national language model, opened to the public in July 2026. It has about 9 billion parameters, is adapted from EuroLLM-9B and is funded with €5.5 million from the EU recovery fund plus €1.5 million to 2027 (SAPO Tek). The National AI Agenda 2026–2030 lists 32 initiatives worth more than €400 million (Government of Portugal). On compute, Start Campus is building a second building of about 180 MW at Sines (ECO), where Microsoft has committed $10 billion (Start Campus).

Name for this market: portugues.si

Ukraine

Ukraine's national language model is late. Kyivstar and the Ministry of Digital Transformation signed the project in June 2025 (VEON) and picked Google's Gemma as the base model in December (GlobeNewswire). More than 136,000 people voted in the Diia app to name it Siaivo (Ukrinform), but the promised September 2026 release slipped, and the head of a separate Ukrainian model project expects another delay (dev.ua). The model's budget has not been disclosed.

Name for this market: ukrainians.si

Israel and Turkey

Israel

In June 2026 the government approved a National AI Program targeting 100,000 processing units, without a budget for that target (TPS). A work plan approved in May makes 5,000 advanced GPUs a year available from 2027 to 2032 (J-Wire), and the first phase of the national supercomputer, built by Nebius with 1,000 B200 chips, went live in January (Jerusalem Post). For Hebrew, DictaLM 3.0 offers open models at 24, 12 and 1.7 billion parameters (arXiv).

Name for this market: israelis.si

Turkey

Turkey's AI Vision and Action Plan 2026–2030, unveiled in June, targets at least $10 billion of mostly private investment and at least 1 GW of data centres by 2030 (Invest in Türkiye). In July the industry minister said BİLGE, a 37-billion-parameter Turkish model from TÜBİTAK, would open to developers soon (Ekonomim). Turkcell and Google Cloud are building a sovereign cloud region due at full capacity in 2028 (Anadolu Agency), and T3 AI, from the T3 Foundation and Baykar, has been in beta since 2025 (Anadolu Agency).

Name for this market: turks.si

Asia-Pacific

India

The IndiaAI Mission, with an outlay of about ₹10,372 crore, had made more than 45,000 GPUs available by June 2026 and selected 20 foundation models from 506 proposals (PIB). At the India AI Impact Summit in February, 92 countries and organisations endorsed its declaration and the government pledged 20,000 more GPUs (PIB). Sarvam AI, the Bengaluru lab first chosen to build India's sovereign model (PIB), launched 30- and 105-billion-parameter models at the summit, trained from scratch on IndiaAI compute and covering 22 Indian languages (TechCrunch). In June it raised $234 million at a $1.5 billion valuation (MediaNama). Google Ads estimates put searches for Sarvam AI in India at about a million in February alone.

Singapore

Singapore's Budget 2026 created a National AI Council chaired by the Prime Minister and a 400% tax deduction for AI spending (EDB), on top of more than S$1 billion committed to AI over five years in 2024 (MDDI). The ASPIRE 2B supercomputer launched in June with more than 1,500 H200 GPUs (NSCC). AI Singapore's SEA-LION models cover Southeast Asian languages; version 4.8, released in September, is the first built on NVIDIA's Nemotron 3 (SEA-LION).

Name for this market: singaporeans.si

Indonesia

Indonesia is building through state companies. The sovereign wealth fund Danantara and Telkom are developing sovereign AI data centres, with Telkom's NeutraDC preparing up to 1 GW (Liputan6). The government and Indosat launched the Sahabat-AI app nationwide in February (Telecom Review Asia), built on a 70-billion-parameter model from GoTo and Indosat (GoTo). The national AI Roadmap 2026–2029 is still waiting for the President's signature (Selular).

Name for this market: indonesians.si

Australia

Australia is the laggard in this tracker. Its National AI Plan came out in December 2025 (CommBank), but the 2026–27 budget funds a $70 million AI Accelerator and no dedicated sovereign compute or model (SmartCompany). Private projects fill some of the gap: OpenAI and NEXTDC signed a memorandum for a Sydney campus (NEXTDC), and Maincode opened its Matilda model to public beta (Defence Connect). Firmus and CDC have ended Project Southgate with 42 MW deployed of 1.6 GW planned, according to Rampart (Rampart).

Name for this market: aussies.si

Latin America

Brazil

Brazil's national AI plan, PBIA, commits R$23 billion for 2024–2028 (Convergência Digital). In September 2026 the REDATA data-centre regime became Law 15.504/2026, with tax breaks worth R$5.2 billion this year (Teletime). In August the government announced a supercomputer in Macaíba, a Portuguese-language model developed with Huawei and iFlytek and a "Nuvem Brasileira" sovereign cloud; outlets report different totals for the package (Telesíntese). The Chamber vote on the AI bill, PL 2338, has slipped to after the October elections (Mobile Time).

Names for this market: brazilians.si and portugues.si

What the tracker shows

• Gigawatts versus GPUs. Saudi Arabia, the UAE, Indonesia and Turkey plan in gigawatts; most of Europe and Asia count GPUs in the thousands. The money gap is wider still: MGX's single fund is larger than the EU's entire public budget for gigafactories.
• Few models are fully home-grown. Sarvam's models and Denmark's Mimir were trained from scratch. Most others adapt a base model from the US, China or an EU project.
• Delays are normal. Ukraine's Siaivo, OpenEuroLLM's flagship, Turkey's BİLGE, Indonesia's roadmap and Brazil's AI bill all ran past their dates. Announcements outrun delivery.
• Sovereign AI is now an export. G42 builds data centres in Vietnam, Palantir and NVIDIA sell an engine for running models in sovereign and air-gapped environments (Business Wire), and NVIDIA now reports sovereign AI as its own revenue line.

Building for one of these markets?

A national AI company needs a name that reads as local and global at once. A demonym or language name on .si, such as saudis.si, dansk.si or brazilians.si, does both, and .si now reads as superintelligence (why .si matters). The names from this guide are listed below with public prices. Our sales tracker shows what comparable .si names have sold for.

FAQ

What is sovereign AI?

Sovereign AI is a country's ability to build and run artificial intelligence on its own terms: data centres on its own soil, models trained in its languages, data kept under its laws and funding it controls. NVIDIA popularised the term in 2024, and by mid-2026 CNAS counted 184 sovereign AI projects in 67 countries.

Which country spends the most on sovereign AI?

By money already raised, the Gulf leads: Abu Dhabi's MGX closed a $49 billion fund in 2026, and Saudi Arabia's HUMAIN targets 1.9 GW of capacity by 2030. The EU's InvestAI has the largest headline target, €200 billion, but most of that is private money still to be raised.

What is HUMAIN?

HUMAIN is Saudi Arabia's national AI company, launched by the Crown Prince in May 2025 and owned by the Public Investment Fund. It builds AI data centres, with AMD chips live since August 2026, and in September 2026 released humain-m3, a 428-billion-parameter Arabic model built with China's MiniMax.

What is G42?

G42 is an Abu Dhabi AI group and one of the backers of MGX, the UAE's AI investment firm. It co-developed the K2 Think V2 model with MBZUAI, is a partner in Stargate UAE, and in 2026 signed a $1 billion deal to build three sovereign AI data centres in Vietnam.

Is sovereign AI the same as open-source AI?

No. Many national models are released openly, such as Germany's Kolibri and Denmark's Mimir, but sovereignty is about who controls the infrastructure, the data and the deployment. The biggest Saudi and Emirati projects are state-controlled, and several national models are built on foreign base models.