AI-driven weather forecast models are now more accurate and faster than the best physics-based systems. Extending these advances to seamless global weather–climate prediction poses broader challenges, but progress is rapid. Several models trained on satellite-era reanalysis capture historical variability and trends, with some coupled to simple AI ocean components for seasonal forecasts. The key question is whether such systems can generalize to project future climate reliably. Purely data-driven extrapolation remains elusive, but emulators of physics-based models trained across multiple climates are emerging as a promising route. These can generate ensembles of ocean-coupled simulations that are statistically consistent with their reference models but at orders-of-magnitude lower cost.

I present results from the open-source Ai2 Climate Emulator (ACE), which emulates daily weather variability and climate at 100 km resolution, running ~1600 years/day on a single GPU—about 100× faster than comparable physics-based models. It has been trained on reanalysis or customized outputs from any climate model, paired with AI downscaling to provide realistic km-scale detail, or coupled to a slab-ocean model to capture CO2-driven climate change responses. Most recently, when coupled to Samudra, a full-depth ocean emulator, ACE reproduces stable coupled climate states and realistic El Niño–Southern Oscillation variability.  I conclude with remaining generalization challenges toward CMIP-type applications.