Boston-based climate-tech firm Tomorrow.io has unveiled the instrument architecture behind DeepSky, a next-generation satellite platform designed to collect high-frequency weather observations for AI-driven forecasting. Axios reported the reveal on September 10.
Built for models, not meteorologists
Traditional weather satellites were designed around human forecasters and physics-based numerical models. DeepSky treats AI forecasting systems as the primary customer. That changes what gets measured, how often and at what latency: microwave sounders and radar payloads are tuned for cadence and coverage that a machine-learning model can ingest without the aliasing and gaps that limit today's AI-driven products.
Why cadence is now the constraint
AI weather models have jumped forward — Google DeepMind's GraphCast, Nvidia's FourCastNet, Huawei's Pangu — but their skill ceiling is set by the observations they can consume. Radar coverage is patchy over oceans, hurricane development zones and much of the Global South. Filling those gaps with a purpose-built constellation is the piece that could extend severe-weather lead times by hours, not minutes.
A pattern beyond weather
DeepSky is another example of a broader 2026 shift: companies are now building physical infrastructure specifically around AI model requirements, not the other way around. IBM and NASA did it with a lunar foundation model. Nvidia's 2 GW Australian factory did it with power. Tomorrow.io is doing it with orbital coverage.
Reporting based on coverage from Axios.