The Startup Making the Weather a Hot Investment

The weather has never been hotter.

The past few years have seen a boom in the weather prediction industry, with AI-based weather models from the likes of Google DeepMind, Huawei, and Nvidia consistently outperforming traditional models. Most of that work has been research-oriented, but today the startup Brightband emerged from stealth with $10 million in Series A funding and a unique plan to commercialize generative AI weather modeling. Instead of trying to go up against Weather.com, Brightband is tailoring models to specific industries such as insurance, finance, agriculture, energy, and transportation. The round was led by Prelude Ventures.

Weather forecasting has traditionally been the domain of the public sector, with the most widely used models coming from the U.S. National Weather Service and the European Center for Medium-Range Weather Forecasts. Brightband’s CEO Julian Green told me that private companies haven’t been able to break in “because it has cost so much to have billion dollar supercomputers,” which are required to run today’s so-called “numerical” weather models.

These models rely on complex atmospheric equations based on the laws of physics to predict future weather patterns, and because of their computational intensity, are usually only updated four times daily. It’s possible then that AI-based weather prediction could thus actually reduce energy demand — because while it takes a lot of energy to train an AI model, after that’s done, generating forecasts is simple. “So instead of six hours to have a forecast, it takes under a second. Instead of using a billion dollar supercomputer, you’re using a laptop,” Green told me.

AI models like Brightband’s are trained on decades worth of past weather data, and when fed a snapshot of current conditions, can predict what will come next, much like ChatGPT does with text. “Think about the weather AI prediction problem as predicting the next frame in a radar sequence,” Green told me.

He said that customizing forecasts for particular industries will also be as simple as querying a large language model. A wind farm operator could, for example, “just take an attached file of historical wind energy production, and throw it in there and say, hey, tell me what the wind energy is going to be like next week.” Likewise folks in the aviation industry could have the model tell them if their plane’s wings are likely to ice up, utilities could get detailed insight into expected energy demand and generation, and finance companies could get up-to-the-minute information about weather-sensitive commodities. Previously, companies would’ve had to build their own forecasting teams or hire third-party advisors to get such specific predictions.

Brightband wants to further differentiate itself from the types of models that tech companies have already built by using only raw data inputs to generate its forecasts, from sources such as satellites, weather balloons, and radar systems. Perhaps surprisingly, this is not the way that most models currently work. Because there are data gaps, such as over oceans and in the developing world, the datasets used to train today’s AI weather models, Green explained, “smear the available data over a three-dimensional grid of the globe,” diluting the accuracy of both the real-time weather and presumably the resulting forecasts.

It’s hard to say how much more accurate using only raw data inputs will be, because “that’s what nobody has done yet,” Green told me. Data gaps are still an issue of course, but Green told me that Brightband’s approach will also allow the company to better analyze when and where filling these gaps would add the most value.

Brightband says it hopes to publish a paper by year’s end with an open-source version of its forecast model, alongside evaluation tools to assess its performance.