The landscape
Independence labs
Private companies focusing on strategic autonomy from the US and Chinese AI stacks. Frontier or near-frontier capability under domestic ownership.
Mistral (FR) · Cohere–Aleph Alpha (CA/DE) · Domyn (IT) · Cosine (UK)
National programs
Publicly funded models, anchored in language and culture. Models that speak the nation’s languages and reflect its law and values.
OpenEuroLLM (EU) · Korea’s MSIT program · ALLaM (SA) · SEA-LION (SG) · PLLuM (PL)
Sovereign infrastructure
Compute and cloud under local jurisdiction and vendors selling “sovereignty” as a product line.
EuroHPC AI Factories · Isambard-AI (UK) · STACKIT (DE) · NVIDIA, Oracle, Microsoft, AWS
What is Sovereign AI?
Sovereign AI is a nation's ability to build, run and control AI on its own infrastructure, data and talent instead of renting from foreign providers. AI now matters enough strategically that countries no longer want their critical systems sitting inside a handful of large tech companies, usually American or Chinese, that they cannot govern.
There are several layers to consider. Infrastructure sovereignty is owning or controlling the physical compute: the data centres and GPUs sitting inside a country's borders. Data sovereignty: keeping a nation's data under its own laws and on domestic servers rather than letting it flow abroad. Models/capability: whether a country can train or fine-tune its own foundation models. Under all three sits the talent and supply chain: the researchers, engineers and access to critical components.
Beyond capability, nations are motivated by four factors. National security: governments don't want critical systems running on foreign infrastructure that can be cut off or surveilled. Economic competitiveness: AI is treated as foundational to future productivity, so importing it means exporting the value. Cultural and linguistic concerns: for countries whose languages are not the focus of the big commercial models. Regulation: local control makes domestic privacy and governance rules easier to enforce.
In reality, sovereignty is a spectrum of managed dependencies. Nations can decide themselves how much control each layer needs.
The case for — and the case against
For
Resilience. A state that cannot run critical AI workloads without foreign permission is strategically exposed.
Bargaining power. Mistral’s Arthur Mensch frames sovereignty as leverage: without production capacity, Europe has “nothing to put on the table” — and risks becoming a “vassal state”.
Value capture. If AI spending heads toward hundreds of billions a year, importing it all adds straight to the trade deficit.
Language and culture. Models trained on Anglo-American inputs impose foreign defaults.
Rule-of-law alignment. Putting critical AI beyond foreign legal reach and being in control to prevent surveillance, subpoena. Defence and health are the strongest ground for sovereignty. As seen in the Schrems II legislation France’s air-gapped ASGARD system.
Against
Subscale economics. US hyperscalers guided ~$725B of AI capex for 2026; the EU’s public gigafactory facility is €20B, the UK’s whole programme ~£2B. Critics call the gaps unbridgeable.
Execution gaps. IndiaAI released under 4% of its budget in two years; Krutrim, India’s first GenAI unicorn, quit foundation models; OpenAI paused Stargate UK; AI21 collapsed into a niche.
The hardware problem. Nearly every program runs on NVIDIA chips, TSMC fabs and ASML lithography. Sovereignty purchased from the same chokepoints.
Provenance theatre. Many “sovereign” models are fine-tunes of foreign bases, such as China’s Qwen. Frontier capability also depreciates in months: a national model funded through a budget cycle is superseded before procurement completes.
Opportunity cost. New Zealand explicitly declined to build models; diffusion scholars argue adoption, not duplication of pre-training, determines national gains.
The European picture
Mistral AI is the strongest non-US/China frontier lab, valued at €11.7B after ASML, Europe's chip champion, led its €1.7B Series C (Sep 2025) and took 11%. Since then: ARR moved past $400M, an $830M debt raise for its Essonne datacenter (Mar 2026), a reported ~€3B round in talks at ~€20B, and a defence framework that makes Mistral the backbone of French military AI on the air-gapped ASGARD supercomputer. Around it sits a full stack: Kyutai’s open-science voice lab, Yann LeCun’s new AMI (a record $1.03B seed, Mar 2026), LightOn’s air-gapped deployment platform, and OVHcloud selling shelter from the US CLOUD Act.
The boldest structural move is the Cohere–Aleph Alpha merger (announced 24 Apr 2026; ~$20B expected combined value at close, H2 2026): a transatlantic champion financed by Germany’s Schwarz Group ($600M Series E lead), anchored on its STACKIT sovereign cloud, the flagship of the Canada–Germany Sovereign Technology Alliance. Schwarz is meanwhile building a €11B, 200MW, ~100,000-GPU campus in Brandenburg.
Brussels has moved from regulation to industrial policy with mixed delivery. Delivered: 19 EuroHPC AI Factories, 13 antennas; JUPITER, Europe’s first exascale machine, now #5 worldwide; an adopted AI Act simplification that delays high-risk obligations to 2027–28. Pending: the €20B AI Gigafactories are still unselected, with 77 bids across 16 member states waiting. The Commission’s Frontier AI Grand Challenge already picked its winner (Jun 2026): the Domyn-led EUROPA consortium will train a fully open 400B+ model in all 24 EU languages with up to 2.5% of EuroHPC’s entire compute.
National-language models now exist across the bloc: Spain’s ALIA, Poland’s PLLuM (now inside the mObywatel citizen app), the Netherlands’ lawfully-sourced GPT-NL, Bulgaria’s BgGPT — with Switzerland’s radically open Apertus as the reference “public AI,” already in clinical testing at CHUV. OpenEuroLLM, the flagship open-model consortium, has shipped only reference models against a 31 July 2026 deadline: compute, not ambition, is its bottleneck.
The British picture
Britain focused not on a state-built model, but “control and leverage”. A state compute plus state equity in private champions. Isambard-AI (Bristol, 5,448 GH200s, 21 AI-exaflops, world #13); the £500M Sovereign AI Unit, an explicit state VC, wrote its first cheques in June 2026 with up to 1M GPU-hours per company; the AI Security Institute is a world-leading sovereign evaluation capability; and a £1.1B AI Hardware Plan (Jun 2026) puts £400M into direct government purchases of UK-made AI chips becoming the first real counter to a decade of losing champions (Arm and Graphcore to SoftBank, Faculty to Accenture for >£600M in Jan 2026).
The flagship bet is Cosine’s Lumen Sovereign: a ~1.35-trillion-parameter coding model trained from scratch on Isambard-AI, co-designed by BT, Lloyds, NatWest, LSEG, BAE Systems, Leonardo and Babcock, deployable air-gapped, targeted for end-2026 — Britain’s one true frontier-training exception.
On the other hand, OpenAI paused Stargate UK in April 2026, citing energy prices and the still-unresolved copyright regime, weeks after ministers celebrated £31B of US “Tech Prosperity” pledges. Nscale’s much-touted Loughton “sovereign supercomputer” site was still a scaffolding yard in June. At ~£2B, UK public AI spend is roughly 1% of what a single US hyperscaler lays out in capex each year.
Five findings
01From-scratch vs fine-tune is the fault line.▶
Most “sovereign” LLMs sit on foreign open bases — increasingly China’s Qwen (~1 billion downloads; more than half of global open-model downloads). Korea’s national program disqualified Naver and NC AI in round one for exactly this; only independently trained models advance. Meta’s Llama still underpins 36% of open-based national models (CNAS).
02Sovereignty still means data, language and jurisdiction — not silicon.▶
But the exceptions got real in 2026: China’s GLM-5 was trained on ~100,000 Huawei Ascend chips (self-reported) and served entirely on domestic silicon; Korea fields Rebellions NPUs; Japan’s PFN runs its own MN-Core; the UK committed £400M to buy British chips.
03The label is unstable — and increasingly audited.▶
Australia’s Maincode dropped “sovereign” as divisive; New Zealand declined to build models at all; US hyperscalers sell “sovereign cloud” to everyone while Microsoft told the French Senate, under oath, that it cannot guarantee French data stays out of US hands. Sovereignty by architecture is not sovereignty by ownership.
04Language is the Global South’s driver.▶
Spanish and Portuguese are ~2–3% of mainstream training data; hence LatamGPT (launched Feb 2026), Brazil’s Sabiá, Nigeria’s N-ATLAS, Türkiye’s T3AI, and the Arabic families from Falcon to Fanar. This is the axis where a few hundred thousand dollars demonstrably buys relevance.
05The economics are hard — now with receipts.▶
IndiaAI had released under 4% of its five-year budget by Feb 2026; Krutrim quit models for GPU rental; AI21 shed 60% of staff after an acquisition that never was; OpenAI paused Stargate UK over energy and copyright. The winners so far sell picks and shovels: NVIDIA’s sovereign revenue tripled to $30B+.
The landscape, region by region
A summary of organisations and programs including what it is, who funds it, what it has achieved, and the sovereignty overview. Figures as of 6 July 2026.
Entries in our landscape research, by region — July 2026. Europe includes the EU, EU-level programs, UK, Switzerland and Norway. Click a bar to filter the entries below.
The capital extremes are striking. Saudi Arabia’s HUMAIN is the most capital-intensive program anywhere with up to 600,000 NVIDIA GPUs over three years, a $10B AMD deal, $3B into xAI, and a 6.6GW datacenter roadmap. At the other end, LatamGPT launched in February 2026 across 15+ countries on roughly $550K. Korea bought 260,000+ GPUs nationally and tripled its AI budget to ₩10.1T; Japan runs GENIAC as a 70-project accelerator.
About Mirai Collective
Mirai Collective is an AI development studio. We build AI products end-to-end, train whole teams from zero to one, and guide organisations through AI adoption — practitioners, not theorists. The full landscape dataset behind this briefing (~110 profiled organisations with funding, capability and sourcing detail, plus the claim-level verification log) is available on request: hello@miraicollective.io