Early May brought a quiet landmark: a joint ESA-NASA workshop on foundation models for Earth observation, where the two agencies compared notes on a question neither could answer alone — how to build frontier AI for planetary-scale data.
The framing that emerged is worth keeping: Earth observation as a proving ground for AI. The data is genuinely hard — multispectral, multisensor, noisy, seasonal, biased in ways that are hard to see. A model that performs well on EO imagery has survived a stress test most benchmarks cannot offer.
The agencies face the same practical wall: petabytes of open data, a small number of labeled examples, and user questions that arrive as sentences rather than queries. Foundation models — TerraMind on the European side, NASA's own efforts on the other — are the shared answer.
What makes this collaborative rather than competitive is that both agencies publish. Open weights, open datasets, open benchmarks — a rare configuration in the AI race, where the stakes are scientific credibility rather than market share.
For engineers watching from outside, EO is the doorway into applied AI that actually matters: real sensors, real constraints, measurable public benefit. The workshop's real output was not a paper but a network — the people building the layer between satellites and decisions, now talking to each other across the Atlantic.
It is also worth noting what makes EO unique as a testbed: failure is visible and expensive in the real world, not just in benchmarks. A flood model that underestimates an inundation area misallocates emergency resources — an environment that enforces honest evaluation in an industry sometimes accused of grading its own homework.