DeepMind GenCast predicts weather and extreme conditions up to 15 days ahead - more accurate than ECMWF's leading system in 97.2% of cases, and on top of that, in eight minutes, on one chip.
- DeepMind GenCast: AI weather forecasting up to 15 days ahead, described in Nature (December 4, 2024).
- More accurate than ECMWF's leading system in 97.2% of tested targets (99.8% at a horizon beyond 36 hours).
- One TPU v5 produces a 15-day forecast in 8 minutes - against hours on a supercomputer.
Eight minutes on one chip - and you have a forecast fifteen days out. That's what it takes GenCast, the new model from Google DeepMind, to compute not one forecast but a whole ensemble of probabilistic scenarios, aimed squarely at cyclones, heat waves and strong winds.
Easy on the excitement. But one thing here really is big. Not that AI 'understands' the weather - but that it computes it in minutes on one chip, instead of hours on a supercomputer. That shifts the question of who can even afford such a forecast.
Early disaster warning stops being a luxury of rich weather services. A fast, cheap forecast means more people around the world with those critical hours before the storm in hand. (Honestly: the model is trained on past data - an unseen climate regime remains its blind spot.)