AI Weather Forecasting, Built on High-Tech Balloons

In October 2024, Hurricane Milton turned into one of the fastest-growing storms on record over the Atlantic Ocean. The hurricane’s rapid gain in intensity caught meteorologists off guard, which meant the affected communities were surprised too. The storm ultimately claimed 15 lives and caused US $34 billion in damages as it tore across Florida.

Why was Milton’s explosive growth so hard to anticipate? This failure stemmed from a lack of good weather data. The kind of data you can get only by flying a suitably outfitted aircraft straight into a developing storm. This type of mission requires human pilots to put their lives at risk to release dropsondes—sensors dangling from parachutes—that will gather critical atmospheric measurements. If meteorologists can get that precious data in time, they can often use it to produce life-saving predictions.

But hurricane hunters can fly only so many missions, and most storms develop in places that aircraft can’t safely reach, such as over vast ocean expanses. So we are left with massive data gaps precisely where the most dangerous weather begins.

WindBorne Systems, the company I cofounded in 2019, is pioneering a better way to predict the weather. Our approach starts with cutting-edge weather balloons and ends with our proprietary AI weather-forecasting system. Hurricane Milton’s dramatic arrival last year gave us our first opportunity to observe such a weather system directly and to predict a hurricane’s path as the storm evolved.

The WindBorne crew arrives before dawn to set up a balloon launch at Bodega Bay, Calif. Christie Hemm Klok

At WindBorne, based in Palo Alto, Calif., we’ve developed a sophisticated type of long-duration weather balloon. These Global Sounding Balloons (GSBs), as we call them, can maneuver through the atmosphere and follow dynamic flight paths by surfing the winds. In the lead-up to Milton, we launched six of these balloons, carrying dropsondes, from a safe distance away, in Mobile, Ala. Within the next 24 hours, the balloons were able to enter the hurricane and release their dropsondes to measure temperature, pressure, and humidity, along with wind speed and direction—information that potentially could have helped forecasters determine exactly how the hurricane would behave.

The sensors that collect weather data for each Global Sounding Balloon are encased in plastic. Christie Hemm Klok

This dropsonde deployment, the first ever by weather balloon, demonstrated that it’s possible to release airborne sensors without the usual costs and risks to human life. And when our team ran the collected data through our AI-based forecasting model, WeatherMesh, its predictions of Milton’s path were more accurate than those from the U.S. National Hurricane Center. Alas, because our dropsonde launch was an experiment meant to test our technology’s capabilities, the results we obtained couldn’t be disseminated to the public in real time. But it was nevertheless a great accomplishment: WindBorne proved definitively that AI forecasts can outperform the kind of weather models our society has relied on for decades.

Our mission at WindBorne is to build what we liken to a “planetary nervous system”—an end-to-end AI-based forecasting system that can gather vast amounts of weather data and transform that data into accurate and timely forecasts. Just as a person’s nervous system constantly sends information from all parts of the body to the brain, our planetary nervous system gathers observations from all over the Earth and sends them to our AI brain.

Our system, which requires both advanced data-collection hardware and sophisticated AI modeling, can radically improve how people use weather predictions to make decisions in areas such as national defense, renewable energy, and agriculture. With climate change increasing the frequency and cost of extreme weather events like Milton, we hope to provide better forecasts to help society navigate this new reality.

WindBorne’s Stanford Origins

WindBorne started as a 2015 project in the Stanford Student Space Initiative, when Andrey Sushko (now WindBorne’s CTO) and some other students became interested in extending the flight duration of conventional weather balloons. Most weather balloons burst after just a couple of hours in flight, collecting data for only a single up-down cycle as they ascend, pop, and then drop back down to the ground. These balloons almost never go far beyond their continental launch sites, leaving the air above oceans, deserts, and other remote regions underobserved. That’s problematic because weather is global: A disturbance that starts near the west coast of Africa can develop into the next catastrophic storm to hit North America.

While working on the project, we discovered that the flight limitations of conventional weather balloons mean that they’re observing only about 15 percent of the globe. We realized that if we improved the hardware and control systems, we could create weather balloons that self-navigate and intelligently “surf” the wind, allowing them to stay aloft much longer than conventional balloons—think weeks instead of hours.

John Dean cofounded WindBorne in 2019. Jason Henry/The New York Times/Redux

I cofounded the company in 2019 with four of my peers from Stanford, and later took on the role of CEO. At that time, we were still in the early R&D stages for our balloons. The result of that work was a design for autonomous, long-duration balloons that communicate with operators via satellite. In 2024, we introduced our first AI forecasting model, WeatherMesh, to ingest the data from the balloons and give them high-level instructions on where to fly next to fill in specific data gaps.

The main envelope of a WindBorne balloon is made from a thin, transparent film just 20 micrometers thick—less than half the thickness of a human hair—and the whole assembly weighs less than 2 kilograms. Each balloon has a bag of sand used as ballast; the balloon can release sand to rise higher or vent gas to descend to a different wind current. Each balloon’s onboard autonomous system plots how to use the winds at different elevations to reach the locations specified by its WeatherMesh instructions.

Our GSBs, which collect orders of magnitude more data than single-use dropsondes, make up Atlas, our global constellation. Today, our GSBs can fly for well over 50 days at altitudes ranging from ground level up to around 24 kilometers. Atlas, which typically has hundreds of balloons in the air at any time, collects more in situ data each day than the balloons managed by the U.S. National Weather Service.

Following our time at Stanford, the WindBorne team built a business by scaling our Atlas constellation and providing weather data as a service. At first, the balloons’ navigation was guided by results from a traditional numerical weather-prediction model that ran on a supercomputer. But running that model required hundreds of times as much computing power as AI weather models do. As our constellation proved capable of collecting vast amounts of data, we knew we needed to build a model that could not only efficiently direct our balloon constellation but also assimilate its massive datasets.

The Limitations of Traditional Forecast Methods

Currently, most weather forecasts rely on physics-based numerical weather prediction. In the United States, this job is handled by the federal government’s Global Forecast System (GFS), which ingests data from satellites, ground stations, radar systems, and a worldwide network of conventional weather balloons. It runs on a supercomputer four times a day, using a technique called data assimilation to produce forecasts that extend up to 16 days out. Data assimilation interprets new data alongside historical data to come up with the most accurate forecast possible.

But therein lies the problem: Forecasting models are only as accurate as the data they are fed. With much of the global atmosphere not being regularly probed by balloons, current forecasts are hamstrung by the sparseness of the datasets available to them. You’ve probably seen a hurricane’s forecast cone shift dramatically from one day to the next. That volatility comes in part from the incomplete data driving these models. What’s more, physics-based models require enormous computing resources, which translate into high operational costs.

For the launch, the balloon is mounted on a ring that’s aligned with the wind. Christie Hemm Klok

Over the last few years, AI models have disrupted weather forecasting, proving that they can generate faster, less costly, and more accurate predictions when compared with the prior gold standard of physics-based numerical weather models. When the Chinese company Huawei introduced its Pangu-Weather model in 2023, it served notice that AI forecasting could not only compete with physics-based models, but it could even outperform them. Other recent AI weather models include Google DeepMind’s GraphCast and AIFS from the European Centre for Medium-Range Weather Forecasts. But our system outperforms all of them, sometimes by a very large measure.

While they continue to smash records, AI models (including ours) still make use of traditional physics-based models in several ways. For starters, all AI models are trained on historical weather data and predictions produced by conventional systems. Without them, the model would have to rely on raw, real-time observations for training data, without historical context.

AI models also inherently lack an advanced understanding of physics, so traditional models provide a baseline to ensure that AI-generated predictions are physically plausible. This assistance is especially important during extreme weather events, when physics-based models can help AI models simulate rare conditions based on atmospheric principles.

How We Built our AI Weather-Forecasting Model

When the WindBorne team set out to build the initial version of WeatherMesh, we had three main goals. First, it had to be inexpensive to run. Second, it needed to be at least as accurate as the top physics-based models. Third, it had to deliver forecasts with a high spatial resolution, providing fine-grained predictions on the scale of tens of kilometers.

We decided to use an architecture based on what are called transformers—the same technology that powers large language models like ChatGPT—because transformers can process huge datasets efficiently once they’re trained. This architecture includes what AI mavens refer to as an encoder-processor-decoder structure. The encoder transforms raw weather data—things like temperature, wind, and pressure—into a simpler compressed format known as latent space, where patterns are easier for the model to work with. The processor then runs calculations in this latent space to predict how the weather will change over time. To create longer-range forecasts, we simply run the processor step multiple times, with the output of the last prediction step serving as the input for the next. Finally, the decoder translates the results back into real-world weather variables.

We trained our first weather model at our headquarters using a cluster of a few dozen Nvidia RTX 4090 graphics processing units (GPUs), which cost far less than relying on cloud-computing services to handle hundreds of terabytes of atmospheric data. Setting up our own machines paid off. The hardware set us back about $100,000, but had we run all our training experiments in the cloud instead, it easily would have cost four times as much.

The balloon is initially doubled up [top] to make it more maneuverable before launch. Then Andrey Sushko, cofounder and CTO of WindBorne Systems, releases the balloon. A screenshot [bottom] shows data gathered by the balloon in real time. Photos: Christie Hemm Klok; Screenshot: WindBorne

The first version of WeatherMesh was smaller, faster, and cheaper to operate than the AI weather models created by tech giants. During training, it used about one-fifteenth the computing power of DeepMind’s GraphCast and one-tenth that of Huawei’s Pangu-Weather. Its small size makes its stellar performance all the more notable: It outperformed both those AI models and traditional physics-based models.

The early accuracy gains of WeatherMesh can be attributed to our data-collection method. Our GSBs collect 30 to 50 times as much data as do conventional balloons, and we feed that data directly into WeatherMesh. We measured our model’s accuracy based on frequency of errors when compared with other physics- and AI-based models. In 2024, we beat both Huawei’s Pangu-Weather and DeepMind’s GraphCast to become the most accurate AI forecasting model in the world. At the time this article is being published, in October 2025, WeatherMesh retains the lead.

Our initial version of the model took in data and output forecasts at 0.25-degree resolution (about 25 kilometers per grid cell) to match the resolution of ERA5, a widely used historical weather dataset. Today, WeatherMesh also includes a component that can provide forecasts for selected locations at a resolution of about 1 km.

Most AI weather models train on historical datasets like ERA5, which organizes decades of atmospheric data into a consistent framework. But we also wanted WeatherMesh to run “live,” ingesting real-time balloon observations and up-to-date analyses from the U.S. and European agencies. That transition was challenging, because most AI models perform worse when they shift from carefully curated historical data to messy real-world feeds.

To address this issue, we built specialized adapters based on a type of neural-network architecture known as U-Net, which excels at learning spatial features across different scales. These adapters translate real-time data into the same internal format used for WeatherMesh’s training data. In this way we preserved the benefits of training on ERA5 while still delivering accurate real-time forecasts.

Building On Success With WeatherMesh-4

Following the success of our initial WeatherMesh model, we released the second, third, and fourth versions of the model in rapid succession. WeatherMesh-4 predicts standard atmospheric variables at 25 vertical levels throughout the atmosphere. It also predicts a wide range of conditions at the surface, including temperature and dewpoint at 2 meters from the ground, wind speed at 10 meters and 100 meters, minimum and maximum temperatures, precipitation, solar radiation, and total cloud cover. It can produce a full forecast every 10 minutes based on the latest observations. In contrast, traditional global weather models update every 6 hours.

We’ve run extensive benchmarks to compare the latest version of WeatherMesh with other popular forecasting systems. We’ve found that the model’s predictions for the Earth’s surface and atmosphere are up to 30 percent more accurate than those from a traditional model from the European Centre for Medium-Range Weather Forecasts, and also surpass results from DeepMind’s latest model, GenCast, on most evaluations.

Building an end-to-end system means the entire pipeline must work in harmony. Our balloon constellation can’t afford to wait 12 hours for a new forecast; it needs near-constant refreshes to navigate the skies. Meanwhile, the AI model uses fresh atmospheric data from the balloons to improve the accuracy of its forecasts. Balancing these requirements forced us to get creative about how we moved the data and ran the model, but ultimately we produced a powerful system that’s fast and responsive.

What’s Next for WindBorne

In the coming years, our goal is to expand our Atlas balloon constellation to about 10,000 GSBs flying at any time, launched from about 30 sites worldwide. To achieve that goal we’ll need roughly 300 launches per day, or 9,000 per month. By 2028, we believe the entire globe could be under near-continuous observation by Atlas, from the remote Pacific to the polar ice caps. And we continue to test the boundaries of what is possible: WindBorne recently kept a balloon aloft for a record-breaking 104 days.

We’re not aiming to make physics-based weather models obsolete. We see a future where AI and traditional methods operate side by side, each reinforcing the other. Governments, researchers, and corporations can lean on these improved forecasts to guide disaster preparedness, aviation, supply-chain logistics, and more. Our planet’s weather challenges are only going to intensify as the climate continues to change, and improved forecasts are key to helping us prepare.

Each WindBorne balloon contains ballast that can be released to gain altitude. Christie Hemm Klok

A technician connects sensors to a valve (white and blue circle) that vents gas to reduce altitude. Christie Hemm Klok

Looking back at Hurricane Milton, it still feels surreal that our balloons managed to ride into a storm of that scale. Yet that was the moment WindBorne proved that a new and agile system could deliver real value where legacy methods fall short. In a world where an extra 12 or 24 hours of warning can mean the difference between safety and devastation, end-to-end AI forecasting offers a revolution in how people can observe, predict, and protect themselves from the most powerful forces on Earth.

In October 2024, Hurricane Milton turned into one of the fastest-growing storms on record over the Atlantic Ocean. The hurricane’s intensity caught meteorologists off guard, which meant the affected communities were surprised too. The storm ultimately claimed 15 lives and caused US $34 billion in damages as it tore across Florida.

Why did weather forecasters miss the danger this storm presented until it was too late? This failure stemmed from a lack of good weather data. The kind of data you can get only by flying a suitably outfitted aircraft straight into a developing storm. This type of mission requires human pilots to put their lives at risk to release dropsondes—sensors dangling from parachutes—that will gather critical atmospheric measurements. If meteorologists can get that precious data in time, they can often use it to produce life-saving predictions.

But hurricane hunters can fly only so many missions, and most storms develop in places that aircraft can’t safely reach, such as over vast ocean expanses. So we are left with massive data gaps precisely where the most dangerous weather begins.

At WindBorne Systems, in Palo Alto, Calif., the company I cofounded in 2019, we’re pioneering a better way to make weather predictions. Our approach starts with cutting-edge weather balloons and ends with our proprietary AI weather-forecasting system. Hurricane Milton’s dramatic arrival last year gave us our first opportunity to observe such a weather system directly and to predict a hurricane’s path as the storm evolved.

WindBorne has developed a sophisticated type of long-duration weather balloon. These Global Sounding Balloons (GSBs), as we call them, can maneuver through the atmosphere and follow dynamic flight paths simply by leveraging the wind. In the lead-up to Milton, we launched six of these balloons, carrying dropsondes, from a safe distance away, in Mobile, Ala. Within the next 24 hours, the balloons were able to enter the hurricane and release their dropsondes to measure temperature, pressure, and humidity, along with wind speed and direction—information that potentially could have helped forecasters determine exactly how a hurricane would behave.

Forecasting models are only as accurate as the data they are fed.

This dropsonde deployment, the first ever by weather balloon, demonstrated that it was possible to release airborne sensors without the usual costs and risks to human life. And when our team ran the collected data through our AI-based forecasting model, WeatherMesh, its predictions of Milton’s path were more accurate than those from the U.S. National Hurricane Center. Alas, because our dropsonde launch was an experiment meant to test our technology’s capabilities, the results we obtained couldn’t be disseminated to the public in real time. But it was nevertheless a great accomplishment: WindBorne proved definitively that AI forecasts can outperform the kind of weather models our society has relied on for decades.

Our mission at WindBorne is to build what we liken to a “planetary nervous system”—an end-to-end AI-based forecasting system that can gather vast amounts of weather data and transform that data into accurate and timely forecasts. Just as a person’s nervous system constantly sends information from all parts of the body to the brain, our planetary nervous system gathers observations from all over the Earth and sends them to our AI brain.

Our system, which requires both advanced data-collection hardware and sophisticated AI modeling, can radically improve how people use weather predictions to make decisions in areas such as national defense, renewable energy, and agriculture. With climate change increasing the frequency and cost of extreme weather events like Milton, we hope to provide better forecasts to help society navigate this new reality.

WindBorne’s Stanford Origins

WindBorne started as a 2015 project in the Stanford Student Space Initiative, when Andrey Sushko (now WindBorne’s CTO) and some other students became interested in extending the flight duration of conventional weather balloons. Most weather balloons burst after just a couple of hours in flight, collecting data for only a single up-down cycle as they ascend, pop, and then drop back down to the ground. These balloons almost never go far beyond their continental launch sites, leaving the air above oceans, deserts, and other remote regions drastically underobserved. That’s problematic because weather is global: A disturbance that starts near the west coast of Africa can develop into the next catastrophic storm to hit North America.

While working on the project, we discovered that the flight limitations of conventional weather balloons result in only about 15 percent of the globe being adequately observed. We realized that if we improved the hardware and control systems, we could create weather balloons that self-navigate and intelligently “surf” the wind, allowing them to stay aloft much longer than conventional balloons—think weeks instead of hours.

I cofounded the company in 2019 with four of my peers from Stanford, and later took on the role of CEO. At that time, we were still in the early R&D stages for our balloons. The result of that work was a design for autonomous, long-duration balloons that communicate with operators via satellite. In 2024, we introduced our first AI forecasting model, WeatherMesh, to ingest the data from the balloons and give them high-level instructions on where to fly next to fill in specific data gaps.

Each balloon has an antenna that enables it to communicate via satellite. Christie Hemm Klok

A technician assembles the valve used to vent gas. Christie Hemm Klok

The main envelope of a WindBorne balloon is made from a thin, transparent film just 20 micrometers thick—less than half the thickness of a human hair—and the whole assembly weighs less than 2 kilograms. Each balloon has a bag of sand used as ballast; the balloon can release sand to rise higher or vent gas to descend to a different wind current. Each balloon’s onboard autonomous system plots how to use the winds at different elevations to reach the locations specified by its WeatherMesh instructions.

Our GSBs, which collect orders of magnitude more data than single-use dropsondes, make up Atlas, our global constellation. Today, our GSBs can fly for well over 50 days at altitudes ranging from ground level up to around 24 kilometers. Atlas, which typically has hundreds of balloons in the air at any time, collects more in situ data each day than does the U.S. National Weather Service.

Following our time at Stanford, the WindBorne team built a business by scaling our Atlas constellation and providing weather data as a service. At first, the balloons’ navigation was guided by results from a traditional numerical weather-prediction model that ran on a supercomputer. But running that model required hundreds of times as much computing power as AI weather models do. As our constellation proved capable of collecting vast amounts of data, we knew we needed to build a model that could not only efficiently direct our balloon constellation but also assimilate its massive datasets.

The Limitations of Traditional Forecast Methods

Currently, most weather forecasts rely on physics-based numerical weather prediction. In the United States, this job is handled by the federal government’s Global Forecast System (GFS), which ingests data from satellites, ground stations, radar systems, and a worldwide network of conventional weather balloons. It runs on a supercomputer four times a day, using a technique called data assimilation to produce forecasts that extend up to 16 days out. Data assimilation interprets new data alongside historical data to come up with the most accurate forecast possible.

But therein lies the problem: Forecasting models are only as accurate as the data they are fed. So with 85 percent of the global atmosphere not being regularly probed, current forecasts are hamstrung by the sparseness of the datasets available to them. You’ve probably seen a hurricane’s forecast cone shift dramatically from one day to the next. That volatility comes in part from the incomplete data driving these models. What’s more, physics-based models require enormous computing resources, which translate into high operational costs.

By 2028, we believe the entire globe could be under near-continuous observation by Atlas.

Over the last few years, AI models have disrupted weather forecasting, proving that they can generate faster, less costly, and more accurate predictions when compared with the prior gold standard of physics-based numerical weather models. When the Chinese company Huawei introduced its Pangu-Weather model in 2023, it served notice that AI forecasting could not only compete with physics-based models, but it could even outperform them. Other recent AI weather models include Google DeepMind’s GraphCast and AIFS from the European Centre for Medium-Range Weather Forecasts. But our system outperforms all of them, sometimes by a very large measure.

While they continue to smash records, AI models (including ours) still make use of traditional physics-based models in several ways. For starters, all AI models are trained on historical weather data and predictions produced by conventional systems. Without them, the model would have to rely on raw, real-time observations for training data, without historical context.

AI models also inherently lack an advanced understanding of physics, so traditional models provide a baseline to ensure that AI-generated predictions are physically plausible. This assistance is especially important during extreme weather events, when physics-based models can help AI models simulate rare conditions based on atmospheric principles.

How We Built our AI Weather-Forecasting Model

When the WindBorne team set out to build the initial version of WeatherMesh, we had three main goals. First, it had to be inexpensive to run. Second, it needed to be at least as accurate as the top physics-based models. Third, it had to deliver forecasts with a high spatial resolution, providing fine-grained predictions on the scale of tens of kilometers.

We decided to use an architecture based on what are called transformers—the same technology that powers large language models like ChatGPT—because transformers can process huge datasets efficiently once they’re trained. This architecture includes what AI mavens refer to as an encoder-processor-decoder structure. The encoder transforms raw weather data—things like temperature, wind, and pressure—into a simpler compressed format known as latent space, where patterns are easier for the model to work with. The processor then runs calculations in this latent space to predict how the weather will change over time. To create longer-range forecasts, we simply run the processor step multiple times, with the output of the last prediction step serving as the input for the next. Finally, the decoder translates the results back into real-world weather variables.

We trained our first weather model at our headquarters using a cluster of a few dozen Nvidia RTX 4090 graphics processing units (GPUs), which cost far less than relying on cloud-computing services to handle hundreds of terabytes of atmospheric data. Setting up our own machines paid off. The hardware set us back about $100,000, but had we run all our training experiments in the cloud instead, it easily would have cost four times as much.

Copper wires threaded through the plastic help control the gas-venting system. Christie Hemm Klok

The balloon material is only 20 micrometers thick, and each balloon weighs less than 2 kilograms when fully assembled. Christie Hemm Klok

The first version of WeatherMesh was smaller, faster, and cheaper to operate than the AI weather models created by tech giants. During training, it used about one-fifteenth the computing power of DeepMind’s GraphCast and one-tenth that of Huawei’s Pangu-Weather. Its small size makes its stellar performance all the more notable: It outperformed both those AI models and traditional physics-based models.

The early accuracy gains of WeatherMesh can be attributed to our data-collection method. Our GSBs collect 30 to 50 times as much data as do conventional balloons, and we feed that data directly into WeatherMesh. We measured our model’s accuracy based on frequency of errors when compared with other physics- and AI-based models. In 2024, we beat both Huawei’s Pangu-Weather and DeepMind’s GraphCast to become the most accurate AI forecasting model in the world. At the time this article is being published, in October 2025, WeatherMesh retains the lead.

Our initial version of the model took in data and output forecasts at 0.25-degree resolution (about 25 kilometers per grid cell) to match the resolution of ERA5, a widely used historical weather dataset. Today, WeatherMesh also includes a component that can provide forecasts for selected locations at a resolution of about 1 km.

Most AI weather models train on historical datasets like ERA5, which organizes decades of atmospheric data into a consistent framework. But we also wanted WeatherMesh to run “live,” ingesting real-time balloon observations and up-to-date analyses from the U.S. and European agencies. That transition was challenging, because most AI models perform worse when they shift from carefully curated historical data to messy real-world feeds.

To address this issue, we built specialized adapters based on a type of neural-network architecture known as U-Net, which excels at learning spatial features across different scales. These adapters translate real-time data into the same internal format used for WeatherMesh’s training data. In this way we preserved the benefits of training on ERA5 while still delivering accurate real-time forecasts.

Building On Success With WeatherMesh-4

Following the success of our initial WeatherMesh model, we released the second, third, and fourth versions of the model in rapid succession. WeatherMesh-4 predicts standard atmospheric variables at 25 vertical levels throughout the atmosphere. It also predicts a wide range of conditions at the surface, including temperature and dewpoint at 2 meters from the ground, wind speed at 10 meters and 100 meters, minimum and maximum temperatures, precipitation, solar radiation, and total cloud cover. It can produce a full forecast every 10 minutes based on the latest observations. In contrast, traditional weather models update every 6 hours.

Traditional weather balloons stay aloft for only a few hours and don’t go far from their launch sites. Annie Mulligan/Houston Chronicle/Getty Images

We’ve run extensive benchmarks to compare the latest version of WeatherMesh with other popular forecasting systems. We’ve found that the model’s predictions for the Earth’s surface and atmosphere are up to 30 percent more accurate than those from the traditional model from the European Centre for Medium-Range Weather Forecasts, and also surpass results from DeepMind’s latest model, GenCast, on most evaluations.

Building an end-to-end system means the entire pipeline must work in harmony. Our balloon constellation can’t afford to wait 12 hours for a new forecast; it needs near-constant refreshes to navigate the skies. Meanwhile, the AI model uses fresh atmospheric data from the balloons to improve the accuracy of its forecasts. Balancing these requirements forced us to get creative about how we moved the data and ran the model, but ultimately we produced a powerful system that’s fast and responsive.

What’s Next for WindBorne

In the coming years, our goal is to expand our Atlas balloon constellation to about 10,000 GSBs flying at any time, launched from about 30 sites worldwide. To achieve that goal we’ll need roughly 300 launches per day, or 9,000 per month. By 2028, we believe the entire globe could be under near-continuous observation by Atlas, from the remote Pacific to the polar ice caps. And we continue to test the boundaries of what is possible: WindBorne recently kept a balloon aloft for a record-breaking 104 days.

We’re not aiming to make physics-based weather models obsolete. We see a future where AI and traditional methods operate side by side, each reinforcing the other. Governments, researchers, and corporations can lean on these improved forecasts to guide disaster preparedness, aviation, supply-chain logistics, and more. Our planet’s weather challenges are only going to intensify as the climate continues to change, and improved forecasts are key to helping us prepare.

Looking back at Hurricane Milton, it still feels surreal that our balloons managed to ride into a storm of that scale. Yet that was the moment WindBorne proved that a new and agile system could deliver real value where legacy methods fall short. In a world where an extra 12 or 24 hours of warning can mean the difference between safety and devastation, end-to-end AI forecasting offers a revolution in how people can observe, predict, and protect themselves from the most powerful forces on Earth.