The hurricane models got better at predicting where storms go. They still cannot predict how strong they will be when they get there.
The 2025 Atlantic hurricane season produced 23 rapid intensification events — cases where a storm's wind speed spiked by at least 35 miles per hour in 24 hours, accounting for 10 percent of all forecasts for the year, according to the NHC's 2025 season verification report. That is the number that should keep coastal residents up at night, because every major forecasting model — including DeepMind's — still struggles with the physics that make a storm explode.
DeepMind's GraphCast model did achieve something real: it became the first AI system to run inside the National Hurricane Center's forecast cycle and beat the NHC's own official track predictions at the 12-hour through 72-hour windows, according to the NHC's verification report. At five days out, the model placed cyclones an average of 140 kilometers closer to their actual positions than the ECMWF ensemble, the previous gold standard in global weather modeling. The model generates 50 possible storm scenarios out to 15 days and produces a five-day forecast in roughly one minute on a single specialized computer chip.
These are real operational numbers, not preprint theater. The NHC tested GraphCast as part of its 2025 hurricane season workflow — a milestone that puts AI forecasting squarely inside the government institution that issues the watches and warnings that coastal residents rely on.
But the model's advantage narrows sharply when the conversation turns to intensity. DeepMind's internal tests showed its model outperforming NOAA's Hurricane Analysis and Forecast System on intensity error. "There were 23 cases of rapid intensification, a much higher proportion than in an average season," the NHC report noted. Every major model still struggles with the physics — the precise interplay of ocean heat, wind shear, and inner-core processes that can add 50 miles per hour to a storm's winds in 24 hours. That problem has challenged conventional models for decades. It is not a software bug. It is a fundamental gap in how the field represents a specific class of atmospheric instability.
Track accuracy has improved across the industry — NHC forecasts were up to 14 percent better than the five-year average in 2025. But a hurricane that veers slightly off course is a near-miss. One that undergoes rapid intensification before landfall is a disaster. The 50-scenario ensemble approach helps communicate uncertainty, and the speed advantage means forecast cycles can update more frequently without computational bottlenecks. The model is a very good instrument for estimating where a storm is going. It is not yet a reliable instrument for estimating how bad it will be when it arrives.
The 2025 Atlantic season itself was a peculiar test. It produced 13 named storms, three of which reached Category 5, yet no storm made landfall in the United States — the first such quiet landfall season in ten years. That pattern tests track models differently than a year with multiple Gulf and East Coast hits would. The numbers are real; their generalizability to a more destructive season remains an open question.
The operational transition has begun. For the first time, a frontier AI model is part of the cycle that generates the advisory messages that emergency managers, utilities, and coastal residents act on. The track improvement is real and consistent. The intensity problem is real and persistent. Both things are true at the same time, and knowing which one to worry about is the point of the story.
† Consider rephrasing to: 'DeepMind reported that in internal tests, its model outperformed NOAA's Hurricane Analysis and Forecast System (HAFS) on intensity error' and add a † footnote.
†† Update the hyperlink to point to one of the registered sources (weather.com or NHC PDF) where this data appears, rather than the unregistered third-party site.
† Consider rephrasing to: 'DeepMind reported that in internal tests, its model outperformed NOAA's Hurricane Analysis and Forecast System (HAFS) on intensity error' and add a † footnote.
†† Update the hyperlink to point to one of the registered sources (weather.com or NHC PDF) where this data appears, rather than the unregistered third-party site.
