AI Factory
KWeather Weather AI Factory
A system that applies weather big data to GPU AI infrastructure and AI weather prediction models to produce and distribute high-value weather intelligence tokens for industries such as autonomous driving, smart cities, and energy.
Wellbian AI
Wellbian AI builds services on top of the Weather AI Factory
Wellbian AI is a company dedicated to weather AI services. It takes the AI forecasts produced by the KWeather Weather AI Factory and the readings from the Wellbian sensor network, delivers a hyper-personal weather AI agent tuned to one person's schedule, location, and activities alongside weather AI services for industry, and grows its inference infrastructure step by step as demand is confirmed.
60 daysHyper-local forecast range (FourCastNet3)
1.5 kmKorea ultra-high-resolution grid (CorrDiff)
10 minRadar precipitation nowcast refresh (Stormscope)
3,420Measurement devices already sending data
4,000AI Factory access passes (3 years)
01Live
Hyper-personal weather AI agent
A B2C agent you start in conversation, with no sign-up. Ask 'Is hiking Bukhansan okay this weekend?' and it answers with a verdict for that activity, drawn from a 60-day forecast for your exact location.
- Activity domains: outdoor, golf, hiking, fishing, baseball games, domestic and overseas travel, fashion and beauty, air quality
- Input: text, photos (vision intent), location (GPS or address), natural time expressions ('next Friday afternoon')
- Output: verdict headline, condition chips, hourly timeline, evidence metric cards, alert banner, and voice (TTS) read-out
- Personalization: remembers your activities, heat, cold, and fine-dust sensitivity, and frequent places, and notifies you first when conditions change
- Expansion: Korean today, then English, Japanese, Chinese, and Spanish in turn
02Live · improving
60-day and nowcast forecasts on NVIDIA models
Wellbian retrains and calibrates the NVIDIA Earth-2 outputs of the KWeather Weather AI Factory at point level to produce the forecasts the agent uses.
- 60-day forecast: FourCastNet3 global 25 km / 6 h, downscaled with CorrDiff to 1.5 km / 1 h over Korea, then point-calibrated with 30 years of KWeather observations and Wellbian sensor readings. Refreshed four times a day
- Nowcast: Stormscope radar precipitation at 1 km / 10 min (2 h) and 1 km / 1 h (12 h), fused with live sensor readings to give neighborhood-level rain start and end times
- Dual initial conditions: GFS and KIM outputs are used side by side to show forecast uncertainty
- Verification: monthly hit rates against Korea Meteorological Administration observations, published (target)
03Partly live · expanding
Weather AI services
The same forecast engine that powers the agent is opened to industry and developers as services.
- Industry agents: energy (generation output), construction (schedule and work-stop decisions), retail (demand forecasting), public sector (disaster response)
- Inference API: index, forecast, and risk-index endpoints, city or point level, API-key authentication
- AI Factory access pass NFT: 4,000 passes, 3 years, pure utility for content payment (wellbian.io/products/ai-factory-pass)
- Data sales: raw and processed data are sold separately on Weather Market (weathermarket.io)
04Phased
Infrastructure expansion
Infrastructure grows only as fast as confirmed demand. Phase 1 uses GPU output from the KWeather factory; as pass sales and API revenue accumulate, Wellbian adds its own inference nodes.
- Phase 1 (now): inference outputs from the KWeather Weather AI Factory (H200, B200); agent serving on serverless infrastructure (Vercel, Supabase)
- Phase 2 (target 1H 2027): dedicated Wellbian inference nodes (B300) running the downscaling, point-calibration, and personalization models in-house, funded by the 4,000 access passes
- Phase 3 (target 2027-2028): next-generation (Vera Rubin) capacity secured jointly with the KWeather factory; city-level downscaling abroad to serve the four additional language markets
- Input expansion: Wellbian sensor network nodes (indoor and outdoor air quality and weather) folded into forecast assimilation
'Live' and target labels are as of September 2026. Target dates are confirmed by the operator before they become commitments.
The KWeather Weather AI Factory
The world's first AI factory dedicated to weather and climate
Using NVIDIA AI infrastructure, KWeather applies its weather big data to AI weather prediction models and is building the world's first global weather 'AI factory', one that produces and distributes intelligence tokens such as ultra-high-resolution global grid AI forecasts.
What is an AI factory?
Just as a refinery turns crude oil into petroleum, an AI factory is a data center that processes and standardizes diverse raw data to produce high-value intelligence tokens.
Synergy between KWeather's weather big data and AI weather prediction models
The KWeather Weather AI Factory uses the models of NVIDIA Earth-2, NVIDIA's AI weather prediction platform, to simulate everything from hourly nowcasts to 60-day long-range forecasts and to visualize them at ultra-high resolution, delivering the best possible results.
AI weather prediction models developed by global big tech companies, such as Google's GenCast and Microsoft's Aurora, are also actively used.
Weather intelligence tokens for logistics, energy, construction, insurance, and more
Weather intelligence tokens support decision-making across industries: optimal logistics routing, power generation forecasting, early disaster response, and more.
Technology partnership with NVIDIA and joint research on AI weather prediction models
Through its partnership with BNi&C, an NVIDIA partner, the KWeather Weather AI Factory runs on NVIDIA's high-performance GPUs such as the H200 and B200, and the next-generation Vera Rubin architecture is scheduled to support the project as well. KWeather is also actively engaging NVIDIA's global technology collaboration channels, including joint research on AI weather prediction models.
The AI factory process
Observation, satellite, public, and industry data go in. A GPU supercomputing cluster refines, fuses, and computes them, and out come intelligence tokens: ultra-high-resolution AI weather forecasts, climate analysis, energy output predictions, and more. Unlike the storage-centric data centers of the past, an AI factory produces intelligence.
- 1
Raw Data Mining
Ingest observation, satellite, public, and industry data
- 2
Filtering & Cleaning
Remove gaps and errors, apply quality control
- 3
Tokenization & Formatting
Structure the data into standard grids and formats
- 4
Intelligence Generation
Run AI weather prediction models and produce intelligence tokens
NVIDIA Earth-2 core technologies
The five core technologies of the NVIDIA Earth-2 platform used by the KWeather Weather AI Factory.
Earth-2 Atlas
Mapping global observation data
Earth-2 Nowcasting
Very short-range prediction
Global Data Assimilation
Global data assimilation
Earth-2 CorrDiff
High-resolution prediction
Earth-2 FourCastNet 3
Ensemble prediction
Industry applications
Eight industries where weather intelligence tokens are put to work.
- Logistics
- Energy
- Agriculture
- Aviation
- Insurance & risk
- Smart city
- Construction
- Mobility & autonomous driving
AI weather forecast products
The Weather AI Factory's first AI weather forecast products
Ten AI weather forecast products covering the globe and the Korean Peninsula, built on the models of NVIDIA Earth-2, NVIDIA's AI weather prediction platform.
What sets KWeather's NVIDIA AI weather forecast products apart
Global domain
Forecasting weather change across the globe.
Long-range forecastFourCastNet3 model
NVIDIA 60-day Global Forecast
- Up to 60 days
- 6-hour interval
- 25 km resolution
Medium-range forecastAtlas model
NVIDIA 15-day Global Forecast
- Up to 15 days
- 6-hour interval
- 25 km resolution
Korean Peninsula domain
A closer look at the Korean Peninsula.
Short-range forecastCorrDiff model
NVIDIA 3-day Korea Ultra-high-resolution Forecast
- Up to 3 days
- 1-hour interval
- 1.5 km resolution
NowcastStormscope model
NVIDIA Radar Precipitation Forecast
- Up to 2 hours
- 10-minute interval
- 1 km resolution
- Up to 12 hours
- 1-hour interval
- 1 km resolution
Product details
| No. | Product | AI model | Input data | Domain | Output | Max lead time | Interval | Spatial resolution | Link |
|---|
| 1 | NVIDIA 60-day Global Forecast (GFS) | FourCastNet3 | GFS initial conditions | Global | 72 variables incl. wind, pressure, temperature, precipitation, specific humidity | 60 days | 6 h | 25 km | Open → |
|---|
| 2 | NVIDIA 60-day Global Forecast (KIM) | FourCastNet3 | KIM initial conditions | Global | 72 variables incl. wind, pressure, temperature, precipitation, specific humidity | 60 days | 6 h | 25 km | Open → |
|---|
| 3 | NVIDIA 15-day Global Forecast (GFS) | Atlas | GFS initial conditions and forecast fields | Global | 75 variables incl. wind, pressure, temperature, specific humidity, surface pressure, sea surface temperature, precipitation | 15 days | 6 h | 25 km | Open → |
|---|
| 4 | NVIDIA 15-day Global Forecast (KIM) | Atlas | KIM initial conditions and forecast fields | Global | 75 variables incl. wind, pressure, temperature, specific humidity, surface pressure, sea surface temperature, precipitation | 15 days | 6 h | 25 km | Open → |
|---|
| 5 | NVIDIA 3-day Korea Ultra-high-resolution Forecast (FCN3-GFS) | CorrDiff | FCN3 forecast fields (GFS) | Korean Peninsula | Surface temperature, surface wind, precipitation, dew point, feels-like temperature | 3 days | 1 h | 1.5 km | Open → |
|---|
| 6 | NVIDIA 3-day Korea Ultra-high-resolution Forecast (FCN3-KIM) | CorrDiff | FCN3 forecast fields (KIM) | Korean Peninsula | Surface temperature, surface wind, precipitation, dew point, feels-like temperature | 3 days | 1 h | 1.5 km | Open → |
|---|
| 7 | NVIDIA 3-day Korea Ultra-high-resolution Forecast (ATLAS-GFS) | CorrDiff | Atlas forecast fields (GFS) | Korean Peninsula | Surface temperature, surface wind, precipitation, dew point, feels-like temperature | 3 days | 1 h | 1.5 km | Open → |
|---|
| 8 | NVIDIA 3-day Korea Ultra-high-resolution Forecast (ATLAS-KIM) | CorrDiff | Atlas forecast fields (KIM) | Korean Peninsula | Surface temperature, surface wind, precipitation, dew point, feels-like temperature | 3 days | 1 h | 1.5 km | Open → |
|---|
| 9 | NVIDIA 2-hour Radar Precipitation Forecast | Stormscope | Radar, satellite | Korean Peninsula | Radar, satellite | 2 h | 10 min | 1 km | Open → |
|---|
| 10 | NVIDIA 12-hour Radar Precipitation Forecast | Stormscope | Radar, satellite | Korean Peninsula | Radar, satellite | 12 h | 1 h | 1 km | Open → |
|---|
The 'KWeather Weather AI Factory' content below is translated from the introduction on KWeather's weather data page. kweather.co.kr →