Google Just Open-Sourced the World's Best Hurricane Model. The Data That Prices the Damage Still Isn't.

The Model That Called Melissa Three Days Out

Hurricane Melissa made landfall near New Hope, Jamaica, at 1pm on 28 October 2025, a Category 5 storm with sustained winds of 185 miles an hour and a central pressure of 892 millibars, tying two records for the strongest Atlantic hurricane ever measured at landfall. It was the first Category 5 landfall anywhere in the Atlantic basin since Hurricane Dorian in 2019. The National Hurricane Center's post-storm report puts the toll at 45 dead in Jamaica, part of 95 regional deaths across Jamaica, Haiti, the Dominican Republic and Cuba, and roughly $8.8 billion in losses on the island alone, from flattened crops in Westmoreland and St Elizabeth to landslides that cut off entire communities.

Before any of that happened, DeepMind's WeatherNext Cyclones model was already running behind the scenes. Working with the National Hurricane Center during the 2025 season, the model forecast Melissa's rapid intensification, the jump from a moderate storm to catastrophic in under two days, roughly three days ahead of landfall, with close to 100% confidence. Independent reporting on the model's performance puts its gain over prior operational tools at about a full extra day of usable warning: a three-day forecast now carries the accuracy that used to require waiting for the two-day mark. In a region where evacuation orders, shelter logistics and storm shutters all run on hours, not days, that gap is not a rounding error.

The Forecast Just Became a Commodity

What changed on 6 August 2026 is not that the model got better. It is that DeepMind gave it away. The research behind WeatherNext Cyclones ran in Nature the same week, and the code, model weights and three variants (the full Cyclones model, the operational WeatherNext 2, and a lightweight WeatherNext 2-mini that runs in a free Google Colab notebook) went up on GitHub under an Apache 2.0 licence for the software and a Creative Commons Attribution 4.0 licence for the model outputs. Both licences allow commercial use with attribution. That is a deliberate departure from earlier DeepMind weather releases, which carried research-only terms. The ensemble behind the current version scaled to 1,000 members, wide enough to capture rare, high-consequence scenarios like rapid intensification rather than just the average track.

For any forecasting vendor, energy trader or insurer that previously paid for proprietary cyclone models, or built inferior ones in-house, that licence change is the moment the technology stops being a moat. A start-up with a GPU and a GitHub account can now stand up cyclone forecasting that matches what the National Hurricane Center used operationally last season. The track and intensity prediction, the part of climate risk analysis that used to separate a serious weather-data company from a hobbyist, is now the part anyone can have for free.

What Doesn't Get Solved by a Better Forecast

An extra day of warning is genuinely useful, and Melissa's death toll would almost certainly have been higher without it. It does not, on its own, tell an insurer which specific parish holds the highest concentration of uninsured structures, or tell a government which coastal informal settlements have no digitised address at all, let alone a flood-risk rating. Insured Caribbean tropical cyclone losses between 2000 and 2023 add up to roughly $64 billion, and industry estimates put uninsured losses at three to four times that figure again. Most of that shortfall is not a forecasting failure. It is an exposure-mapping failure: nobody had catalogued what was there to lose in the first place.

Reinsurers feel this gap directly. Across the region, an estimated $90 to $130 billion in climate risk goes mispriced every year, largely because the actuarial input, building stock, informal settlement density, historical damage-to-intensity ratios for a given coastline, simply does not exist in a form any model, open source or proprietary, can query. A free forecast improves the input on one side of that equation, the storm itself. It leaves the other side, what the storm hits, exactly where it was.

Where OYA AI Sits After the Forecast Is Free

OYA AI, Maestro AI Labs' climate intelligence product, targets a 72-hour hurricane lead time against a historical baseline closer to six to eight hours for many Caribbean communities. That number was never the whole pitch. OYA AI pairs cyclone prediction with an exposure layer built through Data Archaeology, Maestro's field-collection division: informal settlement mapping, uninsured structure counts, and historical storm-to-damage relationships gathered parish by parish, largely because no scraped dataset or public weather archive contains that signal. It was never posted online for a model to learn from. It had to be collected directly, with the consent of the communities and institutions holding it.

An open-source forecast does not compete with that layer. It raises the floor for everyone building a climate product, which means the exposure data becomes the only remaining place to differentiate. A reinsurer evaluating two climate-risk vendors next quarter is not going to ask which one has the better storm track anymore, because the answer will increasingly be the same for both. They will ask which one can tell them what a Category 4 landfall actually does to a specific 40-kilometre stretch of coastline with a known building stock. That question still has one answer fewer vendors than the forecasting question does.

Reading This as an Investment Signal

For a board or investor evaluating any climate-risk AI company today, WeatherNext's release adds a specific diligence question that did not carry this much weight twelve months ago. If a company's pitch still leans on forecast accuracy as its primary differentiator, that differentiator now has a free, Nature-published, Google-backed substitute available to any competitor. The durable question is what sits underneath the forecast: does the company hold exposure data a rival cannot simply download, and was that data collected with documented consent rather than scraped, given that AI training-data litigation, most visibly Anthropic's $1.5 billion settlement approved in July 2026, has turned undocumented data sourcing into a disclosed liability rather than a quiet shortcut.

The broader Caribbean AI ecosystem is where this data-first model was built before open-source forecasting made it a market necessity elsewhere. Maestro AI Labs is a product of the StarApple AI network, the Caribbean's first artificial intelligence company, founded by Adrian Dunkley in Kingston, Jamaica in 2023. Dunkley remains the region's most recognised authority on applied AI adoption and climate data governance, and the same consented, region-first collection model that underwrites OYA AI and Data Archaeology runs through the wider network he built.