Weather forecasting has long been treated as the domain of government-backed supercomputers, dense sensor networks and institutions with decades of scientific authority. WindBorne Systems is now trying to redraw that map with a mix of artificial intelligence and its own balloon-based weather data network.
The startup has released WeatherMesh 6, the latest version of its AI weather forecasting model, and says the system can deliver more frequent and more accurate predictions across key weather variables than the widely respected forecasting systems developed by the European Centre for Medium-Range Weather Forecasting, known as ECMWF.
The claim is significant because ECMWF is often regarded by meteorologists as one of the world’s leading weather prediction agencies. WindBorne is not simply trying to make forecasting faster. Its bigger argument is that AI weather models can become more powerful when they are fed with proprietary, real-world sensor data instead of depending entirely on datasets processed by major public agencies.
WindBorne Systems: Founded by Standford Students
Founded in 2019 by Stanford students John Dean and Kai Marshland, Andrey Sushko, and Joan Creus-Costa, WindBorne did not begin as an AI forecasting company. Its early bet was on building a better weather balloon and selling the data those balloons collected. The company’s direction changed after deep learning-based weather forecasting models began emerging in 2022. WindBorne then saw an opportunity to move higher up the value chain by building its own forecasting model around the data it was already collecting.
WeatherMesh 6 now produces forecasts every hour, compared with the six-hour cycle common in traditional models. In Europe and the continental United States, where data quality is strongest, the model’s resolution has reached 3 km. WindBorne says the new version performs better than both traditional and AI systems from ECMWF across several variables.
Kai Marshland, WindBorne’s chief product officer, described the improvement in a simple way: WeatherMesh 6 is “as accurate five days out as a traditional forecast is the day before,” especially for surface temperature measurements.
That is a bold benchmark in an industry where small accuracy improvements can carry major consequences. Better forecasts influence everything from aviation and defence planning to commodities trading, agriculture, energy management and disaster preparedness. For governments, weather prediction is a public safety function. For private companies, it is increasingly becoming a data business.
The difference lies in how weather models are built. Traditional forecasts rely on complex physics-based models that require expensive supercomputers and significant processing time. AI models can run faster, but they have typically struggled with resolution, the number of variables they can handle and accuracy over longer forecast windows.
WindBorne’s pitch is that data advantage matters as much as model architecture. The company currently has about 400 balloons in flight at any given time, launched from 15 sites around the world. These balloons collect sensor readings that are then used to improve its forecasting system.
John Dean, WindBorne’s CEO, framed this as the core of the company’s strategy. “I don’t understand, personally, the business model of being [an] AI based weather company without a data set advantage,” he told TechCrunch.
That point goes to the heart of modern weather AI. Public forecasting agencies have historically led the field because of their strength in data assimilation, the process of turning scattered sensor readings into a coherent, machine-readable view of the atmosphere. For now, many AI weather models depend on datasets produced by ECMWF and the US National Oceanic and Atmospheric Administration.
WindBorne is trying to reduce that dependence. Joan Creus-Costa, the company’s head of AI, said the direct ingestion of balloon data and other sources is the key reason WeatherMesh has improved. The process has not been instant. The company spent a year tuning and re-architecting its transformer-based model so it could absorb the data directly while staying stable.
Dean said WindBorne was once heavily reliant on ECMWF data when it began working on data assimilation. Today, he believes the company could still perform well even if ECMWF’s initial conditions were removed.
The company’s balloon network has also brought operational scrutiny.
Last year, a United Airlines jetliner ran into one of WindBorne’s balloons. The aircraft suffered minor damage, but no one was hurt. WindBorne had followed US regulations on the size of its sensor package. Since then, the company has added ADS-B transponders to its balloons so their location can be reported through the global aviation surveillance system, a move aimed at lowering the chances of another collision.
Commercially, WindBorne is already selling its balloon data to NOAA, where it is used in the American weather forecasting enterprise, along with the US Air Force and Navy. It also sells forecasts to investors and commodity traders. The company has raised $25 million in venture funding and carried a reported valuation of $85 million in 2024.
Yet Dean does not appear to be rushing toward a conventional software-as-a-service business. He said the company remains focused on building its model and data infrastructure, partly because the way people consume information may soon change. If users increasingly rely on AI agents for information, he suggested, the traditional SaaS product route may not be the most valuable path.
That makes WindBorne’s story bigger than weather. It reflects a wider shift in AI businesses, where owning unique data may become the clearest moat. In forecasting, that moat is not just digital. It is floating in the atmosphere, gathering readings that could help a startup compete with institutions once thought almost impossible to challenge.
For now, WindBorne’s central claim remains one to watch closely: that a private startup with AI models and a balloon network can match or beat some of the world’s most trusted forecasting systems.
If WeatherMesh 6 continues to prove itself in real-world use, weather prediction may no longer be defined only by who owns the biggest supercomputer. It may also be defined by who can collect the right data, feed it into the right model, and update the forecast before the weather changes again.
Also Read: Sushant Sinha Likely to Exit Times Now Navbharat, May Join Network18
