Ryan Orbuch, who writes from Lowercarbon Capital, a climate-focused venture firm, has published a 125 minute essay arguing that AI labs are building the wrong foundation model. Large language models learn from the record of human output: books, code, the internet. Orbuch’s proposal, which he calls Natural General Intelligence (NGI), would instead train on direct observations of the Earth system itself: the atmosphere, oceans, ice, soil chemistry, and biosphere, and how they interact.
The argument starts from a limit no amount of compute removes. Natural resources, Orbuch writes, are not produced by human intelligence and cannot be automated into abundance the way knowledge work can. The Earth system that produces them already runs on its own feedback loops, with or without us. Fish reproduce, nutrients cycle through soil, water moves through a watershed. Smarter models do not replace those processes; they can only help people steer them, which he frames as stewardship rather than automation.
Orbuch backs the urgency case with recent events rather than projections. A glacier collapse in Nepal in August killed hundreds, he notes, and a summer heat wave across Europe killed at least 35,000 people and forced France to take nuclear plants offline as rivers ran too warm to cool them. A screwworm parasite once eradicated from Texas cattle has reappeared there, and warm water building in the Pacific could, by his account, produce one of the largest El Nino events on record this winter. These are illustrations of his broader claim, not data the essay independently verifies.
The essay’s most concrete contribution is a survey of what already exists. Weather and climate modeling has pushed supercomputing limits for decades, and Orbuch notes that current AI weather systems, including NVIDIA’s Earth-2, Microsoft’s Aurora, and Google’s WeatherNext, mostly do not train directly on raw observations. Instead they learn from “reanalysis” data: reconstructions already produced by a physics-based simulation. He treats that as a gap NGI would need to close by training closer to the raw sensor record.
He also quantifies the scale of the raw material available. The global Earth-observing network, by his estimate, produces roughly 40 petabytes of data per year, arriving through satellite downlinks, automated ground networks, and field campaigns. Daily readings run into the hundreds of millions, he writes, but most of the atmosphere still escapes measurement at any given moment. That gap between data volume and actual coverage is the practical bottleneck the essay spends much of its length on, more than compute or model architecture.
This is a self-published argument from someone associated with a fund that invests in climate technology, not a peer-reviewed research paper or a lab’s technical report, and it should be read that way. Orbuch names no timeline, no budget, and no organization currently building NGI as he describes it; the essay closes by inviting readers and AI agents to email him with proposed contributions. What distinguishes this from a generic call to take climate seriously is the specificity of the data infrastructure survey, which draws on citizen-science networks like eBird, permanent seismic stations, and acoustic monitoring projects such as HARK. Whether that survey adds up to a buildable foundation model or a well-researched thought experiment is exactly the question Orbuch has left open.
For AI infrastructure investors, the essay is worth reading as a signal of where climate-adjacent capital wants frontier labs to point their next large training run, not as evidence that such a model is already underway.
Ryan Orbuch published this essay, “Natural General Intelligence,” on naturalgeneralintelligence.ai in September 2026.
