The U.S. Air Force has agreed to purchase weather data gathered by WindBorne Systems' network of autonomous balloons, a deal that could bring the startup close to $50 million. Right now, roughly 700 of the company's weather balloons are aloft worldwide, gathering atmospheric sensor measurements that WindBorne sells to clients and also incorporates into its own AI-driven forecasting model.

Based in Redwood City, California, and founded in 2015, WindBorne stands among the earliest organizations to feed empirical data directly into a deep-learning weather model. While the company has maintained a close working relationship with the U.S. government—especially the National Weather Service, the Air Force, and the Navy—this represents the largest single contract it has ever been part of.

The agreement carries a maximum value of $49 million, though the Air Force only pays as it requires data, placing orders for particular regions at particular times. To date, it has placed orders totaling $1.7 million, with executives anticipating additional orders down the line. The contract was awarded without a competitive bidding process, which, as noted, reflects "a determination by the government that for this capability of collecting global data, we were the only company able to provide it in the way that met their needs for these operations."

Earlier this summer, WindBorne rolled out the newest iteration of its forecasting model, WeatherMesh 6, which the company claims delivers forecasts more accurate than the conventional model operated by the European Centre for Medium-Range Weather Forecasts (ECMWF)—a system that simulates the atmosphere through the laws of physics. Since transformer model architectures emerged and the ECMWF released a vast weather dataset a few years back, AI weather forecasting has grown into one of the more fruitful applications of deep-learning technology. In the past, meteorologists depended on costly supercomputers crunching detailed equations that approximate atmospheric behavior. Today, end-to-end learning can absorb decades of weather data and produce predictions more rapidly. Models developed by frontier labs and government agencies are demonstrating greater accuracy than traditional forecasts, though the frontier remains uneven, and both forecast types are still needed to build a useful understanding of the weather. Google DeepMind's newest weather model introduces significant new capabilities, and its cyclone forecasting tools are playing an increasingly vital role in how the U.S. government handles hurricane season.

In August, WindBorne became the first organization to deploy an unpowered flight system—one without an engine—to gather data inside the eye of a typhoon. The company steers its balloons by changing their altitude, releasing air, or dropping sand held as ballast, directing the vehicles into winds that carry them to their destinations. "That's a pretty novel step, and the two ingredients for that are one, having good wind and weather models, and knowing how to navigate them; and two, having a large enough constellation that you have balloons in the right places," Sushko said. "It's one of the first times that a deep-learning model was used for actual operational decisions."