Sovereign AI in three layers, compute, models, and data and rules, with what the EU, UK, Canada, India and Switzerland have funded in each.
The phrase turns up in two kinds of document. Companies that sell cloud, chips or software use it on product and glossary pages, usually with a one-paragraph definition. Governments use it in the plans that fund national programmes, and those come with sums, dates and named schemes that anyone can check.
The clearest official statement sits in the UK's Compute Roadmap of July 2025. It says sovereignty means having "the ability to act independently and effectively where it matters most", then lists what that covers: allocating compute to national priorities, protecting sensitive data, and supporting research, innovation and public services on the country's own terms. Isolation and self-sufficiency are ruled out. India's cabinet, approving its own mission in 2024, wrote of ensuring the "tech sovereignty of India", and Canada put the word into the name of a funding programme.
The OECD saw the gap early. Its blueprint on national compute, published in February 2023, found that no country had "data on, or a targeted plan for, national AI compute capacity". It filed sovereignty under resilience, next to security and sustainability.
Government spending on sovereign AI falls into three layers:
Compute is where the largest published sums sit. For how the chips themselves differ, see GPU vs TPU.
European Union. The Commission's AI Continent Action Plan of 9 April 2025 builds on a network of AI factories around EU supercomputers. Thirteen were being deployed when the plan launched; the Commission's page now lists at least 19. Next come up to five AI gigafactories with about 100,000 advanced AI chips each, which the Commission puts at four times the current AI factories and links to keeping "the EU's strategic autonomy in critical industrial sectors and science". InvestAI, launched in February 2025, aims to mobilise 200 billion euros of investment in AI, including 20 billion euros for the gigafactories. The Commission's own initial share comes from existing programmes such as Digital Europe, Horizon Europe and InvestEU, and it put support for AI factories at 10 billion euros, co-financed by the EU and member states.
United Kingdom. The Compute Roadmap commits up to 2 billion pounds between now and 2030. Over 1 billion pounds of that expands the AI Research Resource twentyfold, from 21 AI exaFLOPS in 2025 to 420 by 2030, and up to 750 million pounds buys a new national supercomputer at the Edinburgh Parallel Computing Centre, due online in 2027. Some capacity is ring-fenced. The AI Security Institute and the Sovereign AI Unit get dedicated tracks.
Canada. Budget 2024 set aside 2 billion Canadian dollars over five years, starting in 2024-25, and the government now runs it as the Canadian Sovereign AI Compute Strategy, split three ways: up to 700 million for commercial AI data centres through the AI Compute Challenge, up to 1 billion for public supercomputing, including a new system described as "Canadian-owned and located", and up to 300 million to help Canadian businesses buy compute. Canada funds supply and demand in one programme.
India. The cabinet approved the IndiaAI Mission on 7 March 2024 with an outlay of 10,371.92 crore rupees (a crore is ten million), in furtherance of the vision of "Making AI in India and Making AI Work for India". Its compute pillar called for 10,000 or more GPUs built through public-private partnership. By 30 May 2025 the government reported 34,333 GPUs empanelled with industry partners and offered as a shared platform for training and inference.
Ownership is where these programmes split. India's shared GPUs are offered by private providers, who bid prices for each category of chip and are empanelled by the mission, while the UK and Canada are also paying for new public supercomputers.
India selects startups. Its call for foundation models drew 506 proposals by 30 April 2025. Sarvam AI was picked first, on 26 April 2025, to build what the government calls India's Sovereign LLM Ecosystem: an open source model of 120 billion parameters for governance and public services. Three more came in May. Soket AI is building a 120 billion parameter open source model, Gnani AI a 14 billion parameter voice model and Gan AI a 70 billion parameter multilingual one.
Switzerland used its public research institutions. On 2 September 2025 EPFL, ETH Zurich and the Swiss National Supercomputing Centre (CSCS) released Apertus in two sizes, 8 billion and 70 billion parameters. It was trained on 15 trillion tokens across more than 1,000 languages, 40% of the data non-English, and its architecture, weights, training data and recipes are all published. The release names its funding as over 10 million GPU hours on Alps at CSCS plus support from the ETH Board, with no franc figure attached. The developers say the design ensures "full reproducibility of the training process".
The UK invests like a fund. The Sovereign AI Unit sits in the Department for Science, Innovation and Technology, and the roadmap gives it up to 500 million pounds, in partnership with the British Business Bank. Its own site now lists a fund of more than 600 million pounds, cheques of 1 million to 10 million pounds and up to one million GPU hours per startup. Foundation models are one of its six priority areas.
The EU ties models to compute. The Commission expects the gigafactories to "lead the next wave of frontier AI models".
Before comparing these figures, check what each one counts.
Companies meet these programmes in two places. The first is compute, which some of them can use: Canada's access fund supports purchases by Canadian businesses, the EU's AI factories are meant to serve startups, industry and researchers, and the UK unit offers GPU hours to the startups it backs. The second is the rules: the AI Act applies to deployers of AI systems established in the EU and to providers placing systems on the EU market, wherever those providers are based.
Government documents tend to use "sovereign" for something they fund, such as Canada's Sovereign AI Compute Strategy or India's Sovereign LLM Ecosystem, and "sovereignty" for the policy aim, as in India's "tech sovereignty". In practice the two overlap, and neither has a legal definition.
Compute and data centres located in a country and under its control or law. Canada calls sovereign public AI infrastructure "essential for ensuring Canada's competitiveness and sovereignty", and the UK roadmap reserves capacity for government AI functions. Canada and the EU also back commercial data centres on their own soil: the EU aims to at least triple its data centre capacity within five to seven years.
Not in the five programmes above. The UK roadmap rules it out in writing, India's shared GPUs come from private providers, the EU's gigafactories need private capital, and Apertus is offered to business customers through Swisscom.
Data sovereignty is one part of it: where data is stored and which law applies to it. Sovereign AI adds the compute that processes the data and the models trained on it.
The AI infrastructure panel at the Davos Super Intelligence Summit 2027 covers AI chips and compute, data centres, energy and scaling, and sovereign AI infrastructure. The summit takes place on 19 January 2027, from 16:00 to 20:00 CET, at the Grandhotel Belvedere during Davos week. It is organised by Davos Premier, and access is by invitation only. The programme sets out all four keynote-led panels.