The Assembly Lab

Who controls the AI
your country depends on?

Today, a handful of US corporations decide which governments can access frontier AI, what languages it speaks, and whose laws it follows. The CLOUD Act means your citizens' data is one subpoena away from a foreign court. Export controls shift overnight. API access gets revoked without warning. This is not a market problem. It is a sovereignty crisis. This tool explores what happens when middle powers stop buying AI and start building it together.

Policy Brief Executive Summary (2-minute read)

The Problem. ChatGPT drives ~80% of global chatbot referral traffic. The EU's largest AI lab, Mistral, holds ~2% market share, even in France. Middle-power countries (Canada, France, Germany, Japan, South Korea, UK, and others) each invest in national AI champions, but none can individually compete with the scale advantages of US and Chinese frontier labs. The result: fragmented efforts, duplicated costs, and accelerating dependency on foreign AI infrastructure.

The Precedent. In 1970, four European governments faced the same problem in aviation. Boeing dominated, and no single European firm could compete. They pooled resources into Airbus, which went from 0% to 50% market share. Today Airbus employs 157,000 people and has an enterprise value of ~$159B.

The Proposal. Merge the most promising middle-power AI companies (Mistral, Cohere, DeepL, Stability AI, Aleph Alpha, Multiverse Computing) into a Public AI Company: a coordinated public-private frontier AI lab with shared compute, R&D, and market strategy. Combined, these six companies represent $24B+ in value, ~3,900 employees, 232,000+ customers, and coverage across 100+ languages.[23] Government investment in compute, procurement, and talent programs could close the remaining gap.

The Stakes. Every year of inaction widens the gap. EU researchers continue migrating to US labs (60%+ of top EU-educated AI talent now works in the US). Cloud dependency deepens (70% of EU cloud infrastructure is US-owned). The capital gap grows ($14B EU AI investment vs. $109B US). The window for building a competitive alternative is closing. The question is not whether middle powers need a coordinated AI strategy, but whether they will act before it is too late.

Further reading: The Bennett School policy brief (Tan, Jackson, Berjon & Coyle, 2025) and the companion Oxford Blueprint paper, which adds geopolitical analysis: the US controls ~75% of global AI compute capacity, China ~15%, and the EU ~5%. The Blueprint identifies three structural advantages for "bridge powers": pooled computing infrastructure, concentrated AI talent with ties to allied nations, and domain-specific data assets for expert labeling.

Deep Data

The Numbers Behind the Narrative

Six research-backed visualizations exploring Europe's AI sovereignty gap.

Capital Flow — Where EU AI Money Ends Up
SOURCED

EU raised $14B in AI funding vs $109B in the US (2024). 57% of relocating European founders move to the US. The pipeline flows one way.

$14B
EU AI Funding (Atomico, 2024)
$109B
US AI Investment (Stanford HAI, 2024)
57%
EU Founders Relocate to US
8:1
US:EU Investment Ratio
Atomico State of European Tech 2025 Dealroom AI Europe Report 2024
Talent Migration — Where EU AI Researchers End Up
SOURCED

Top-tier AI researchers (NeurIPS authors): where they studied vs. where they work. Europe educates talent that the US employs.

~60%
Top AI researchers work in US
15%
EU founders relocated HQ (Atomico)
26K→52K
Net tech inflow to EU halved (2022→24)
MacroPolo Global AI Talent Tracker Stanford HAI AI Index 2025
Compute Infrastructure — What Europe Actually Has
SOURCED

EuroHPC supercomputers vs. hyperscaler data centers. Europe has research compute. It does not have commercial-scale AI infrastructure.

EuroHPC Supercomputers
JUPITERJuelich, DE
LUMIKajaani, FI
LeonardoBologna, IT
MareNostrum 5Barcelona, ES
DeucalionMinho, PT
11
Supercomputers + 12 AI Factory sites planned
Global Hyperscale Data Centers
United States54%
China16%
Europe~12%
1,136+
Global hyperscale DCs (US owns majority)
EuroHPC JU Synergy Research: Hyperscale DCs
Language Coverage — Digital Equality Across EU Languages
SOURCED

24 official EU languages. Only 3 are well-supported by frontier AI models.[18] 20+ official and many regional languages face "digital extinction."[19]

3
Well-supported (EN, FR, ES)
~8
Partial support (DE, IT, NL, PL...)
13+
Weak or no support
Stanford HAI AI Index 2025 European Language Equality (ELE)
AI Patent Filings — Innovation Output Gap
SOURCED

China files 70% of all AI patents.[20] But US patents are cited 7x more.[21] Europe produced only 3 notable AI models in 2024 vs 40 from the US.[22]

38,000+
China GenAI inventions (2014-23)
~6,300
US GenAI inventions (2014-23)
7x
US patents cited more than CN
3
Notable EU AI models in 2024
WIPO Technology Trends: AI Patents Stanford HAI AI Index 2025
AI Company Valuations — US vs Coalition ($B)

Standalone AI company/division valuations. Alphabet and Meta figures are estimated AI division value, not full market cap.

Source: public filings, funding announcements, and industry estimates (2025-2026)
Coalition Company Revenue Growth
SOURCED

These aren't hypothetical startups. The coalition companies are scaling fast, with real enterprise customers and real revenue trajectories.

20x
Mistral ARR Growth (12 mo.)
11x
Cohere ARR Growth (24 mo.)
100+
DeepL Languages
12.6B
Stable Diffusion Images
MLQ AI (Mistral) CNBC (Cohere) Wikipedia (DeepL) Quantumrun (Stability AI)
Who Controls Open-Source AI?
LOADING

Live data from HuggingFace: the top 50 most-downloaded AI models, classified by region. Updated every page load.

HuggingFace Model Hub
Coalition Sovereign Infrastructure
SOURCED

Beyond the core companies, the coalition nations control critical infrastructure that no US hyperscaler can replicate or route around.

ASML (Netherlands)
Revenue (2025)€32.7B
Order Backlog€38.8B
High-NA EUV Per Machine$400M
EUV Market Share100%
Market Cap~$527B
Every advanced AI chip (TSMC, Samsung, Intel) requires ASML machines.
Helsing (Germany/France)
Valuation$12B
Total Raised$1.6B+
Ukraine Drones (contracted)10,000+
Employees900
ChairmanDaniel Ek
AI integrated into Eurofighter Typhoon. Mistral partnership for defense AI models.
Mistral Compute (France)
Nvidia GB300 GPUs13,800
Data Center (Paris)44MW
Debt Financing (Mar 2026)$830M
Nuclear Grid Share~70%
France's nuclear grid provides near-zero carbon power for AI workloads.
MareNostrum 5 / BSC (Spain)
Peak Performance314 PFlops
Nvidia H100 GPUs4,480
Storage248 PB
Salamandra Languages35
Open-source Salamandra LLMs (2B/7B/40B) trained from scratch in European languages.
Wikipedia (ASML) Wikipedia (Helsing) TechCrunch (Mistral Compute) BSC (MareNostrum 5) HuggingFace (Salamandra)
The Growth Gap

How US labs pulled ahead: 2017-2025

The gap didn't open overnight. It compounded year after year as US labs attracted capital, talent, and compute at a pace no European company could match.

Historical chart showing AI lab valuations from 2017-2025. OpenAI grew from $1B (2019) to $300B (2025). Anthropic grew from $1B (2021) to $60B (2025). European labs like Mistral reached $6B and Cohere $5.5B by 2025.
Sources: PitchBook, Crunchbase, TechCrunch, Bloomberg funding rounds. Valuations as of last priced round in each year.
The Cost of Inaction

What happens if middle powers don't consolidate

Every year of fragmentation widens the gap. These projections trace the diverging trajectories of inaction vs. coalition.

2030
Sources: MacroPolo Global AI Talent Tracker, Synergy Research Group, Stanford HAI AI Index 2025, EuroHPC Joint Undertaking, Atomico State of European Tech 2025. Projections are editorial estimates for illustrative purposes.

They said it couldn't be done

In the late 1960s, European aerospace was fragmented. France, Germany, the UK, Spain — each had their own companies, each too small to challenge Boeing's dominance. This is the "middle powers' dilemma" — too small to compete alone, too proud to surrender.

So they did something radical: they merged everything. Not a loose consortium. A real company with shared R&D, shared manufacturing, one mission. A coordinated public-private partnership that bootstrapped a product and focused relentlessly on market strategy.

The result? Airbus went from 0% to roughly 50% of the global commercial aircraft market.[3] Over 150,000 employees.[4] Deliveries and profitability far exceeding Boeing's.[5] A ~$160B enterprise.[6] As Tan, Jackson, Berjon & Coyle argue: the playbook exists.[7]

Four problems, one solution

Middle powers face four interlocking failures: scale (no single nation can match US compute), limited domestic markets (too small to sustain a frontier lab), foreign capture (ChatGPT drives ~80% of chatbot referral traffic globally;[1] Mistral's Le Chat doesn't register on tracking even in France[2]), and fragmented sovereignty (dozens of national AI strategies pulling in different directions).

But across Canada, France, Germany, Japan, the Netherlands, Singapore, South Korea, Spain, Sweden, Switzerland, and the UK are extraordinary AI companies — private champions like Mistral and Cohere, public labs like BSC and RIKEN, and coalitions like New Nordics AI. Each world-class in a specific domain. (Note: ASML, a publicly traded global company, is an upstream dependency for all AI chip production, not a company that could be directly merged.)

What if they formed a Public AI Company? Not a loose consortium, but a real entity — with shared compute, shared R&D, and a market strategy. Assemble the entity below and find out.

The Components

The pieces already exist. Nobody has put them together.

Across eleven candidate nations — Canada, France, Germany, Japan, the Netherlands, Singapore, South Korea, Spain, Sweden, Switzerland, and the UK — are extraordinary AI companies. Mistral grew revenue 20x in 12 months.[12] DeepL translates 100+ languages, preferred 1.3x more often than Google Translate in blind tests.[13] Stable Diffusion has generated an estimated 12.59 billion images.[14] And in the Netherlands, ASML (€32.7B revenue, €38.8B backlog[15]) holds a 100% monopoly on EUV lithography — every advanced AI chip requires their machines.[8]

In defense, Helsing ($12B valuation[16]) has integrated AI into the Eurofighter Typhoon and delivered thousands of strike drones to Ukraine. In compute, Barcelona's MareNostrum 5 runs 4,480 H100 GPUs training Salamandra, open-source LLMs in 35 European languages.[17] None can challenge ChatGPT's ~80% referral traffic share[1] alone. But together, they cover the full stack.

11
Candidate Nations
Combined Value[23]
Combined Customers[23]
100+
Languages (DeepL)
12.6B
Images Generated (SD)
€32.7B
ASML Revenue
Interactive Tools

Build the entity. Test it. Finance it.

Use the interactive tools below to assemble a Public AI Company from real companies, stress-test it against US hyperscalers, and explore how to finance it.

Assemble the Entity Run the Stress Test Explore Financing
The Roadmap

From announcement to parity in ten years.

Airbus took three decades to reach parity with Boeing.[3] AI moves faster. With coordinated investment, a merged Public AI Company could reach competitive parity with leading US labs within a decade. The roadmap below traces that trajectory — from the first letters of intent to a 50,000-person, $350B entity serving 30+ allied nations.

This is illustrative, not predictive. But the Airbus playbook[7] shows that sovereign consolidation can work in concentrated markets, and the building blocks exist today.

The 10-Year Roadmap

From first handshake to competitive parity. Each milestone builds on the last.

Live Dashboard

Model Sovereignty Index

How sovereign are the AI models nations depend on? This index scores open-weight models across six dimensions: training data privacy, local training capability, control of weights, model independence, weight privacy, and country-specific knowledge.

6
Sovereignty Dimensions
0-100
Score Range per Model
Live
Hugging Face Metadata
Loading Model Sovereignty Index...
Open full dashboard in new tab · View source on GitHub
Why model sovereignty matters for the Public AI Company

The Airbus for AI initiative argues that middle powers need to build, not just buy, frontier AI. But how do you measure whether a country actually controls the AI it depends on? The Model Sovereignty Index provides a concrete, data-driven answer.

Training Data Privacy measures whether a model's training data can be audited and controlled, critical for GDPR compliance and preventing foreign surveillance through training data.

Local Training Capability assesses whether a model can be trained or fine-tuned on domestic infrastructure, the core argument for pooled European compute.

Control of Weights evaluates whether the model's parameters are accessible and modifiable, essential for air-gapped government deployments like Aleph Alpha's PhariaAI.

Model Independence scores how reliant a model is on foreign APIs, cloud services, or proprietary components.

Weight Privacy assesses protection against unauthorized extraction, relevant for defense and critical infrastructure use cases.

Country-Specific Knowledge measures linguistic and cultural coverage, the dimension where DeepL's 100+ languages and Cohere's Aya models provide the coalition's strongest advantage.

Risks & Limitations

What could go wrong

The Airbus analogy has limits. Airbus consolidated companies that all built the same product (aircraft) in a two-player market, and it still took an estimated $22B in government subsidies[24] and 33 years. AI companies build very different products (models, translation, compression, defense AI, open-source tools). Integrating them is harder, not easier, than merging aircraft manufacturers.

Governance across 11 nations is extraordinarily difficult. Airbus nearly failed multiple times due to Franco-German management conflicts. Adding Japan, Canada, South Korea, Singapore, and seven more countries multiplies the coordination challenge.

The revenue gap is vast. The coalition's combined ARR (~$900M) is 27x smaller than OpenAI alone (~$24B annualized). Incumbents are also accelerating; the gap may widen faster than a coalition can close it.

Talent may not follow. Top AI researchers leave Europe for US labs because of compute, publication culture, and resources. A government-backed merged entity could worsen this dynamic if it prioritizes sovereignty over research freedom.

Competitive scores in this simulator use a synergy model that assumes merging creates coordination benefits. These are editorial estimates, not engineering assessments. Use the custom data import to substitute your own figures.

Why Now

The window is closing. But the barrier to entry is lower than you think.

Building a Public AI Company does not require coordinating between 20 different bodies. It can start with a small team and two labs. A 10-person team working for 3 months, building on existing collaborations and open-source models, supplied with about 100K GPU-hours (~$150K USD), can deliver a minimum viable product and establish first proof points around product, coordination, and market demand.

The world does not currently lack good-enough open-source models. Training a $2 billion "leading frontier model" is not the minimum viable product. The Public AI Company should focus early on generating local demand for existing open models, bootstrapping data flywheels, and launching in partnership with upcoming public AI projects like Swiss AI or OpenEuroLLM.

But the urgency is real. Frontier AI's pace of development is accelerating, and current frontier training costs remain within the reach of coordinated bridge powers. Latecomers to AI development risk lasting strategic weakness given increasing barriers to entry. Every quarter of delay is a quarter where talent migrates, infrastructure dependency deepens, and the cost of catching up grows.

The question isn't whether.
It's when.

Every month of fragmentation widens the gap. Airbus proved sovereign consolidation works.[3] The playbook exists.[7] The companies exist. The political will is forming across 11 nations.

Rebuild the Entity

Support this initiative

Join researchers, policymakers, and technologists calling for a coordinated public AI strategy. We will notify you when the formal coalition letter launches.

Your information will only be used for this initiative. No spam, no sharing.

Communications Kit for Partners

Help spread the word. Use these resources to brief your network, write about the initiative, or share on social media.

100-Word Summary

A group of middle-power countries face a choice: consolidate their fragmented AI efforts into a single competitive entity, or accept permanent dependency on US and Chinese frontier labs. Inspired by how Airbus challenged Boeing, the "Airbus for AI" initiative proposes merging companies like Mistral, Cohere, DeepL, and others into a Public AI Company, backed by coordinated government investment in compute, procurement, and talent. Combined, these six companies represent $24B+ in value and 100+ languages. The playbook exists. The companies exist. The question is whether political will forms before the window closes.

Press Release Template

FOR IMMEDIATE RELEASE: Coalition of [X] nations announces feasibility study for a "Public AI Company," modeled on the Airbus consortium that transformed European aerospace. The initiative, backed by [partner organizations], proposes consolidating middle-power AI investments into a coordinated public-private frontier AI lab. "The question is not whether we need sovereign AI capability, but whether we will act before the window closes," said [spokesperson]. Full details: airbusforai.org

Social Media Card Text

Airbus went from 0% to 50% of the aircraft market by pooling European resources. What if we did the same for AI? Six companies. Eleven nations. $24B+ combined. Meet the Public AI Company: airbusforai.org

Partner Organizations

Interested organizations include: Pour Demain, Oxford Martin Programme on AI Governance, LawZero, Ada Lovelace Institute, AI Safety Canada, Trajectory Labs, New America. Contact us to join as a partner.

Assemble the Entity

Click companies to merge them into the reactor. Watch the power grow.

Available Companies
Click a company to begin the merger
Live Merger Stats
Achievements
Data: public estimates (2025-2026)
Load custom data (JSON)
The Airbus for AI Fund

How to finance a Public AI Company

Building a frontier AI lab requires patient, strategic capital. The Airbus for AI Fund is a proposed transnational investment vehicle, modeled on the NATO Innovation Fund's multi-sovereign LP structure and EuroHPC's pooled compute logic, designed to acquire, integrate, and finance shared AI infrastructure.

The fund's edge is governance-as-core-technology: holding-company structures that separate IP, operations, and deployment; contractual veto rights on capex duplication to prevent wasteful parallel spending; pre-agreed deadlock resolution mechanisms; and ring-fenced national interfaces with shared core infrastructure. This governance architecture enables integration without requiring full political merger.

Anchor Investment

Government takes a cornerstone equity stake and provides compute infrastructure + procurement contracts. Model: Franco-German Compute JV, ~EUR 35M cornerstone equity consolidating two national GPU programs.

Strategic Merger

Two or more publicly funded model teams merge into a shared IP holding company with single compliance infrastructure under the AI Act. Model: Northern Europe Foundation Model Roll-Up, ~EUR 25M minority stake.

Collective Equity

Novel legal instrument where multiple nations pool equity stakes in shared AI infrastructure without full political merger. Multi-state procurement framework with cash flow within 24 months. Model: Regulated-Sector AI Platform, ~EUR 20M project finance. See collectiveequity.com.

How It Works

A three-tier architecture for sovereign AI

The Public AI Company is not a monolith. It operates through three tiers that balance scale with sovereignty.

Tier 1
The Consortium Core

Shared model building (especially pretraining), shared compute and expertise, shared market strategy and branding. This is the unified lab that competes on the global stage.

Tier 2
National Entities

Each country identifies or organizes at least one participating entity: an existing national champion (Mistral, Cohere), a public lab (Spain's BSC, Japan's RIKEN), or a new public-private partnership. These link the consortium to public compute, local markets, and national populations.

Tier 3
Local Layer

Each national entity handles localization, in-country deployment, and public engagement. This ensures the Public AI Company serves local needs, languages, and regulatory contexts while benefiting from shared frontier capabilities.

The Stress Test

Now test it against the incumbents.

A merged entity looks impressive on paper. But can it actually compete? OpenAI hit ~$20B in revenue in 2025[9] and now runs at ~$24B annualized. Anthropic hit $14B ARR in February 2026, up from $1B just 14 months earlier.[10] Google Gemini sits atop the world's largest compute infrastructure. xAI reached a $230B valuation in January 2026.[11]

The gap is real. But it's not permanent. Use the scenario levers below to model what happens when governments invest: compute buildouts, procurement contracts, capital injections, talent programs. Every lever you pull changes the outcome.

The Stress Test

Pick a US hyperscaler to challenge. See where the merged entity wins and loses. Then invest government resources to close the gap.

Scores out of 99 — editorial estimates based on public data. Merged scores include a synergy bonus that assumes coordination benefits. These are illustrative, not predictive. Tap or hover any round to see the evidence. Import your own data.
Government Investment in Public AI Co. — pull levers to boost the merged entity