As Typhoon Dolphin bore down on China’s coast this week, the country’s bet on artificial intelligence to read the skies looked less like a research curiosity and more like national infrastructure.
Extreme weather is intensifying, the cost of getting a forecast wrong is rising, and Beijing is pouring resources into models that promise to see storms coming sooner and cheaper than the supercomputers that have long done the job.
Three home-grown systems lead the effort. Fengwu comes from the Shanghai AI Laboratory, Pangu from Huawei, and Fuxi from Fudan University, a spread that puts a state lab, a tech champion, and a leading university into the same race.
It is a very Chinese arrangement, and a deliberate one, since forecasting has quietly become a field where AI supremacy carries real-world stakes.
The headline claims are striking. Fengwu reportedly outperformed Google DeepMind’s GraphCast across roughly 80% of the weather variables it was tested on, and pushed skilful forecasts beyond the ten-day mark that has traditionally marked the edge of usefulness.
If the numbers hold up to independent scrutiny, they suggest the gap with the West is not merely narrowing but, on some measures, closing.
Dolphin offered a vivid test. Five days before the storm made landfall on the mainland, Fengwu predicted the time and place of impact to within 30 minutes and 30km, or about 19 miles, according to Sun Zhi of the firm Techwind.
That is the kind of precision that turns an evacuation order from guesswork into logistics, and it is precisely what governments want when a typhoon is still a working week away and every hour of warning buys another town the time to move people and boats out of harm.
AI forecasters generate their predictions far faster than conventional numerical weather prediction, and they lean on a fraction of the computing power, which is why the approach has spread so quickly from research papers to operational trials.
For a country that runs some of the world’s most weather-exposed agriculture and coastline, speed and cost are not footnotes.
“With more extreme weather, people need information to make decisions, both local governments, the national government, also the average person, farmers and fishermen,” Sun said, framing the work less as a benchmark contest than as a public service for people whose livelihoods hinge on the next front coming in off the sea.
Yet the caveats are real, and China’s forecasters are candid about them. AI models still lag traditional methods at predicting storm intensity, which is the difference between a manageable blow and a genuine catastrophe, and they remain untested against major, longer-term climate developments that fall outside the historical data they learned from.
A model brilliant at where a storm lands can still be shaky on how hard it hits, and a warming climate keeps serving up conditions no training set has seen before.
The wider contest gives the effort its edge. China is chasing a Western field that has moved fast, from DeepMind’s probabilistic GenCast system to Nvidia-backed FourCastNet and the European Centre for Medium-Range Weather Forecasts’ AIFS, with well-funded startups piling in as well.
One Swiss firm has even claimed its forecaster beats Microsoft and Google, a sign of how crowded and combative the space has become.
What makes weather an unusually revealing arena for the AI rivalry is that the referee is physics. A chatbot can be graded on taste, but a typhoon forecast is either right or it is not, and the ground truth arrives on schedule.
As the planet warms and the storms grow fiercer, China is wagering that the side which reads the atmosphere fastest will hold an advantage that is measured not in benchmarks but in lives.
If you liked the article, do not forget to share it with your friends. Follow us on Google News too, click on the star and choose us from your favorites.
If you want to read more like this article, you can visit our Technology category.