AI in Weather and Climate Prediction

Narrative

The commercial appetite for AI-driven weather data has grown sharply. Bloomberg reported in September 2025 that the number of commercial firms licensing data from ECMWF rose almost 20% in 2024 to over 800, with nearly half owned by energy companies including traders and renewable power producers. AI is explicitly driving the boom, as customers feed ECMWF data into new machine-learning tools; energy traders in Europe's increasingly wind- and solar-driven power markets were among the earliest and most aggressive adopters. By March 2026, Bloomberg was reporting that European energy traders had gone further still, using AI and machine-learning tools not merely to predict temperatures but to "forecast the forecast" itself. Google DeepMind's November 2025 WeatherNext 2, profiled by Bloomberg, was specifically designed with additional tools for energy traders, demonstrating how the largest AI labs are now commercialising weather-model research directly.

On the venture and startup side, the clearest signal of investor conviction is Tomorrow.io's February 2026 close of a $175 million equity round led by Stonecourt Capital and HarbourVest at a valuation above $1 billion, with total disclosed funding reaching roughly $500 million. The round was earmarked for DeepSky, a proliferated low-Earth-orbit satellite constellation designed to address the data-density bottleneck that constrains current AI models. Clients cited include Amazon and BNSF, whose supply-chain executives framed atmospheric data as mission-critical infrastructure on a par with any other operational input. Smaller rivals such as Atmo, which has raised around $19.7 million, focus on national-government contracts in markets such as the Philippines and Tuvalu, illustrating how reduced inference costs have opened sovereign customers that once lacked the compute to run global models.

The regulatory and institutional picture is sharply bifurcated. ECMWF went fully operational with its AI Forecasting System (AIFS Single) on 25 February 2025, and on 1 July 2025 added AIFS ENS, a 51-member ensemble, running both in parallel with the physics-based IFS. ECMWF described energy-sector pricing forecasts, insurance, security, and shipping as key beneficiaries, and the centre made near-real-time forecast data freely downloadable shortly thereafter. By contrast, NOAA in the United States saw roughly 20% of its approximately 13,000 employees cut in early 2025 under the Trump administration's DOGE programme, with proposed budget reductions of around 27% for 2026. Former NOAA officials and meteorologists warned that reduced weather-balloon launches and suspended satellite data streams were already degrading forecast inputs, while private companies including WindBorne, Tomorrow.io, and Sofar Ocean Intelligence stepped in to plug data gaps, raising governance concerns about public safety and proprietary dependency.

The broader structural story is one of rapidly falling marginal costs alongside persistent ceiling effects. Running a 10-day GraphCast forecast costs a few dollars of GPU time versus hundreds of millions of dollars of supercomputing for equivalent IFS runs, a cost compression that is democratising medium-range guidance for lower-income national meteorological services. However, Bloomberg's September 2025 analysis noted that AI models still struggle with low-frequency, high-impact extreme events, which are precisely the scenarios of greatest commercial and societal importance. ECMWF's own 2024 annual report acknowledged that the AIFS underestimated extreme rainfall partly due to coarser resolution, and later versions incorporated physical-consistency constraints to address precipitation biases. The frontier visible in 2026 is hybrid physics-ML pipelines, ML-driven data assimilation, and kilometre-scale regional models, though independent verification of vendor-reported accuracy claims remains patchy and industry observers note the gap between benchmark performance and operational reliability in out-of-distribution events.