The brief
- Bets that AI weather needs its own data: 600+ balloons aloft at once, sampling skies over oceans that few sensors reach.
- WeatherMesh-6 (June 2026) claims up to 38% lower error than ECMWF's flagship model; independent checks are pending.
- Raised a $37M Series B in Aug 2026 (Khosla, Galvanize); over $62M in total.
Technical approach
Weeks aloft, steering by altitude
A few pounds, mostly sand ballast. Rising or sinking into different winds steers it; one flight gives about 50 profiles vs 1 for a normal balloon.
Data where there was none
Balloons cross oceans and remote regions where weather stations and normal balloons are scarce, the gaps that limit global forecasts.
WeatherMesh deep-learning models
Its AI models are trained on public weather records and fed its own balloon data at forecast time.
