DeepSky Explained: How Tomorrow.io Is Building the First AI-Native Weather Satellite Constellation

DeepSky is Tomorrow.io's next-generation low-Earth-orbit weather satellite constellation, and it represents one of the clearer architectural step-changes in commercial space-based Earth observation in 2026. The shorthand framing — 'an AI-native weather satellite constellation' — is more than a marketing label: it captures a deliberate set of architectural choices that are derived from what AI forecasting models need to ingest in order to deliver skill improvements over legacy numerical weather prediction. Understanding why DeepSky is structured the way it is requires walking through both the data needs of modern AI weather models and the limits of the satellite infrastructure that currently supplies atmospheric observations to those models.

The Observation Bottleneck for AI Weather Models

AI weather forecasting has advanced rapidly over the past several years. Foundation-model architectures trained on global reanalysis datasets have demonstrated forecasting skill that matches or exceeds traditional physics-based numerical weather prediction (NWP) models in several skill categories — most notably medium-range forecasts of large-scale atmospheric variables. But AI model performance increasingly hits a ceiling that is not architectural — it is observational. AI models can only learn from the observations they consume during training, and they can only forecast accurately for the regions, altitudes, and time scales for which observation density is sufficient. The data-sparse parts of the atmosphere — over oceans, over polar regions, at upper-tropospheric and stratospheric altitudes, and at short refresh cycles — are exactly where AI model performance degrades, and those are exactly the regions and time scales that legacy satellite infrastructure underserves.

Legacy government weather satellites were architected for a different consumer. NOAA's GOES geostationary platforms, the JPSS polar-orbiting platforms, the EUMETSAT MetOp series, and equivalent systems internationally are designed around large, expensive, individually customized spacecraft launched on multi-decade refresh cycles, with a relatively small number of high-capability instruments per platform optimized for a specific set of channels and applications. That architecture has served NWP and operational forecasting well for decades. It does not serve AI-native forecasting well — AI models want larger numbers of smaller, multi-sensor satellites, refreshed on shorter cycles, providing higher revisit rates and denser observation coverage over the regions that drive forecast skill. The architectural mismatch is the bottleneck DeepSky is engineered to relieve.

Gen1: The Foundation

Tomorrow.io's first-generation constellation — Gen1 — was structured as a deliberate first step toward solving the observation bottleneck. Gen1 satellites are 6U-class cubesats carrying microwave sounder instruments derived from the same instrument class used on flagship government weather satellites. Microwave soundings are particularly valuable for AI forecasting because they observe atmospheric temperature and humidity profiles through cloud cover (where infrared sounders cannot), and they provide the atmospheric vertical structure information that AI models rely on heavily. Tomorrow.io has launched 13 Gen1 satellites to date, of which 11 are operational microwave sounders, and the constellation has reached a 60-minute global revisit rate for microwave atmospheric soundings — a meaningful improvement over the multi-hour revisit cycles of the legacy polar-orbiting platforms. Reported observation-to-product latency is approximately 30 seconds, with meter-level horizontal accuracy on relevant atmospheric profile products.

Gen1 served two strategic purposes. First, it gave Tomorrow.io operational satellite infrastructure to feed its AI forecasting platform and demonstrated that proprietary microwave sounder observations meaningfully improve model skill in customer-relevant verticals. Second, it gave the company end-to-end satellite design, build, launch, ops, and data-processing experience, and a deployed asset base that anchors the company's claim to be a vertically integrated weather intelligence operator rather than a software-only consumer of third-party data. With Gen1 fully deployed, the strategic question shifted from 'can we operate satellites' to 'what is the constellation architecture that maximizes AI forecasting skill,' and the answer is DeepSky.

DeepSky Architecture: Car-Sized, Multi-Sensor, Many Satellites

DeepSky represents a step-change in nearly every architectural dimension. CSO Rei Goffer has described the DeepSky satellites as 'closer to the size of a car' — meaningfully larger than the 6U cubesats of Gen1 and approaching the scale of a small commercial bus. The size increase enables the headline DeepSky innovation: 3–5 co-located sensors per satellite, rather than the single instrument per cubesat of Gen1. Co-location is the architecturally important property — when multiple sensors observe the same atmospheric column at the same time, the observations are physically consistent in a way that observations gathered from different satellites and different times cannot be. Multi-sensor satellite designs are the standard architecture for flagship government weather satellites for exactly this reason, and DeepSky brings that architectural pattern into the commercial constellation cohort at multi-satellite scale.

The expected DeepSky sensor mix has not been fully disclosed in detail, but the category logic is straightforward. Microwave sounders (the Gen1 anchor instrument) will continue to be central — they provide the all-weather atmospheric profile observations that AI models depend on most heavily. Infrared sounders and imagers complement microwave by providing higher-vertical-resolution observations in cloud-free conditions and supporting cloud-top characterization. Radio occultation (GNSS-RO) receivers provide independent atmospheric profile observations with strong vertical resolution and global coverage. Additional candidate sensors include scatterometry for ocean wind observation, lightning detection, water vapor radiometry, and trace-gas observation. The DeepSky 3–5-sensor design allows Tomorrow.io to assemble sensor packages that fill specific observation gaps where AI model skill is currently most constrained, rather than launching single-instrument cubesats that contribute only one observation type at a time.

Constellation Scale and Revisit Rate

Tomorrow.io has indicated that DeepSky will scale to 'dozens' of satellites at full deployment, with the buildout targeted for completion by the end of the decade. The constellation scale is calibrated to deliver the revisit rates and observation density that AI forecasting models need. Combining multi-sensor satellites with high constellation count produces an aggregate observation density that legacy single-satellite architectures cannot match. A microwave sounder on a DeepSky satellite delivers one observation per satellite per orbit per column observed; a constellation of dozens of DeepSky satellites at appropriate orbital geometry collectively delivers many observations per atmospheric column per hour. That is the data density that an AI weather model can convert into forecast skill, particularly for regional and short-range forecasts where the marginal observation matters most.

Importantly, the revisit rate that DeepSky targets is not equally valuable across all atmospheric variables. Some variables — surface temperature, sea surface temperature, broad-scale circulation patterns — are well-observed already and benefit only modestly from additional satellite revisit. Other variables — short-duration convective systems, severe storm initiation, atmospheric river dynamics, tropical cyclone intensification — are dramatically underserved by current revisit cadences and are exactly the categories where DeepSky's high-revisit multi-sensor architecture should deliver outsized skill improvements. Severe weather events, in particular, are both the highest-economic-stakes forecasts (driving insurance losses, aviation diversions, energy demand spikes, emergency management decisions) and the most observation-limited under current infrastructure. DeepSky's commercial value capture is anchored in exactly that intersection.

Why Tomorrow.io Owns the Constellation

A natural question for a software-anchored AI weather company is why it would invest the capital required to design, build, launch, and operate its own satellite constellation rather than purchasing observations from third-party operators (Spire, PlanetiQ, GeoOptics) and focusing on the AI software layer. Tomorrow.io's strategic answer is that the satellite architecture is the differentiator — if AI forecasting skill is constrained by observation density and diversity, then the company that controls the observation architecture controls the model performance ceiling, and the company that depends on third-party observation supply is permanently capped by what third-party operators choose to provide. DeepSky's specific design choices — instrument selection, orbital geometry, revisit cadence — are derived from what Tomorrow.io's AI models need, not from what third-party operators have chosen to deploy. Vertical integration of the observation network and the AI software layer is the architectural bet, and the $210 million Series F is the capital that operationalizes the bet at scale.

Frequently Asked Questions