Blog

The AI Weather Revolution: Who Is It Actually Built For?

Jan Gunash · Co-Founder & CGO · March 16, 2026

The best questions are not always the comfortable ones. Like: when we say AI is transforming weather forecasting, who is that transformation actually for? When we say the models are getting better; better for whom, measured where, and built on whose data?

These are not rhetorical questions. They have empirical answers,and those answers sit quietly underneath some of the most impressive numbers in climate technology right now.

A Genuine Breakthrough

Something remarkable has happened in weather prediction. NOAA recently deployed a new suite of AI-driven global forecasting models - the efficiency numbers are almost impossible to believe. A single 16-day global forecast from their new system uses only 0.3% of the computing resources of its traditional predecessor (source: NOAA). A hybrid model combining AI and physics-based forecasting now outperforms either approach alone. For the first time, an operational weather center is running a system that is genuinely better than what came before it.

The implications for energy are direct. Better forecasts mean better predictions for wind and solar output, smarter grid management before extreme weather arrives, and longer planning windows that reduce reliance on fossil fuel backup. Every hour of improved forecast skill translates into real operational value across the energy system.

What the Headlines Do Not Say

Here is the thing about AI weather models that almost no one is talking about: they learn from data. Specifically, from historical reanalysis datasets - reconstructions of past atmospheric conditions that blend real observations with physics-based model output. The gold standard is ERA5, produced by the European Centre for Medium-Range Weather Forecasts, which most frontier AI models use as their training foundation.

ERA5 looks like a uniform global dataset. But where real observations are sparse - ground stations, ocean buoys, radiosondes - the reanalysis fills gaps with model estimates. In data-rich regions, ERA5 is tightly constrained by actual measurements. In data-sparse regions, it is essentially a model talking to itself (source: arXiv 2603.05710). An AI trained on this is not learning from reality in those areas. It is learning from another model's best guess about reality - inheriting the biases of the physics model it was built to replace, without that model's ability to flag its own limitations. The result is not intentional. It is the automated reproduction of an existing infrastructure gap.

That gap shows up in the numbers. Research has found that a seven-day weather forecast in a wealthy nation outperforms a one-day forecast in a low-income country (source: arXiv 2603.05710). The gap in forecast skill between rich and poor countries is not narrowing with AI. In some analyses, it is widening. The communities receiving the least accurate weather forecasts are the ones most exposed to the consequences of getting them wrong. Cyclone preparedness in Bangladesh. Drought monitoring across the Sahel. Early flood warnings in Central Africa.

A Model of a Model

It gets more complicated. Most frontier AI climate and weather systems are not trained on raw observations. They are trained on reanalysis products that blend real observations with physics-based model output. In regions where real observations are sparse, the reanalysis is filling gaps with model-generated estimates. An AI trained on ERA5 is not learning from reality in those regions. It is learning from another model's best guess - creating what researchers call a model-of-a-model dependency, where biases from the original physics model become embedded in the AI system trained to replace it, compounding the uncertainty in exactly the places where it is already highest (source: arXiv 2603.05710).

A Data Problem With Policy Consequences

The equity problem does not stop at weather. It runs straight into climate adaptation planning. Governments making decisions about sea walls, drought-resistant crops, or early warning systems are working from climate data. If that data systematically underestimates risk in certain regions, the adaptation plans built on it will be wrong in ways that are hard to detect (source: MDPI, ECAS-7 2025). Biased training data skews projections toward regions that already dominate the dataset - underestimating risks elsewhere and driving adaptation resources toward areas that are already better served (source: Frontiers in AI, 2025).

The Global South is not just underrepresented in training data. Only 32 countries globally host the hyperscale AI infrastructure required for cloud-based deployment of these models. The communities that most need better forecasting have the least capacity to build it (source: arXiv 2603.05710).

The Path Forward

The root cause is not the algorithm. It is the observation network. Fixing this requires investment in ground stations across sub-Saharan Africa, ocean buoys in the Indian Ocean, and radiosondes across Central Asia and the Amazon. These are not glamorous investments. They do not generate benchmark scores. They are the foundational inputs that determine whether AI climate models work for the whole world, or just the parts of it that were already well served.

It also requires changing how performance is measured. Global average accuracy scores conceal regional disparities. Evaluation frameworks disaggregating results by region and income level would make the equity problem visible in the metrics that model developers and funders actually use. Climate Change AI has begun cataloguing the specific data gaps that most constrain AI climate applications - identifying which missing observations, if collected, would unlock the highest-value use cases globally (source: Climate Change AI).

The AI weather revolution is real. The efficiency gains are real. But the trajectory of that improvement is not neutral. Without deliberate intervention, the models will continue to improve fastest in the regions that already have the best data. The algorithm learns from what it is given. Give it better data, from more places, and it will learn better. That is not a technical argument. It is a choice about what kind of climate information system we are building - and for whom.

Sources:

Want to contribute your thoughts or findings in data and AI with climate, energy, or nature? Message Jan Gunash on LinkedIn or send an email to jgunash@projectcoeus.com.