Jua
Jua, a Zurich start-up founded in 2022, builds an AI weather model, EPT-2, that it says forecasts more accurately and far more cheaply than conventional numerical weather prediction.
With 51 to 200 staff, it announced an 11 million US dollar Series A in June 2025; TechCrunch had earlier reported 16 million raised. It markets 20-day forecasts with hourly updates for flight planning, among other uses, and claims to beat Microsoft's and Google's weather models. No aviation customer is named.
No verified deployments yet
Jua is profiled in the companies register, but no deployment has been verified against published evidence. Records are added as evidence is reviewed and graded.
Products & Solutions
2 productsEPT-2 (Earth Physics Transformer Foundation Model)
A Large Physics Model AI weather foundation model that maps sequences of past atmospheric states to future conditions using an end-to-end transformer architecture. Trained on petabytes of observational data including satellite imagery, weather stations, and ERA5 reanalysis.
Traditional numerical weather prediction models require billion-dollar supercomputers and produce forecasts with limited resolution and temporal frequency. EPT-2 delivers equivalent or superior accuracy at a fraction of the compute cost, with hourly updates instead of 4x-daily, enabling faster and more precise decision-making for weather-dependent industries.
- Native hourly global forecasts (0-240h horizon)
- 0.083-degree spatial resolution (~9x9 km at equator)
- 30-member probabilistic ensemble (EPT-2e)
- Rapid Refresh variant updated 24 times per day via satellite assimilation
6 more features
- EPT2-HRRR: 5.5 km resolution for Europe updated hourly
- Any-lead-time forecasting via dynamic lead time conditioning
- Outperforms ECMWF HRES and Microsoft Aurora on wind speed and temperature benchmarks
- 25% faster inference than Aurora
- Trained using 75% less compute than Aurora
- Output available via API (JSON, CSV, GRIB), Python SDK, and web platform
Technical specifications (11)
- Model family: Earth Physics Transformer (EPT-2, EPT-2e, EPT2-RR, EPT2-HRRR)
- Training data: petabytes of satellite imagery, weather stations, topography, ERA5 reanalysis
- Global spatial resolution: 0.083 degrees (~9 km at equator)
- EPT2-HRRR resolution: 5.5 km (Europe)
- Temporal resolution: native 1-hour forecasts
- Forecast horizon: 0-240h standard; up to 60 days for EPT-2e
- EPT-2e: 30-member ensemble via perturbation of initial conditions
- Rapid Refresh: 24 updates/day via direct satellite observation assimilation
- Key variables: 10m wind speed, 100m wind speed, 2m temperature, surface solar radiation, geopotential, precipitation
- Training hardware: 8x NVIDIA H100 GPUs for 10 days (pre-training)
- Technical report: arXiv 2507.09703 (Jul 2025)
Earth Intelligence Platform
A web-based decision platform and API that provides access to Jua's proprietary EPT models alongside third-party weather models (ECMWF IFS, NOAA GFS, Microsoft Aurora).
Energy traders and grid operators previously relied on fragmented data sources and manual processes to interpret weather model outputs for trading decisions. The platform consolidates multiple weather models in a single interface with integrated analytics, automated workflows, and AI-assisted querying.
- Access to all Jua EPT models and major third-party models in one interface
- Power Trading Analytics (beta): weather-based energy production forecasts and capacity analysis
- Oracle: natural language query interface for custom trading and weather use cases
- Dynamic global weather maps and point forecast analysis
3 more features
- Athena AI analyst for automated model benchmarking and portfolio monitoring
- Usage-based API pricing with Python SDK
- Supports automated trading decision workflows
Technical specifications (5)
- Access modes: web platform (app.jua.ai), REST API, Python SDK
- API output formats: JSON, CSV, GRIB, map tiles
- Integrated models: EPT-2, EPT-2e, EPT2-RR, EPT2-HRRR, ECMWF AIFS, ECMWF IFS, NOAA GFS, Microsoft Aurora
- Power Trading Analytics: weighted energy forecasts, renewable capacity insights (beta)
- Oracle: natural language interface for custom query and workflow automation
Register notes
Notes held in the Aerostack register. Aerostack has not independently verified them.
Jua is building Large Physics Models (LPMs), AI systems trained to simulate the physical world, starting with weather.
Read the rest
Its EPT-2.0 model benchmarks ahead of competitors from Microsoft, Google DeepMind, and ECMWF on standard accuracy metrics while using 75% less compute.
The primary market is energy trading (solar and wind generation forecasting) but the technology has direct aviation applications: weather routing, turbulence prediction, and approach planning.
Jua's European-sovereign positioning (Swiss base, European investors, ECMWF scientific heritage) is a differentiator for European aviation authorities and operators wary of US big tech dependency.
Key competitors in AI weather: Google DeepMind (GraphCast), Microsoft (Aurora), The Weather Company (IBM), and Tomorrow.io.
Profile data from the Aerostack companies register; deployment records from the verified register. Evidence grades: T1 first-party confirmation, T2 credible reporting. See Our approach for how records are graded.