We Are the Frontier

Seattle's Ai2 Builds Open AI Climate Model

Google-funded NeuralGCM+ aims at decades-long projections, a horizon beyond Google's one-year weather model, with a three-year timeline.

Updated on Nov. 4, 2026 in Tech

Seattle's Ai2 Builds Open AI Climate Model

Live Poll

Do you trust long-range climate projections to guide local water planning decisions?

Google.org is providing about $5 million in funding and Google is providing $1.8 million in cloud credits to the Seattle-based nonprofit AI institute Ai2 for a three-year effort to build NeuralGCM+, an open-source climate projection model. This is a project announcement; the model is in development and has not yet been released.

Why it matters

Google's existing NeuralGCM work focuses on weather forecasts looking ahead a year at most, while decades-long climate projections are a different problem that Google has less interest and expertise in. Ai2 is taking on that longer-horizon gap, which signals AI-based climate modeling moving from short-range forecasting toward the multi-decade projections used for long-term planning.

The project targets climate projections spanning five to 50 years, a horizon far longer than the one-year range of Google's NeuralGCM weather work. The 7-degree Fahrenheit ocean warming test is the stress condition under which NeuralGCM's 2024 predictions failed to remain physically sensible.

The players

Ai2

A Seattle-based nonprofit AI institute founded in 2014 by Paul Allen, whose climate team builds the Ai2 Climate Emulator (ACE) model family.

Google.org

Google's philanthropic arm, providing about $5 million in funding for the Ai2 climate project.

NeuralGCM

Google's hybrid physics-and-AI climate and weather model, published in Nature in 2024, on which Ai2's NeuralGCM+ builds.

The details

NeuralGCM mixes traditional physics equations with machine learning, a hybrid approach in which conventional physical rules constrain the AI component. Ai2 plans to connect NeuralGCM+ to an ocean and sea ice model so that multi-decade simulations stay stable, and to build more established physics into the model. The team will also check the model's predictions for a warmer world against traditional climate models. The original physics-plus-AI plan was reportedly set aside at one point because AI could not be trained well inside older Fortran-based climate models.

Timeline

  1. 2014: Paul Allen founded the Ai2 nonprofit AI institute.

  2. Around 2016: Paul Allen's climate modeling effort began at Vulcan.

  3. 2021: The climate modeling group moved to Ai2.

  4. 2024: Google published NeuralGCM research in Nature.

  5. November 4, 2026: Google's funding for Ai2's climate model was announced.

The Tech Race

Google's NeuralGCM was built for forecasts looking ahead a year at most, leaving decades-long climate projections an open gap. Ai2's NeuralGCM+ effort positions a Seattle nonprofit to compete in that longer-horizon space, alongside its existing ACE model family.

The model is in development, so no climate projections are available to use yet. Ai2 expects to release NeuralGCM training code within the first year, which would let outside researchers build on the work before the full model is finished.

The takeaway

Watch whether Ai2 can keep NeuralGCM+ physically stable in a warmer-world test against traditional climate models, which the project's timeline suggests could extend beyond three years. The first-year release of training code is a concrete checkpoint for whether the open-source commitment holds.

What happens next

Ai2 expects to release NeuralGCM training code within the first year of the three-year project.

Further reading

For more on the research behind this push, see our Artificial Intelligence coverage from Seattle.

Source note: This article includes information reported by GeekWire.

Live Poll

Do you trust long-range climate projections to guide local water planning decisions?