A Stanford-founded startup says it has built an AI weather forecasting system that outperforms the world’s leading government weather agencies. WindBorne Systems announced the release of WeatherMesh 6, which the company claims is more accurate than both traditional and AI-powered forecasts from the European Centre for Medium-Range Weather Forecasting (ECMWF).
The breakthrough comes at a time when AI is reshaping weather prediction across the industry. Major tech companies like Google DeepMind and government agencies worldwide are racing to integrate machine learning into their forecasting systems, but most still rely on data processed by traditional physics models running on supercomputers.
WindBorne’s advantage comes from its unique approach of combining model development with direct data collection. The company operates about 400 weather balloons simultaneously, launched from 15 sites around the globe, feeding sensor readings directly into its AI models rather than relying solely on government-processed data.
“WeatherMesh 6 is as accurate five days out as a traditional forecast is the day before,” says Kai Marshland, WindBorne’s chief product officer, particularly for surface temperature measurements. The model produces hourly forecasts compared to the standard six-hour intervals of traditional systems, with resolution down to 3 kilometers in Europe and the continental United States.
Traditional weather forecasting relies on complex physics models that require expensive supercomputers and significant processing time. AI models typically run faster but have historically struggled with resolution, variable complexity, and long-term accuracy. However, the technology is improving rapidly and gaining adoption at government agencies worldwide.
WindBorne’s breakthrough centers on data assimilation – the process of converting diverse sensor readings into a comprehensive, machine-readable picture of atmospheric conditions. This has long been the ECMWF’s strength, which is why most AI weather models depend on datasets produced by the European organization and the US National Oceanic and Atmospheric Administration.
Joan Creus-Costa, WindBorne’s head of AI, says direct data ingestion from their balloons and other sources drives the improvements in WeatherMesh 6. The company spent a year tuning and re-architecting its transformer-based model to deliver these forecasts while maintaining stability.
“When we started doing data assimilation we were still very heavily reliant on ECMWF,” CEO John Dean explained. “I predict today, if we removed ECMWF’s initial conditions, we would actually still do pretty good.”
The company faced challenges last year when a United Airlines aircraft collided with one of its balloons, causing minor damage but no injuries. WindBorne has since added transponders to its balloons that report locations through the global aviation surveillance system ADS-B to prevent future incidents.
Founded by Stanford students in 2019, WindBorne initially focused on building better weather balloons to sell data. The emergence of weather forecasting AI in 2022 prompted the team to develop their own models to capture more value from their data collection efforts.
WindBorne has raised $25 million in venture funding with a reported $85 million valuation in 2024. The company sells balloon data to NOAA for use in American weather forecasting and to the US Air Force and Navy. It also provides forecasts to investors and commodity traders, though Dean says the focus remains on building model and data infrastructure rather than commercial products.
“I’m not trying to invest a massive team into building a SaaS product, if the way people want consumer information two years from now is through an agent,” Dean noted, reflecting the rapidly changing information landscape driven by AI advancement.
