The weather forecast you checked this morning may have been generated by AI. Not augmented by AI. Not "AI-enhanced." Generated. Google DeepMind's GenCast outperformed the European Centre for Medium-Range Weather Forecasts, the gold standard for global weather prediction, on 97.2% of evaluation targets. NOAA is now running AI weather models operationally. The infrastructure underneath your daily forecast is being rebuilt.
What GenCast Did
The ECMWF's Integrated Forecasting System has been the benchmark for global weather prediction for decades. It runs on some of the most powerful supercomputers in the world, processing terabytes of atmospheric data through physics-based numerical models. A single forecast takes hours of compute time across thousands of processors.
GenCast runs in eight minutes on a single Google TPU.
97.2%
of evaluation targets where GenCast outperformed the ECMWF gold-standard system
Source: Google DeepMind, Nature, 2024
That alone would be significant. But GenCast also produced more accurate forecasts on 97.2% of the evaluation targets, covering temperature, wind speed, humidity, and pressure across multiple time horizons up to 15 days out. It matched or exceeded human-engineered systems that represent 40+ years of atmospheric science development.
The model is a diffusion-based architecture trained on decades of ERA5 reanalysis data. It doesn't simulate physics equations. It learns weather patterns directly from historical data and generates probabilistic forecasts, providing not just a prediction but a distribution of possible outcomes with associated confidence levels.
The Engineering Achievement
Understanding why this matters requires appreciating what traditional weather forecasting involves.
Numerical weather prediction divides the atmosphere into a three-dimensional grid. At each grid point, differential equations model fluid dynamics, thermodynamics, radiation, and moisture transport. These equations are solved iteratively, stepping forward in time at intervals of minutes to hours. The computational cost is enormous.
8 min
GenCast forecast generation time on a single TPU vs hours on a supercomputer
Source: Google DeepMind, Nature, 2024
Every improvement in resolution (smaller grid cells, more vertical layers) multiplies the compute requirement. The ECMWF's current operational model runs at approximately 9km horizontal resolution. Getting to 1km resolution, which would dramatically improve local forecasts, would require roughly 1,000 times more compute with traditional methods.
AI models sidestep this entirely. They don't solve physics equations at runtime. The physics is encoded implicitly in the training data. At inference time, the model generates a forecast directly from current conditions. The compute cost is fixed regardless of the complexity of the weather pattern.
This means AI weather models can run at higher effective resolution, produce ensemble forecasts (multiple scenarios) cheaply, and update forecasts more frequently. GenCast can generate 50-member ensemble forecasts in the time it takes a traditional system to produce one.
NOAA Goes Operational
The US National Oceanic and Atmospheric Administration didn't wait for academic consensus. NOAA deployed AI weather models into its operational forecasting pipeline in 2025, running them alongside traditional numerical models.
The AI models are particularly strong at medium-range forecasting (3-10 days out) and at identifying the probability of extreme events. For hurricane track prediction, AI models have shown measurable improvements over traditional methods, providing earlier and more precise guidance for evacuation decisions.
NOAA's approach is pragmatic: run AI and physics-based models in parallel, use AI forecasts where they demonstrate consistent advantage, and maintain traditional models as a baseline. The transition isn't instant, but the direction is clear.
The Invisible Infrastructure Pattern
Weather forecasting is a useful case study because it reveals a broader pattern: AI is rebuilding invisible infrastructure.
Nobody thinks about how their weather forecast is generated. They open an app, check the temperature, decide whether to bring an umbrella. The entire system, from satellite data collection through atmospheric modelling to the number on your phone, is infrastructure that works best when you don't notice it.
AI is transforming this infrastructure layer across dozens of domains:
Power grid management. AI models predict demand and renewable generation with enough accuracy to reduce spinning reserves, saving billions in fuel costs. Google DeepMind's work with the National Grid reduced energy needed for data centre cooling by 40%.
Traffic routing. Every major navigation app uses AI traffic prediction models that continuously retrain on real-time data. The routes you're given today are generated differently than they were three years ago.
Supply chain logistics. Container shipping routes, warehouse inventory levels, last-mile delivery timing. AI models now drive operational decisions that were previously based on heuristics and manual planning.
Financial settlement systems. Fraud detection, transaction routing, credit decisions. The pipes that move money have been AI-powered for years.
40%
reduction in data centre cooling energy achieved by DeepMind AI for Google
Source: Google DeepMind / National Grid Partnership
In each case, the end user doesn't know or care that AI is involved. They just get a better forecast, a faster route, a package that arrives on time. The value is in the outcome, not the method.
What Engineers Should Watch
Three technical developments worth tracking:
Foundation models for Earth systems. Google, NVIDIA, and several national weather services are building large-scale foundation models trained on multi-decade Earth observation datasets. These models will handle weather, climate, ocean currents, and atmospheric chemistry in a unified framework.
Real-time data assimilation. Current AI weather models are trained on historical data and run inference on current conditions. The next step is continuous learning systems that update their parameters in real time as new satellite and sensor data arrives.
Resolution scaling. AI models that can dynamically scale resolution, providing global coverage at coarse resolution and zooming to street-level detail for areas of interest. This would make hyperlocal weather forecasting (your specific neighbourhood, not your city) computationally feasible.
The weather forecasting revolution is a leading indicator. When AI can outperform 40 years of physics-based modelling in eight minutes on a single chip, the question for every domain becomes: what other infrastructure is next?