A group of balloons without engines, drifting solely by wind, have just landed a big contract from the U.S. Air Force. WindBorne Systems, a company that makes autonomous weather balloons, announced it has won a new contract with the U.S. Air Force to purchase data collected by its global balloon fleet, with a total value of up to $49 million. For a startup that makes a business out of "peeking into the atmosphere with balloons," this is the heaviest deal it has ever participated in.

WindBorne currently has about 700 weather balloons floating around the world. They continuously read data from atmospheric sensors, selling some to paying customers and feeding the rest into the company's own AI weather forecasting model. The company, founded in Redwood City, California in 2015, was one of the first organizations to directly feed real observational data into deep learning weather models. It has long worked with the government, with the National Weather Service, the Air Force, and the Navy being familiar partners. But this latest Air Force contract truly elevated the collaboration to a new level.

The money isn't spent recklessly. The contract has a maximum value of $49 million, but the Air Force follows a "pay as you go" model — ordering only when data is needed, for specific regions and time periods. So far, only $1.7 million worth of orders have been placed, and company executives expect more to follow. Co-founder and CTO Andrey Sushko described the contract as "mainly a vehicle to support the continuous expansion of the constellation and to invest in R&D to constantly improve the balloon hardware and performance." He specifically mentioned that the contract was awarded without a bid, which the government recognized as "we are the only company capable of providing services in a way that meets their operational needs."

WindBorne's confidence comes from solid technical achievements. CEO John Dean revealed in August that the company had just completed a $37 million B-round financing to continue fueling business expansion; and the WeatherMesh6 forecast model released in early summer this year claims to be more accurate than the traditional physical law-based model used by the European Centre for Medium-Range Weather Forecasts (ECMWF). After the introduction of the transformer architecture and the release of a massive meteorological dataset by ECMWF, AI weather forecasting has become one of the most effective applications of deep learning — previously, meteorologists relied on expensive supercomputers to solve complex equations approximating atmospheric behavior, but now end-to-end learning models can process decades of data and generate predictions faster. Although cutting-edge laboratories and government agencies' models have become increasingly accurate, both types of forecasts still need to complement each other to truly understand the weather; Google DeepMind's latest weather model, for instance, has become increasingly critical in government responses during hurricane seasons.