This AI weather startup is even outperforming government agencies
By improving how sensor readings are integrated into deep learning models, startup WindBorne Systems has launched a new AI weather forecasting tool today. The system provides more frequent and accurate predictions on crucial variables, even outperforming the world-leading model developed by European governments.
WindBorne was started in 2019 by a group of Stanford students who initially focused on creating superior weather balloons to sell weather data. However, when advanced deep learning models for weather forecasting emerged in 2022, the team shifted strategy, realizing they could generate much more value by developing their own proprietary AI model alongside their hardware.
Today, the company launched WeatherMesh, the sixth version of its forecasting model. According to WindBorne, this new iteration delivers higher accuracy than both the traditional and AI-driven forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF)—an intergovernmental organization widely recognized by meteorologists as the global leader in weather prediction.
To put it simply, WindBorne’s Chief Product Officer, Kai Marshland, explains that WeatherMesh-6 can predict weather conditions five days in advance with the same accuracy a traditional forecast achieves just one day prior—especially when it comes to measuring surface temperatures.
While traditional models only update every six hours, WeatherMesh-6 generates a new forecast every single hour. Additionally, its resolution has now improved to 3 km across the continental United States.
Conventional weather forecasts rely on intricate physics models that demand expensive supercomputers and significant processing time. In contrast, AI models—developed by startups and major labs such as Google DeepMind—operate much faster than physics-based systems, though they currently lag behind in resolution and long-range accuracy.
Even so, the rapid evolution of weather AI has already led to its adoption by major global government institutions. Researchers are now focusing on embedding this technology into the systems responsible for compiling meteorological data and generating public weather reports.
Windborne gains a unique advantage by combining model development with its own data collection. At any given moment, the company has about 400 balloons in the sky, collecting sensor data from 15 launch sites around the world. The breakthroughs in its latest models come from its advanced method of integrating this balloon-collected data directly into the system.
WindBorne CEO John Dean expressed skepticism to TechCrunch about competitors in the space, noting that he personally doesn’t understand the business model of running an AI weather company without having a distinct advantage in data collection.
The ECMWF owes its leading position to its expertise in “data assimilation”—the process of converting scattered sensor data into a unified, machine-readable snapshot of global weather. Currently, AI-driven weather models still rely heavily on the datasets provided by both the ECMWF and the U.S. National Oceanic and Atmospheric Administration (NOAA).
However, WindBorne and other groups are actively working to bypass traditional steps and feed data straight into their systems. Joan Creus-Costa, WindBorne’s Head of AI, notes that this direct ingestion of data from their balloons and other sources is the main reason behind the upgrades in the latest WeatherMesh version. Achieving this required a year of fine-tuning and redesigning the transformer-based model to ensure it could deliver these forecasts without sacrificing stability.
WindBorne’s CEO, John Dean, noted that while the company relied heavily on the ECMWF when they first began data assimilation, things have changed significantly. He predicts that even if they completely removed the ECMWF’s initial conditions today, their system would still perform quite well.
The company faced a serious scare last year when a United Airlines jetliner collided with one of its weather balloons. Although the aircraft sustained minor damage, fortunately, no one was injured—largely because WindBorne had strictly adhered to U.S. regulations regarding maximum sensor package sizes. To prevent similar incidents moving forward, the company now utilizes the global aviation surveillance system, ADS-B, to actively steer its balloons out of the path of oncoming aircraft.
Backed by $25 million in venture funding with an $85 million valuation reported in 2024, WindBorne sells its balloon-collected data to NOAA for use in the American weather forecasting system, as well as to the U.S. Air Force and Navy. While the startup also provides forecasts to investors and commodity traders, CEO John Dean emphasizes that the company is prioritizing its model and data infrastructure over commercial products, driven partly by the shifting dynamics of the information landscape.
Dean explained that he is reluctant to dedicate a large team to developing a traditional SaaS product, pointing out that in two years, consumers might prefer accessing information through AI agents instead.
A previous version of this article misstated how WindBorne’s balloons interact with air traffic. While the company actively monitors commercial flights and maneuvers its balloons to avoid them, it has not yet integrated ADS-B transponders directly onto its sensor platforms.
