AI Detected This Wildfire 2.5 Hours Before the First 911 Call

Nearly 1,300 mountaintop cameras, processing more than 7 million images a day, are now catching wildfires as early as 2.5 hours before the first 911 call. That’s the headline result from ALERTCalifornia, a UC San Diego program Microsoft’s AI for Good Lab has worked with alongside fire agencies to build a computer-vision system that spots smoke signatures across multiple camera angles, cross-verifies locations, and filters out false positives from clouds, fog, and lighting changes. In its first two months of operation, the system logged 77 confirmed detections. Microsoft has committed $5 million to the effort — $2 million in in-kind technology development and a $3 million Azure grant — and cites two concrete outcomes: the 2023 Trotter Fire contained at 52 acres versus an estimated 4,000 acres without early detection, and support for evacuating more than 180,000 residents during the 2019 Kincade Fire.

The wildfire-specific case for computer vision has a second, larger-scale data point behind it. Ai2’s OlmoEarth Platform, built to run geospatial foundation models at continent scale, ran a North America wildfire-risk mapping job across 19,600 CPUs and 994 GPUs in parallel — compressing what would have been 4,737 serial compute-hours into 30.5 wall-clock hours, a roughly 155x speedup, at a cost Ai2 describes as fractions of a penny per square kilometer. Where ALERTCalifornia detects fires already burning, OlmoEarth maps risk before ignition; together they sketch the two ends of applied AI for wildfire response, detection and prediction, both running on the same underlying advance — vision and pattern-recognition models cheap and fast enough to run continuously across huge physical areas rather than as periodic manual surveys.

It’s also a preview of the same capability applied to weather more broadly: DeepMind’s WeatherNext models already helped the U.S. National Hurricane Center forecast Hurricane Melissa’s rapid intensification in the 2025 Atlantic season, cutting the time needed for accurate multi-day storm forecasts roughly in half. As Microsoft’s Juan Lavista Ferres put it: “If you detect a fire early, you can stop it with a shovel. If you wait, you’ll need a bulldozer, air tanker, and maybe even a miracle.” For enterprises weighing AI’s real-world ROI beyond productivity tooling, early-warning systems like this are among the clearest, most measurable cases available.