A world-class forecast now fits on one machine.
Weather forecasting has spent fifty years as one of the largest users of supercomputing. AI models trained on the archive now produce comparable global forecasts in minutes, and the weights are published. The measurements they depend on are the part that has not become cheap.
Original sources ↓ · Revision history ↓
Deployed · source published 2025-12-17
What the evidence supports
- Verified in the source
- NOAA states that a single 16-day forecast from its AI model uses 0.3% of the computing resources of the operational GFS and finishes in about 40 minutes. ECMWF has run an AI forecast operationally since 25 February 2025. WMO and SOFF document the observing gap those models cannot close.
- Claimed / not established here
- Comparative skill and energy multipliers published by model developers are their own measurements. Open weights are not an operational service, and NVIDIA's assimilation paper reports analyses about twice as far from the truth as the reanalysis it is compared against.
- Bright editorial interpretation
- The barrier to a world-class forecast has moved from silicon to sensors: the compute is now cheap, and the measurements are not.
Independent support: Two national and international weather services — NOAA and ECMWF — have put AI forecasts into operational service and published their own resource and skill figures, and NOAA states plainly where version 1.0 is worse.
Independent source ↗ · Operational deployment on 17 December 2025; 0.3% of GFS compute, about 40 minutes, and a stated degradation in tropical cyclone intensity.
Independent source ↗ · AIFS operational since 25 February 2025 at 28 km, with gains of up to 20% on tropical cyclone tracks.
Independent source ↗ · Germany has more stations meeting the global baseline than the whole of the African continent.
Primary source ↗ · NVIDIA's own account of the January 2026 Earth-2 open-weight release.
AI-assisted source review, 12 September 2026. Bright first encountered this work through material curated by NVIDIA, then checked it against the sources above. Other records Bright met the same way ↗
Try the interactive explainer ↗ · Authored explanation, not a measured system result.
The human problem
Many countries cannot run their own global forecast, and a forecast made elsewhere is not always tuned to the hazard that arrives at home.
The prior constraint
Observations must first be assembled into a coherent estimate of the whole atmosphere, then stepped forward by equations. Both halves ran on national supercomputers.
AI’s actual role
NVIDIA's January 2026 Earth-2 release published open weights spanning the chain: HealDA turns observations into an initial atmosphere, Atlas forecasts to 15 days, StormScope predicts local weather zero to six hours ahead. ECMWF's AIFS has been operational since 25 February 2025 and NOAA's AIGFS since 17 December 2025.
The documented result
NOAA reports that a single 16-day AIGFS forecast uses 0.3% of the computing resources of the operational GFS and finishes in about 40 minutes. ECMWF reports gains of up to 20% on tropical cyclone tracks from AIFS at a 28 km grid against 9 km for its physics model.
Why it may matter
The compute barrier to a world-class forecast has effectively collapsed. The observing barrier has not: SOFF reports that less than 10% of the required basic weather and climate data are available from least developed countries and small island developing states, and WMO notes that Germany has more GBON-compliant observing stations than the whole of the African continent.
Limitations
Open weights are not the same as an operational service, and NVIDIA's own assimilation paper reports analyses about twice as far from the truth as ERA5.
NOAA states that AIGFS version 1.0 degrades tropical cyclone intensity forecasts, which future versions will address.
Speed and energy comparisons are mostly published by the organisations that built the models; NOAA's fractions come from a real operational pipeline, most multipliers do not.
Every one of these models is trained on a reanalysis built from the observing network. No model can supply a measurement that was never made.
Unresolved questions
Which national services actually gain the capacity to run and trust their own forecasts, and by when?
Does forecast skill on extremes — intensity, not track — close the gap that physics models still hold?
Source history & evidence assessment
- Maturity
- Deployed
- Claim confidence
- unassessed
- Event date
- Not recorded
- Source published
- 2025-12-17
- Captured
- 2026-09-09
- Last source review
- 2026-09-12
- Editorial method
- AI-assisted source review
- Place / relevance
- Not recorded
Re-reported from primary sources in September 2026. The operational facts come from NOAA and ECMWF and the observing-gap figures from WMO and SOFF; vendor performance multipliers were not adopted. The NVIDIA announcement remains cited as a primary account of NVIDIA's own release and is not independent verification. Bright first encountered the Earth-2 release through material curated by NVIDIA and then researched the wider field independently.
Maturity describes the tested or operational setting. Confidence describes support for the particular claim; one does not determine the other.
Original sources
NOAA deploys new generation of AI-driven global weather models ↗ · government
ECMWF’s AI forecasts become operational ↗ · institution
Closing the gaps in the observing network · World Meteorological Organization ↗ · institution
Systematic Observations Financing Facility ↗ · institution
NVIDIA: Earth-2 open weather model announcement ↗ · institution
Institutions: NOAA · ECMWF · World Meteorological Organization · NVIDIA
Explore the underlying question
Related developments
Editorial connections between distinct settings and results; these links do not imply replication.
Revision & correction history
2026-09-09 · Added to Bright as a bounded source claim with visible evidence distinctions.
2026-09-12 · Re-reported around operational adoption at NOAA and ECMWF and the observing gap documented by WMO and SOFF, replacing a record based on a single vendor announcement.
No corrections recorded.