US Air Force signs up to $49 million contract with WindBorne for balloon data

The US Air Force will buy atmospheric data from WindBorne Systems' autonomous balloon fleet under a contract that could reach about $49 million (roughly 180 million shekels), the company announced. Payment is not guaranteed up front: the service orders data for specific regions and times, and has so far committed about $1.7 million (roughly 6.3 million shekels). WindBorne expects orders to expand.

Founded in 2015 and based in Redwood City, California, WindBorne now operates roughly 700 weather balloons worldwide. The balloons gather sensor readings that are sold to customers and also fed into the company's own forecasting model, WeatherMesh. WindBorne says it was the first to pipe empirical data directly into a deep-learning weather model, breaking with the traditional approach that relies on physics simulations run on supercomputers.

The deal is the largest the company has won from the US government, following extended work with the National Weather Service, the Air Force and the Navy. It was awarded sole-source, without competitive bidding. CTO and co-founder Andrey Sushko said the decision reflects a determination that WindBorne is the only company capable of delivering global data collection that meets the service's operational needs. The funding will go toward expanding the constellation and further hardware and performance development.

In August the company released WeatherMesh 6, which it says outperforms the physics-based model of the European Centre for Medium-Range Weather Forecasts (ECMWF). A month earlier WindBorne raised $37 million (about 136 million shekels) in a Series B round to accelerate deployment. The new model joins a wave of transformer-based forecasting systems that became practical after ECMWF released its massive dataset a few years ago.

The emerging picture is what researchers call a "jagged frontier": the new models — from frontier labs like Google DeepMind as well as government agencies — show superior accuracy on many metrics, yet both physics-based and deep-learning forecasts are still needed for a usable picture. DeepMind's cyclone-forecasting tools, for instance, have become a core component of US government hurricane-season preparations.

In August WindBorne became the first organization to send an unpowered, unmanned system into the eye of a typhoon to collect data. Navigation works by changing altitude: venting air or dropping sand ballast to catch winds that carry the balloon to its target. Sushko said the feat rested on two conditions — an accurate wind model that can navigate, and a constellation large enough to ensure balloons are in the right places. He called it one of the first cases where a deep-learning model drove real-time operational decisions.