A startup with weather balloons in the sky just beat the gold standard of forecasting. WindBorne Systems released the sixth version of its AI weather model today, and according to TechCrunch AI, the company says it now produces more accurate and more frequent predictions on key variables than the system run by the European Centre for Medium-Range Weather Forecasts (ECMWF). For meteorologists, ECMWF is the benchmark. Beating it is a real claim.
WindBorne started in 2019 as a group of Stanford students building a better weather balloon. The plan was to sell the data. When deep learning weather models arrived in 2022, the team decided to build their own model too. That model is WeatherMesh, and version six is what shipped today.
What’s Actually New
The headline number is striking. WindBorne’s chief product officer Kai Marshland told TechCrunch AI that WeatherMesh-6 “is as accurate five days out as a traditional forecast is the day before,” especially on surface temperature. A few more specifics from the report:
- It generates a new forecast every hour, versus every six hours for traditional models.
- Resolution is down to 3 km across Europe and the continental U.S., where data quality is best.
- The gains come from feeding balloon sensor data directly into the model, not from a bigger model alone.
That last point is the one to watch. WindBorne flies about 400 balloons at any given time, launched from 15 sites worldwide. The company spent roughly a year re-architecting its transformer-based model to ingest that raw data without losing stability.
Why This Matters
Traditional forecasting leans on physics models that need expensive supercomputers and a lot of time to run. AI models move faster but, until recently, traded away resolution and long-range accuracy. WindBorne is arguing it can have both, and the lever is data, not just compute.
What stands out here is the moat. ECMWF’s edge has always come from “data assimilation,” the hard work of turning scattered sensor readings into one clean machine-readable picture of the atmosphere. Most AI weather models, including ones from labs like Google DeepMind, still depend on starting conditions produced by ECMWF and NOAA. WindBorne is trying to cut that cord.
“I don’t understand, personally, the business model of being an AI based weather company without a dataset advantage,” CEO John Dean told TechCrunch AI. He went further on independence: “I predict today, if we removed ECMWF’s initial conditions, we would actually still do pretty good.”
The Business Behind It
WindBorne has raised $25 million in venture funding at a reported $85 million valuation in 2024. The revenue picture, per TechCrunch AI:
- It sells balloon data to NOAA, where it feeds the U.S. forecasting system, plus the U.S. Air Force and Navy.
- It sells forecasts to investors and commodity traders.
- It is deliberately not racing to build a polished SaaS product.
Dean’s reasoning is worth noting for anyone building consumer tools right now. “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, right?” he said. That’s a bet that distribution itself is about to change.
There’s a safety footnote too. Last year a United Airlines jet hit one of WindBorne’s balloons. The plane took minor damage and no one was hurt, helped by the company keeping its sensor package within U.S. size limits. WindBorne now monitors air traffic via the ADS-B surveillance system and maneuvers balloons out of the way.
The bigger signal is direction of travel. Weather AI is improving fast and already running inside government agencies. If a 400-balloon startup can out-forecast the world’s leading public model on some measures, the question shifts from whether AI forecasting works to who owns the data feeding it. Full details are in the original TechCrunch AI report.