For decades, making sense of satellite imagery meant one of two things: a trained human analyst staring at a screen, or a narrowly focused algorithm built to recognize exactly one thing. One could scale. One could generalize. Neither could do both. We've been sitting on over 200 petabytes of Earth observation data — and until recently, the vast majority of it has been effectively invisible.

In the data world, there's a term for this: dark data. Studies suggest that upwards of 80% of enterprise data is collected but never analyzed. It just sits there, inert. Earth observation may represent the single largest dark data problem in history — petabyte upon petabyte of detailed imagery of our planet, collected at great expense, and largely unread.

Here's how we got here, and how LGND is tackling the challenge.

Phase 1: The Human Era

A skilled geospatial analyst can look at a satellite image and immediately understand it. They see an airport — the runways, the terminals, the construction underway on the north side, the bay on one edge and the forest on the other. They can write a report, build a structured table, flag what's changed since the last pass. The human brain is an extraordinary multi-modal processor.

But it doesn't scale. An analyst can review a few thousand images a day. Satellites generate far more than that.

The semantic parallel here is instructive. For centuries, this is how libraries worked. Humans cataloged books by title, subject, and category. The Dewey Decimal System is elegant — but it's only as good as the person who wrote the label. You were searching metadata about knowledge, not knowledge itself.

Phase 2: Machine Learning and the Rise of Computer Vision

The last 25 years brought powerful new tools. On the semantic side, Google transformed how we find information. The key innovation wasn't just the keyword — it was PageRank: the insight that a page's importance could be measured by how many other pages linked to it. Google wasn't just indexing content; it was understanding relationships between documents. We went from searching titles to searching meaning — and the web became navigable at scale.

In geospatial, we saw an analogous leap with Convolutional Neural Networks (CNNs) or Computer Vision, which introduced new efficiencies in the analyst workflow. Engineers can feed a CNN hundreds of thousands of images of airports, reward it when it's right, correct it when it's wrong — and eventually the CNN learns to recognize airports by understanding what an airport looks like from space: the arrangement of pixels that match runways, hangars, taxiways, and towers. It's reading pixels, not metadata.

This leap was crystallized publicly in 2012, when a CNN called AlexNet won the ImageNet computer vision competition by a margin so large it shocked the research community. The deep learning era had begun. In the geospatial world, we started to have our own ImageNet moment — machines that could see.

But CNNs face  a fundamental constraint: a model trained to find airports finds airports. That's it. Want to find ports? Train a new model. Oil storage tanks? Another model. Data center cooling infrastructure? Another. Each one requires months of work from a specialized team. You've traded the human analyst's generality for the machine's speed — but you're still stuck in a world of single-purpose tools that can only do what you explicitly trained them to do.

Phase 3: Foundation Models and the Geography of Everything

Three years ago, with the release of ChatGPT, the semantic world crossed a threshold. The primary unit of knowledge shifted from the keyword to something called a vector embedding — a dense mathematical representation of meaning, not just words. The shift is profound: instead of matching strings of text, you're matching concepts.

Here's an analogy that makes this concrete. Imagine compressing every book ever written into a single vast library — but instead of organizing books alphabetically or by genre, books are shelved by conceptual closeness. Books about similar ideas sit near each other. Ask a question, and you don't get a list of titles — you get pointed to the shelf where the answer lives. That's what a vector embedding space is. Proximity means similarity. Distance means difference. And the whole of human knowledge becomes instantly traversable.

You stopped getting ten blue links and started getting answers. Language models became general-purpose: one model capable of reasoning across everything.

We are now at that exact inflection point in geospatial.

Large Earth Observation Models: The GeoAI Moment

Large Earth Observation Models (LEOMs) — like Clay, which we built and maintain at LGND — apply the same foundation model approach to Earth imagery (satellite, aerial, drone, any top down view). Instead of training a separate model for every object type, a LEOM learns a rich, general-purpose representation of the Earth's surface. It produces geo-embeddings: the spatial equivalent of language embeddings, encoding not just what a pixel looks like, but what it means in context.

The result is a single model that can find airports, identify suitable data center locations, flag vegetation creating wildfire risk, or detect lithium-bearing geology — all without retraining, all from the same underlying model, in seconds.

But here's what makes this era truly different from what came before: because geo-embeddings are numerical vectors, they are in the same language of this new AI paradigm shift. You can ask in plain English — "show me areas that look like the Australian outback but are in South America" — and the model understands both the semantic and visual intent simultaneously. The bridge between language intelligence and spatial intelligence is being built.

And there's one more dimension to this that deserves emphasis: history. Satellite programs like Landsat imagery go back decades. Much of that data was only ever accessible to governments with large teams of analysts. It contains intelligence we've never fully extracted — changes to coastlines, urban growth, agricultural shifts, environmental signals. LEOMs can now read backwards through time. The unlocking isn't just of the present Earth, but of the historical record of our planet.

This isn't searching a label a human attached to an image. It's reading the image itself — at planetary scale, across decades of history.

What We're Building at LGND

We think of LGND as the search layer for the Earth. Not unlike how Langchain let’s you easily put any LLM to work in your application, or how Perplexity sits on top of multiple language models to make them useful and accessible — we're building the infrastructure and applications that make LEOMs actionable for analysts, enterprises, and AI agents. We work across models and imagery sources – happy work with any of the amazing satellite, aerial, and other imagery creators out there – so the intelligence compounds as the underlying technology improves.

The 200+ petabytes of Earth observation data we've collected — that dark data sitting largely unread — is about to become searchable. Queryable. Alive.

The ability to query our planet the way we query LLM’s is now becoming possible. We're at the beginning of it. And we can't wait to show you what you can do with it.