WindBorne Systems began as a college side project, when a group of students launched weather balloons for fun. “Eventually this became a company and we realized that it could make a pretty major impact in how weather forecasting works,” explained Joan Creus-Costa, one of WindBorne’s founders.

The company’s breakthrough idea was to create weather balloons that last weeks or even months, rather than just hours. This dramatically increases the volume and coverage of weather data, particularly over oceans and remote regions where traditional balloons can’t reach. Today, WindBorne operates the largest balloon constellation in the world while also developing state-of-the-art deep learning models for real-time forecasting.

The Challenge

As WindBorne expanded from hardware to AI-driven modeling, their compute needs grew exponentially. Training deep learning models requires not only powerful GPUs but also massive amounts of data. “We have north of 100 terabytes of training data on DigitalOcean,” Creus-Costa noted. “Weather forecasting is very weird in that the datasets are a lot bigger than most other fields.”

Their team had previously cobbled together an on-premises cluster using gaming GPUs—even buying CPUs and RAM in cash. While scrappy, that setup couldn’t keep up with the scale of modern deep learning. “We’ve mostly outgrown them for actually training models and we just rely on DigitalOcean for that,” Creus-Costa said.

Why DigitalOcean

WindBorne evaluated several cloud providers but found that DigitalOcean uniquely fit their needs. “One thing we saw was that DigitalOcean offered us the ability to have these Gradient™ AI Bare Metal GPUs, which had really high interconnect between the different nodes,” explained ML engineer Anuj Shetty. “That was really important for us as we do a lot of multi-node training runs.”

The ability to store and access data efficiently was also critical. “Being able to have all that storage locally unlocks things that you just could not do by streaming from elsewhere,” said Creus-Costa. The team found that DigitalOcean’s infrastructure allowed them to train models faster and more cost-effectively than other cloud options.

Another differentiator was the system resources available. “We have two terabytes of RAM on each of the nodes, which we leverage to train larger models,” Shetty said. “DigitalOcean’s ability to provide us with all of the resources that we need, which we often can’t find in other places, has been really helpful.”

Products and Use Cases

WindBorne primarily relies on Gradient™ AI Bare Metal GPUs for training its WeatherMesh models, supplemented by CPU Droplets for large-scale processing. In one instance, they spun up more than 100 CPU Droplets to quickly process data. The infrastructure is flexible enough to support both development and production workloads.

Reliability and support have also stood out. Shetty recalled, “Whenever we’ve had issues, support has been really on top of it and highly responsive.” Creus-Costa added a memorable anecdote: “Within 24 hours of having the Gradient™ AI Bare Metal GPUs, we bricked one of them by hammering the disk too much. DigitalOcean quickly gave us endpoints to hard reboot the machines, which let us recover immediately.”

Looking Ahead

WindBorne’s vision is bold: to build a “planetary nervous system” for weather. As Shetty described, “There are two parts to that. One is the sensory organs—global sounding balloons that collect weather data around the globe. And then there’s the brain, which directs the balloons and translates those inputs into weather intelligence.”

Creus-Costa added, “The more we work on this, the more we think there’s so much room to improve forecasts. That improvement comes from two places: the balloons, and better deep learning modeling. Being able to run increasingly bigger training runs with more data is the way to do it.”

With DigitalOcean providing the compute backbone for their modeling efforts, WindBorne is pushing the frontier of weather forecasting—making more accurate predictions accessible to everyone.