Weather has always been measured with ground instruments built for the last century: ground stations, weather balloons, government satellites designed when numerical weather prediction was the only game in town. AI changed that. But AI models need data at a density those legacy systems were never designed to deliver.
The constellation architecture shifts from single-sensor cubesats to car-sized multi-sensor satellites designed from day one to serve AI forecasting models
A new market is forming around AI weather intelligence, with startups like WindBorne and Perceptive Space and industrial players like Fujitsu all building toward the same premise: the observation layer must be rebuilt for the AI era
A new class of commercial space infrastructure is emerging to close that gap. AI-native satellite constellations purpose-built to feed the models they run alongside.
Tomorrow.io's DeepSky is the first. The Boston-based company closed a $210 million Series F in May 2026, $175 million led by Stonecourt Capital and HarbourVest in February and topped up with $35 million from Pitango. It is building what it calls the world's first AI-native space-based weather-sensing network.
The satellites are car-sized, carrying three to five co-located sensors each. Not 6U cubesats. Not single-instrument payloads. Multi-modal sensing across the electromagnetic spectrum, on every pass, over every point on Earth.
This is a re-architecture of weather observation, not an incremental upgrade.
The observation ceiling
The paradox of modern weather forecasting is that the models outrun the data. Google DeepMind's GraphCast, Huawei's Pangu, and NVIDIA's FourCastNet all demonstrated skill improvements over traditional numerical weather prediction. But their ceiling is set by the global observing system, not by model architecture.
Government weather satellites were built for an earlier era. Geostationary birds sit at 36,000 km and see the same disk every 15-30 minutes. Polar-orbiting satellites cross the same spot on Earth twice a day. Neither delivers the temporal density that transformer-based weather models need to improve beyond a certain threshold.
The constraint is no longer compute. It is observation density.
Tomorrow.io's first-generation constellation of 11 microwave sounder satellites already achieved a 60-minute global revisit rate. That is faster than any government system. DeepSky targets revisit rates measured in minutes, not hours, across multiple sensing modalities on the same orbital pass.
DeepSky architecture
The architectural shift is not just about satellite size. It is about vertical integration of the observation layer and the AI software layer.
DeepSky satellites are being designed from day one to serve AI models, not legacy numerical weather prediction workflows. That changes the instrument requirements. It changes the orbital geometry. It changes the data pipeline at every stage from photon to forecast.
Co-founder and Chief Strategy Officer Rei Goffer described the satellites to Orbital Today as carrying "multiple very-high-impact, co-located sensors" of a "completely different caliber" than the Gen1 microwave sounders. The company is not disclosing the full instrument manifest or the number of satellites planned, but the constellation is described as "proliferated," implying dozens of spacecraft at full deployment.
The Gen1 constellation cost tens of millions and relied on 6U cubesat form factors. DeepSky is a different order of capital commitment. The $210 million Series F was backed by Stonecourt Capital, HarbourVest, and Pitango. It signals that the market sees weather observation as infrastructure with a return rather than a public-good science mission.
The market map
Tomorrow.io is the largest and most capitalised player in this new market, but it is not alone.
WindBorne Systems, a Stanford-originated startup, operates around 400 AI-guided weather balloons at any given time, launched from 15 sites globally. The company's WeatherMesh model recently outperformed the European Centre for Medium-Range Weather Forecasts (ECMWF) on key variables, a benchmark that was considered the gold standard for decades. WindBorne raised $25 million and sells data to NOAA, the U.S. Air Force, and the U.S. Navy. It is engineering its balloon fleet to feed observations directly into its AI model without relying on government assimilation data.
They do not plan to stay small. Neither do their defence customers.
Perceptive Space, a Toronto-based startup, is building an AI-driven space weather platform targeting the gap between government bulletins and operator needs. The company raised $2.8 million in pre-seed funding from Panache Ventures, Metaplanet, and others, and is working on lightweight AI models that run onboard satellite edge processors. Their pitch: real-time, orbit-specific space environment intelligence rather than hemispheric advisories hours after the fact.
Fujitsu launched its Space Data Frontier research initiative in April 2025, collaborating with JAXA on explainable AI for space weather prediction. The Japanese company approaches the problem from the industrial-research side rather than the startup track. The premise is the same: existing space weather forecasting lacks the accuracy and lead times that modern operations require.
Geomagnetic storms are not the only risk. Solar flares disrupt GPS and communications. Ionospheric disturbances affect radar systems. The space environment has always been dynamic; what changed is how much infrastructure now operates within it.
Each of these companies addresses a different segment of the same structural gap. The observation layer was built for a different technological era, one defined by sparse government satellites and physics-based weather models. The era changed faster than the infrastructure could adapt. AI models that could deliver better forecasts exist today. What they lack is not algorithmic ingenuity but the density of observations to train and run on.
Dual-use by design
Tomorrow.io explicitly markets DeepSky to "defence and national security organizations." The U.S. Air Force is already a customer of the Gen1 constellation. Perceptive Space describes itself as "dual-use by design," serving commercial operators, allied defence programs, and civil science missions on the same platform. WindBorne sells its WeatherMesh forecasts to the Air Force and Navy alongside NOAA, treating government procurement as a revenue channel rather than a research grant.
The space environment does not distinguish between a commercial communications constellation and a national security payload. Neither do the companies building the observation layer to survive it.
A new infrastructure layer is emerging: purpose-built for AI, deployed in orbit, financed by institutional capital, and consumed by the same models it was designed to feed.
DeepSky first launch date: the gap between announcement and orbital deployment determines whether Tomorrow.io can hold its first-mover advantage against well-capitalised competitors
ECMWF and NOAA response to AI-native weather models: government agencies that currently set the global benchmark are also the largest customers or the largest competitors
Space weather startup funding velocity: if Perceptive Space and peers close follow-on rounds within 12-18 months, institutional capital is signalling conviction in the category beyond a single company thesis
Defence procurement of AI weather intelligence: the U.S. Air Force's Gen1 contract sets a precedent; a DeepSky-specific defence contract would validate the dual-use revenue model at scale
The cost of not knowing
The economic case for AI weather intelligence rests on an asymmetry. Weather events destroy more value annually than the entire satellite industry generates, yet the observation layer has barely changed in two decades.
A single geomagnetic storm in February 2022 cost SpaceX an estimated $100 million in Starlink satellites. The 2023 Capella Space loss of two spacecraft to a surprise solar event destroyed two of that company's on-orbit satellites. Ground-based losses from severe weather, including supply chain disruptions, grid outages, and aviation delays, run into the hundreds of billions globally each year. The gap between what is insured and what is insurable widens as climate volatility and space traffic both increase.
Tomorrow.io's customer list reflects this breadth. The U.S. Air Force uses Gen1 data for operational weather planning. BNSF Railway integrates weather intelligence into network-wide logistics decisions. Amazon's supply chain team treats atmospheric data as mission-critical infrastructure. All of them reached the ceiling of what government weather data can deliver and started paying for a commercial alternative.
If an AI-native constellation can demonstrate it predicts geomagnetic storms or severe weather windows more accurately than government bulletins, parametric insurance products for satellite operators become viable. The space insurance market is small but growing. The current premium pool is estimated at under $1 billion, and the asset base it covers is multiples of that and growing with every new constellation deployment. Better observation data shifts the risk curve for underwriters, which lowers the cost of capital for operators.
That ceiling is structural, not budgetary. Government weather satellites were designed to serve numerical weather prediction models that update on hour-long cycles. AI models can assimilate observations in near-real-time, but they need observation density that the existing satellite architecture physically cannot provide. The bottleneck moved from compute to data, and the data bottleneck is an orbital infrastructure problem.
It is Tomorrow.io's answer: purpose-built hardware, AI-native architecture, and institutional capital deployed at a scale that matches the gap it is designed to close.
The bets behind the numbers
$210 million is a bet that weather observation is infrastructure, not science. That AI models trained on proprietary data will outperform any model trained on public data alone. That the company that owns the sensor network and the AI layer together holds an unassailable advantage over one that owns only software.
Tomorrow.io was founded as ClimaCell in 2016, pivoted from mobile weather data to satellite ownership, and now serves 250 enterprise and government customers. The company was named one of TIME's 100 Most Influential Companies. It has demonstrated revenue from its Gen1 constellation. It is the scale-up, not the experiment.
The premise will be tested by deployment timelines and engineering execution. Satellite manufacturing at scale is a hard problem, and proliferated LEO constellations have a history of slipping. But the structural logic is sound. AI weather models are observation-constrained. The only way to break the constraint is to build a purpose-built observation network. Vertical integration of both layers is the correct strategy.
The ceiling is deployment timelines and engineering execution. Satellite manufacturing at scale is a hard problem, and proliferated LEO constellations have a history of slipping. Tomorrow.io has the capital and the customers. What it needs now is orbital delivery.