For seventy years, weather forecasting has been the domain of states. Supercomputers, satellites, networks of radiosonde stations, armies of atmospheric physicists: the right to predict the weather was reserved for the rare countries capable of paying the price. This monopoly has just collapsed. A neural network trained by DeepMind now predicts global weather ten days in advance in less than sixty seconds on a Cloud TPU v4, whereas the best physical model in history took hours on supercomputers — the contract signed by the ECMWF with Atos for its latest supercomputer (BullSequana XH2000) amounts to more than 80 million euros over four years. And it does it better.

This is not a laboratory demonstration. It is an operational shift that is redrawing the geopolitics of a public service that seven billion humans use without seeing it.

The Essentials

  • GraphCast, DeepMind’s model, surpasses the ECMWF HRES model on 90% of the 1,380 standard verification targets, according to a study published in Science in December 2023.
  • Pangu-Weather by Huawei has been integrated into the official forecasts of the China Meteorological Administration for typhoon forecasting, marking growing adoption of these tools in official weather services.
  • A global ten-day forecast runs in less than sixty seconds on a Cloud TPU v4, compared to several hours on the ECMWF’s supercomputers.
  • GraphCast is available open source (code under Apache 2.0 license, weights under CC BY-NC-SA 4.0) and Pangu-Weather is as well, opening unprecedented access to world-class quality forecasts.
  • Weather services in the Global South gain access to forecasting quality without precedent, but at the cost of new dependence on foreign private operators for a strategic public good.

Seventy Years of Physics Surpassed by Thirty-Six Hours of Training

To understand the magnitude of the change, one must first understand what existed before.

The ECMWF HRES (High Resolution Forecast) model is the Rolls-Royce of global weather forecasting. Developed by the European Centre for Medium-Range Weather Forecasts, based in Reading, England, it runs on one of Europe’s most powerful supercomputers. Its architecture is based on partial differential equations that describe the physical behavior of the atmosphere: energy conservation, thermodynamics, fluid dynamics. Seventy years of atmospheric science condensed into millions of lines of code. Its budget exceeds 100 million euros per year. Its five-day forecasts display high reliability according to its own verification metrics.

GraphCast performs better on 90% of the 1,380 targets tested. These targets cover temperature, pressure, wind, and humidity at different altitudes across the entire globe. The advantage is particularly pronounced at horizons beyond five days, where traditional physical models degrade most rapidly. It is also measurable in the trajectory of tropical cyclones: during the 2023 season, GraphCast predicted the intensification of Hurricane Lee with three additional days of advance compared to the ECMWF.

The difference in method is total. GraphCast has no knowledge of the equations of atmospheric physics. It was trained on forty years of ERA5 reanalysis data produced by the ECMWF itself — in other words, the reference model has, in a sense, served as professor to its successor. From this archive, GraphCast learned to predict the state of the atmosphere at the next time step. Result: a global ten-day forecast in less than sixty seconds on a Cloud TPU v4.

Pangu-Weather or How Huawei Enters Public Infrastructure

DeepMind is not alone. Pangu-Weather, developed by Huawei, FuXi (from the Chinese Academy of Sciences), and NeuralGCM (Google DeepMind), form a second front of AI models competing with comparable or superior performance to the ECMWF model on specific criteria.

A decisive step is taken when the China Meteorological Administration integrates Pangu-Weather into its official forecasts for typhoon prediction. This is no longer a research tool. It is state public infrastructure, powered by a model developed by a private company.

The decision is logical from a performance perspective. Pangu-Weather shows superior accuracy on typhoon trajectories in the western Pacific, a region where forecasting stakes are massive: more than 30 million people exposed to typhoons in China each year, a coastal economy representing nearly 60% of national GDP. When an AI model predicts the passage of a typhoon better, it translates into lives saved and better-targeted evacuations.

But the integration of Pangu-Weather into official services raises a question that weather agencies are beginning to formulate cautiously: what happens when a public organization depends on a private provider for a public safety service? The transparency of the ECMWF’s physical models, documented and auditable, contrasts with AI models whose weights are often proprietary and whose internal reasoning is difficult to interpret.

The Dilemma of Global South Countries: The Best Forecast in History, Delivered by Strangers

This is where progress creates its own paradox.

A national meteorological service in Ethiopia, Bangladesh, or Peru did not, until recently, have the means to exploit a model of ECMWF quality. Access licenses exist, but the local computational capacities to run downscaling models, personnel trained in interpreting high-resolution outputs, budgets to maintain these infrastructures: all of this remained out of reach. Weather forecasting remained a luxury of the North.

AI models redistribute this access. GraphCast is available open source (code under Apache 2.0 license, weights under CC BY-NC-SA 4.0). A national meteorological service can run a world-class global forecast on modest infrastructure, without a supercomputer. For countries where rainfed agriculture represents 40 to 60% of GDP and where early warnings of cyclones save concrete lives, this is a first-order change.

The World Meteorological Organization has explicitly recognized it: the deployment of AI models in weather services of low- and middle-income countries figures among its priorities for 2024-2027. Pilot projects are underway in Kenya, India, Indonesia, and Nigeria, with support from the WMO and the United Nations Development Programme.

But accessing GraphCast in open source does not mean independence. The model is maintained by DeepMind, a subsidiary of Alphabet. If Google decides to modify, restrict, or monetize access, agencies that have built their workflows around GraphCast find themselves in a vulnerable position. This is not a theoretical hypothesis: it is exactly what has happened in other sectors where public services integrated Google tools without a substitution plan.

The problem is not so much the quality of the model as the absence of governance. Weather forecasting has been a global public good since the creation of the World Meteorological Organization in 1950. Radiosonde, satellite, and ocean buoy data circulate freely between states by international treaty. AI models, however, belong to private firms subject to national legal regimes, commercial imperatives, and geopolitical dynamics. The Sino-American technological rivalry observed in nuclear fusion and semiconductors is now settling into the global meteorological infrastructure.

What Public Weather Services Are Doing in Response

The major agencies are not passive. They are integrating.

The ECMWF has launched an internal program to develop hybrid AI models, combining the physical constraints of its traditional models with deep learning architectures. The objective is to obtain interpretable models, whose outputs can be audited and whose errors can be attributed to identifiable causes. Google Research’s Metnet-3, which incorporates physical constraints into an AI model, represents a similar path.

The American NOAA (National Oceanic and Atmospheric Administration) announced in 2024 partnerships with several laboratories to develop foundational AI models for weather forecasting, emphasizing open source and auditability. The objective is not to leave the entire field to private actors.

Météo-France is experimenting with integrating AI outputs into its operational forecasting chains, treating them as one source among others rather than as a replacement for physical models. This ensemble approach, which aggregates outputs from different models to reduce uncertainty, is probably the most robust path in the short term.

These responses show one thing: public agencies are taking the threat to their expertise seriously, but also the opportunity these tools represent to improve their own forecasts. The risk is not that states disappear from meteorology, but that they become integrators of models they do not fully understand, on subjects for which they remain accountable to their populations.

Why Ten-Day Forecasting Is a Geopolitical Stake

We must name what is at stake concretely beyond technical performance.

Medium-range weather forecasting conditions decisions that are not trivial. Agricultural futures markets move on ten-day forecasts. Evacuation strategies before a cyclone are decided seventy-two hours in advance. Electric grid management during heat waves depends on five-day forecasts. Military operations have always integrated weather as a tactical variable. Whoever holds the best forecasting models holds a form of informational advantage on all these fronts.

During the Cold War, the United States and the USSR had agreed to share real-time weather data, including at moments of maximum tension. It was the World Meteorological Organization that organized these exchanges. The logic was simple: atmospheric disturbances do not respect borders, and a model that only works correctly on its own territory is a model that does not work.

This framework of cooperation still holds for raw data. It does not hold for AI models, which are developed outside international treaties, in private companies subject to export controls and government pressures. The fact that Pangu-Weather is developed by Huawei, a company whose links with the Chinese state are the subject of sustained attention from American and European regulators, is not a detail. The same tension between technological sovereignty and dependence on foreign suppliers structures other critical industries, from semiconductors to energy.

The question posed is not theoretical: if a major geopolitical crisis leads to restrictions on access to AI models developed by rival firms, can weather services that depend on them switch to alternatives in real time? Today, the answer is uncertain.

The Challenge of the Next Decade: Who Governs Foundational Models

AI weather models are following a trajectory we are beginning to recognize in other domains. The first models are published open source, demonstrate their superiority, are adopted massively. Then second and third generation models, more powerful, become proprietary. The open source window closes.

GraphCast is open source today. Nothing guarantees that GraphCast 2 or GraphCast 3 will be. The same dynamic is observed in large language models, structural biology models, image generation tools. AI imposes everywhere the same governance question: whoever controls foundational models controls the cognitive infrastructure of a sector.

For meteorology, the answer to this problem probably passes through three parallel paths. The first: the ECMWF and its equivalents develop their own AI models, public and auditable, with international financing. The second: an international governance framework, under the auspices of the WMO, defines transparency and access requirements for models integrated into public services. The third: national weather services maintain a minimum autonomous capacity in traditional physical models, so they do not find themselves without a safety net if AI models degrade or become inaccessible.

None of these three paths is simple or rapid. The first requires public funding in a constrained budgetary context. The second runs up against the technological sovereignty of companies and states. The third runs counter to the economic logic that pushes agencies to reduce their costly computational capacities as soon as cheaper alternatives exist.

What weather forecasting illustrates, in advance of other sectors, is the concrete form that dependence on AI takes when it touches strategic public goods. The performance is real, the progress measurable, the benefits for vulnerable populations tangible. The governance dilemma is equally so. It is public agencies, governments, and international organizations that must decide whether they want to remain architects of the service or merely be its users.

Sources

  1. Articledge — AI Weather Forecasting (ECMWF / Science / Nature synthesis)
  2. Lam et al., “GraphCast: Learning skillful medium-range global weather forecasting”, Science, December 2023 (vol. 382)
  3. Bi et al., “Accurate medium-range global weather forecasting with 3D neural networks”, Nature, July 2023 (Pangu-Weather, Huawei)
  4. ECMWF — annual reports and technical documents on hybrid AI-physics models (ecmwf.int)
  5. World Meteorological Organization (WMO) — Strategic Plan 2024-2027 (wmo.int)
  6. NOAA — press releases on AI partnerships for weather forecasting 2024 (noaa.gov)
  7. Science Article – GraphCast (Lam et al. 2023)
  8. Google DeepMind – Official GraphCast Publication
  9. Google DeepMind – GraphCast Blog
  10. GitHub – google-deepmind/graphcast
  11. Nature – Pangu-Weather (Bi et al. 2023)
  12. ECMWF – Key facts and figures
  13. ECMWF – Atos Supercomputer Contract
  14. WMO – AI Congress for Forecasts (2024)
  15. NOAA EPIC – EAGLE Project
  16. Huawei – Pangu-Weather Nature Press Release