Google is betting that live observations and AI can close stubborn gaps in rain and snow forecasting, especially in regions with sparse ground equipment. But its own rollout acknowledges that faster machine predictions still need to sit alongside the physics-based models forecasters rely on for warnings.

Google introduced WeatherNext 3 as the latest version of its AI forecasting system, saying it can generate a global forecast each hour from recent satellite observations rather than relying solely on older model inputs. The company says that shift delivers a picture five times sharper than its previous model, with some temperature and moisture forecasts resolved down to five kilometres.¹

The central promise is precipitation: fast-moving storms are notoriously difficult to pin down, and Google says forecasts made at least a day ahead can be up to 50% more accurate. The potential gains are greatest outside the US and Europe, where fewer rain gauges leave wider holes in conventional observation networks. “We’re able to leverage fresher and richer observational data sets,” Samier Merchant, a Google Research engineer, said.¹

Google DeepMind framed the launch more broadly as “a major breakthrough” in global forecasting, saying the model learns directly from real-world, real-time observations to produce more localised predictions faster.Demis Hassabis (demishassabis)² DeepMind chief Demis Hassabis amplified that message in a repost.²

The system is now being folded into Search, Maps and Gemini, while Google has also designed forecasts for renewable-energy operators, including wind-speed estimates at turbine height. Ferran Alet, a Google DeepMind research scientist, linked that work to rising power demand, arguing that making renewables “a very appealing opportunity is very important for us.”¹

Yet WeatherNext 3 is not being presented as a replacement for traditional meteorology. It is trained partly on physics-based models, and weather agencies still compare multiple forecasts before issuing public alerts—a reminder that sharper AI output is useful only if it improves decisions when the weather turns dangerous.¹