Why We're Building an LGM

Columbus is an applied research company. We're looking for researchers who want to help define the next chapter in foundation models. Join us.

Fundamentally, we're building foundation AI to understand the context and relationships of the physical world. If LLMs understand the context of language, Columbus understands the context of Earth.

By World Bank estimates, roughly 85% of global GDP activity happens in the physical world through logistics, construction, energy, transportation, agriculture, resource extraction and more. But the last few years of AI progress have been largely about the digital one: language, code, and images. The single largest domain of human activity is still waiting for its foundation models. We think that's the opportunity of the decade, and the world is realizing this. 2026 is being called the year of the "world model," with well-respected researchers now arguing that the next frontier of AI will combine language models, world/physics models, and geospatial foundation models into a single stack.

Why an LGM specifically, and why us?

1. Because the existing models don't work for this.

We converged on the LGM while trying to build a simple and powerful system that could answer any question about a place and observing that off-the-shelf LLMs fall short. The failure isn't just missing data. With perfect retrieval, an LLM still processes coordinates as character strings, with no native representation of distance, adjacency, or terrain coordinates, and often lacks access to the necessary geospatial data sources. Earth-observation models see only what a satellite sees. 3D scene models are good at spatial representations, but not at meaning. The thing we needed did not exist, so we began to research and develop what would become the LGM.

2. Because a critical part is the data, and we've made it our edge.

Most of what's true about the Earth is locked in dormant, broken, or unpaired data. Our digestion engine turns that into usable, well-labeled, time-stamped structured data, cheaply and at scale. Ground truth is the thing every model in every category of physical AI is ultimately constrained by. Read how we turn dormant data, usable.

3. Because the mission is worth completing.

We want Columbus’ AI, Magellan, to be the mind of Earth, a model that reasons like the site-selection team, the geologist, the environmental scientist, the urban planner, and considers what each of them would consider to get to the truth of a place. Where should the next clinic go? Where is copper most likely under this ground? How will this ecosystem respond to that project? These are the same pattern of reasoning, applied to the same kind of data, toward the same goal: understand what the person is really asking, understand the place, reason over it, and give an answer you can act on.

We're an applied AI lab that researches many exciting angles, and we can't pursue every angle alone. But the questions here are some of the most interesting in the field, and we offer researchers the time, environment, and support to actually think.

Earth, understood. Come build it with us.