Data and foundation models are the new commons.
If my data are used to train a model that does not belong to me, and the model then replaces me, that is a form of expropriation. Data-as-Labor (Lanier/Weyl) is the methodologically cleanest answer. The question is who takes the first step.
Where do things stand today, and where do they break?
Foundation Models, the structurally most important AI systems, are today controlled by a handful of actors: OpenAI, Anthropic, Google DeepMind, Meta, xAI, plus Chinese champions (Baidu, Alibaba, ByteDance, Huawei). They were trained on public data: texts, images and code from millions of creators, without those creators sharing in the value their data generate. Compute infrastructure is concentrated among a few hyperscalers (AWS, Azure, Google Cloud). Data protection regulation is fragmented: the EU with GDPR and the AI Act, the US with a patchwork of state-level rules, China with a cyber-sovereignty logic. Personal Data Dividends have been discussed as a concept (Lanier, Weyl in Radical Markets 2018), yet nowhere institutionalised. Data cooperatives (MIDATA Switzerland, Salus Coop Barcelona) exist as a niche, but without structural scale.
Foundation Models will become the dominant productive force over the next 5–10 years. They produce knowledge, automate decisions and replace knowledge work. If the Ownership question this infrastructure remains unresolved, we risk arriving at a techno-feudal structure: 90% consumers, 10% recipients of model rents. At the same time, the data question acute: my own texts, code and images train models that substitute me. That is a specific form of expropriation. The training data carry an economic value that has so far not flowed to the creators. NYT vs. OpenAI (2023) is only the first wave; the next will ask for creator participation at scale.
Paths out of the crisis
The stars show how realistic a path is: 1 star = barely feasible, 5 stars = highly feasible.
Personal Data Dividends (Data-as-Labour)
Data are classified as labour: creators receive shares in the economic value generated by the Foundation Models that use their data.
+ Methodologically sound, democratising ownership.
Technically complex (attribution, valuation).
Trade-off: Accurate remuneration versus operational complexity.
Feasibility: Lanier/Weyl Radical Markets as a theoretical foundation. Microsoft Research and Stanford are actively researching this. Technical hurdles around attribution are real, but not prohibitive. Politically feasible in the EU; difficult in the United States.
European AI Commons (foundation models as a common good)
Foundation Models as a digital commons: open weights, public funding, democratic governance. EU/Swiss scaling of BLOOM, Mistral, EuroHPC.
+ Prevents monopolisation of critical AI infrastructure.
Compute costs: enormous.
Trade-off: Democracy vs. speed — competing with the US and China.
Feasibility: EU EuroHPC (~€7bn), Mistral (€4.5bn valuation 2024), BLOOM as an open-source exemplar. Switzerland with CSCS Lugano has a compute base. Realistic for public administration, science and public education; harder as a replacement for hyperscalers.
Data cooperatives scaling (MIDATA model)
Citizens collectively manage data through cooperatives, share revenues democratically, and license that data to AI providers.
+ Federated, citizen-controlled, GDPR-compliant.
Scaling challenge against hyperscalers.
Trade-off: Sovereignty vs. competitiveness.
Feasibility: MIDATA Switzerland (since 2015, health data), Salus Coop Barcelona (medical data), Solid (Tim Berners-Lee). Switzerland's co-operative tradition (Migros, Coop, Raiffeisen) underpins the model. Politically uncontroversial, technically ready for scale.
Mandatory sovereign stakes in foundation models
The state secures mandatory stakes (15–30%) in strategically important Foundation Models, following the logic of Norway's oil fund.
+ Broadens access to value creation through capital ownership.
Asymmetric negotiating position (US/China dominant).
Trade-off: Sovereignty vs. pace of innovation.
Feasibility: The Norway model ($12tn AuM in sovereign wealth funds globally). Switzerland has its AHV as a precursor structure. Works better for the next generation of AI than for the established Mag-7. Geopolitically sensitive.
Open-weight standard with mandatory licence
Foundation Models above a certain size must publish their weights, analogous to compulsory patent licensing in the pharmaceutical industry.
+ Open science, democratising.
Security risks (misuse), weaker incentives for innovation.
Trade-off: Openness vs. control.
Feasibility: Meta Llama has released its weights (Llama 3 open weights, May 2024); Mistral and BLOOM have done likewise. Broad-based regulation is politically very difficult. More likely to emerge as a niche standard for specific sectors such as science and education.
Trigger-based phases
Trigger: Swiss federal involvement in EuroHPC via the CSCS compute contribution.
2026–2029. Switzerland as an associated EuroHPC partner. CSCS Lugano contributes 5–8% of EuroHPC compute. MIDATA scales to 100,000+ members across three sectors with federal funding. Federal Council proposal for a Swiss Data Trustee Act submitted to parliament.
Trigger: Mistral / Aleph Alpha / EuroHPC consortium reaches GPT-4 level.
2029–2032. European foundation models become the standard in public administration, research and education. Swiss Data Trustee Act enters into force. MIDATA reaches 500,000+ members. Extension of the EU AI Act to cover Personal Data Dividends under discussion.
Trigger: EU regulation on Personal Data Dividends for foundation models exceeding 100bn parameters.
2032–2036. Creator remuneration becomes standard practice. Microsoft Research, Anthropic, and Mistral implement attribution mechanisms. MIDATA evolves into a distribution platform for data co-operatives. Switzerland enacts its own equivalent legislation.
Trigger: More than 50 per cent of public AI applications in the EU and Switzerland run on sovereign infrastructure.
2036+. European AI Commons becomes standard public-sector infrastructure. Data cooperatives achieve mainstream status. Personal Data Dividends flow on a quarterly basis. The Swiss model serves as the European reference.
Real-world pilot programmes
Where the ideas behind this pillar are already being tested. "Learn more" opens the explanation without leaving the page.
What we are watching
A structural sovereign AI position. The 25% threshold establishes a sovereign negotiating position relative to US hyperscalers.
Track in the index →Data cooperatives as mainstream infrastructure. Threshold of 500,000 corresponds to ~6% of the Swiss population.
Creator remuneration institutionalised. Threshold: EU regulation with a clear attribution obligation for foundation models.
Public-sector AI sovereign. The 70% threshold marks the structural shift from hyperscaler to sovereign stack.
Track in the index →Our assessment
At Globalance, we regard the data question and the foundation-model question as the central ownership question of the 21st century., structurally more important than income tax reform or Universal Basic Income (UBI). Whoever controls data and models controls future value creation. All other distributional questions will be consequences of this ownership structure.
Viable is a Three-Layer Strategy: Up European AI Commons as a digital commons (EuroHPC, Mistral, EU-sovereign AI) as a counterweight to US hyperscalers and Chinese sovereign models. Middle Data cooperatives as a citizen-led aggregation layer. Below Personal Data Dividends as creator remuneration with clear attribution. Three complementary layers that together create broad ownership of the century's most important productive force.
Switzerland and the EU should invest jointly in EuroHPC, Mistral, and BLOOM, with Swiss compute participation via CSCS Lugano. The goal is foundation models that are state-controlled, democratically governed and open-source, at minimum for public administration, science and education. This does not replace private providers; it ensures that critical public functions do not depend 100% on US hyperscalers or Chinese champions.
MIDATA needs to scale. Today MIDATA has around 12,000 active members in Switzerland in the health sector. That is a substantial precursor, but too small. With federal funding and a clear legal framework (a Swiss data-trust act), scaling to 500,000–1m members within five to seven years is realistic. Sector expansion into mobility, energy, consumption and education would follow. The Swiss data-cooperative model could then be licensed across Europe.
Personal Data Dividends are methodologically the cleanest approach, yet practically the most difficult. Microsoft Research, Anthropic Constitutional AI and Stanford CISAC are working on attribution methods. Extending the EU AI Act to include a creator-compensation obligation for foundation models with more than 100bn parameters would be the political lever. Switzerland could set an example here with a GDPR-2.0 equivalent, though not drive it.
Cyber-sovereignty is not cyber-isolation. We reject any attempt to remodel the Swiss approach along the lines of the Chinese data-sovereignty sphere. The Swiss position: sovereign infrastructure (CSCS, MIDATA) combined with an open scientific commons (BLOOM, Mistral) and bilateral cooperation with EU partners. Data may leave Switzerland for legitimate purposes — research, commerce — but critical public functions run onshore.
What this means for investors: First, when open models keep pace with closed ones, the price per task falls and margin shifts from the models to the applications and to the owners of good data. Second, data trusts and data cooperatives could weaken platforms' negotiating power over user data. Third, European computing and model infrastructure is receiving public funding. Suppliers of that infrastructure benefit regardless of which model prevails. Fourth, a share of data revenues for the wider population remains politically unresolved. Were that to materialise, it would add to costs for data-driven business models.
Why now: The next wave of foundation models (multimodal, agentic, with tool use) will be decisive for the structural distribution of ownership. Whoever, over the next three to five years, not catches up, it becomes a recipient in the next generation, not a contributor. The EU has laid the foundation with the AI Act and EuroHPC; what is missing is a coherent investment strategy. Switzerland can help shape this substantially, if it chooses to.
The strongest objection is the cost. Training a frontier model today requires sums that no European research budget can raise annually, and the gap between open and closed models has recently widened rather than narrowed. Shared models therefore risk being permanently the second-best choice. Nobody uses second-best infrastructure voluntarily. The data dividend has its own problem: the contribution of individual data to a model's value cannot yet be technically attributed. Without attribution, the dividend is a flat payment and therefore a tax under another name.
What we do not yet know
- How can EU-sovereign AI be balanced against the need to keep pace with US frontier models?
- Who defines the attribution logic for Personal Data Dividends? And how do we prevent it from becoming a gaming exercise?
- Should Switzerland become an EU member of EuroHPC or a bilateral partner with an independent CSCS programme?
- How do we prevent cyber sovereignty from tipping into cyber isolation? How do we balance between Chinese sovereign logic and US platform openness?
Connected to: 03 · Value Creation, 06 · Democracy, 09 · Ecology
Scenarios: Technology & AI · Economics & Capital · Geopolitics & Governance
