Geospatial Foundation Models and the Case for Sovereignty

A few weeks ago, my colleague Miguel Álvarez published a summary of our Geospatial Foundation Models Workshop with the Barcelona Supercomputing Center. That post covered the technical landscape — the models, the architectures, the open challenges. But there’s a question that kept coming up in the room that deserves its own conversation: who controls these models, and where do they run?

The field has moved fast from “can we build geospatial foundation models?” to “we’re building them.” Now the question is whether countries, regions, and organizations will have meaningful control over the geospatial AI that shapes their decisions — or whether they’ll depend entirely on a handful of foreign providers.

This is the sovereignty question, and it matters more than most people in our industry realize.

Geospatial Data Is Critical Infrastructure

Geospatial data underpins decisions in agriculture, urban planning, defense, disaster response, climate adaptation, and public health. When a government monitors land use changes, models flood risk, or plans transportation networks, it relies on geospatial intelligence.

As foundation models become the dominant way to extract insights from satellite imagery, mobility data, and environmental sensors, whoever controls those models gains enormous influence over the decisions that flow from them. A country that depends entirely on foreign-controlled AI models for its geospatial intelligence faces a strategic vulnerability — not unlike depending on a single foreign supplier for energy or semiconductors.

This isn’t hypothetical. The debates around cloud sovereignty and LLM sovereignty are already well underway in Europe, Asia, and Latin America. Geospatial AI is simply the next front.

The Current Landscape: Concentration Risk

Today, the most visible geospatial foundation models come from a narrow set of actors. Google’s Population Dynamics Foundation Model and similar efforts from major US tech companies represent significant investment and capability. US-funded initiatives like Clay have made important contributions to open geospatial AI. This work is valuable and has pushed the entire field forward.

But it also creates concentration risk. When a small number of organizations — mostly headquartered in the United States — train the models that the rest of the world uses to understand its own territory, there’s an asymmetry worth examining.

Regions with unique geospatial challenges may not be well served by models trained primarily on data from the Global North. The agricultural patterns of sub-Saharan Africa, the urban density of Southeast Asian cities, the permafrost dynamics of Scandinavia — these require models that understand local context. A foundation model trained overwhelmingly on North American satellite imagery may perform well on North American landscapes and poorly everywhere else.

This isn’t a criticism of the teams building these models. It’s a structural observation: if you want geospatial AI that works for your region, you need to be involved in building it.

Open Models as the Path to Sovereignty

This is where open-weight and open-source models change the equation. When a geospatial foundation model is released with open weights — like Clay — any organization can download it, host it on their own infrastructure, fine-tune it on their own data, and audit its behavior. You don’t need to send sensitive satellite imagery to a foreign API. You don’t need to trust a provider’s claims about data residency. You control the stack.

Contrast this with proprietary, API-only models. If the provider changes pricing, restricts access, or is affected by export controls or sanctions, your entire geospatial AI pipeline is at risk. You can’t inspect how the model handles your data. You can’t adapt it to local conditions without the provider’s cooperation.

Open doesn’t mean ungoverned. Sovereign geospatial AI still requires evaluation frameworks, quality benchmarks, and responsible deployment practices. But openness gives you the option to govern it yourself, which is the entire point of sovereignty.

Our collaboration with the Barcelona Supercomputing Center through the European AI Factory program is a concrete example of what this looks like in practice. European compute infrastructure, European research institutions, and open models — combined to build geospatial AI capabilities that Europe controls.

What Sovereign Geospatial AI Infrastructure Looks Like

If a country or region wants genuine sovereignty over its geospatial AI, it needs several things working together:

Local compute. Models need to be trained and served on infrastructure within your jurisdiction. The EU’s investment in supercomputing centers like BSC, and national cloud initiatives across Europe and Asia, are building this foundation.

Models trained on local data. The Copernicus programme gives Europe an extraordinary advantage — decades of openly available Earth observation data covering the entire continent. National mapping agencies, cadastral systems, and environmental monitoring networks provide additional training data that no foreign provider has access to. These assets should be used to build and fine-tune foundation models, not just consumed through foreign platforms.

Interoperability with existing workflows. A sovereign model is only useful if it integrates with the tools organizations already use for spatial analysis, visualization, and decision-making. This is where platforms like CARTO play a role — connecting foundation model outputs to the analytical workflows where decisions actually happen, regardless of where the model runs.

Community and ecosystem. Sovereignty isn’t just infrastructure — it’s people. Researchers, engineers, and practitioners need spaces to collaborate, share benchmarks, and build on each other’s work. Events like the Spatial Data Science Conference and workshops like the one we hosted with BSC are part of building that ecosystem.

A Conversation the Geospatial Community Needs to Have

The window for shaping how geospatial foundation models develop is open now. The architectures aren’t settled. The data pipelines aren’t locked in. The governance frameworks are being written. This is the moment for governments, research institutions, and industry to make deliberate choices about sovereignty rather than drifting into dependency by default.

At CARTO, we believe the future of geospatial AI should be open, interoperable, and distributed — not locked behind a single provider’s API. That’s why we’re investing in partnerships like the one with BSC, contributing to open benchmarks, and building tools that work with any model, anywhere.

If you care about this topic, join us at the Spatial Data Science Conference in London on May 14, 2026. These conversations are better in person, and the decisions we make together over the next few years will shape geospatial AI for decades.