China vs USA: Who Is Leading the AI Weather Forecasting Race

For decades, weather forecasting was dominated by the European Centre for Medium-Range Weather Forecasts (ECMWF) and a handful of national meteorological agencies running massive supercomputers. But in the last three years, AI has upended that order — and the two countries pushing hardest are China and the United States. Google DeepMind's GraphCast, NVIDIA's FourCastNet, and China's Fengshun and FuXi models are all producing forecasts faster and, in some cases, more accurately than traditional numerical weather prediction. But the two countries are playing fundamentally different games.

The Contenders: What Each Side Brings

USA: GraphCast, FourCastNet, and the Research Frontier

Google DeepMind's GraphCast, introduced in late 2023, was the breakthrough that made the world take AI weather forecasting seriously. Trained on nearly 40 years of ECMWF reanalysis data, GraphCast can generate a 10-day global weather forecast in under 60 seconds on a single Google TPU v4 — a task that previously required hours on a supercomputer with thousands of CPUs. In a landmark paper published in Science, GraphCast outperformed ECMWF's operational HRES system on 90% of 1,380 verification targets.

NVIDIA's FourCastNet takes a different approach, using a Fourier neural operator architecture that achieves even faster inference speeds. It can generate a 100-member ensemble forecast — essentially 100 different possible weather scenarios — in a fraction of the time traditional ensemble systems require. NVIDIA has also open-sourced the model and made it available through its Earth-2 platform, a digital twin of the Earth's climate system designed for both research and commercial applications.

Microsoft's Aurora and ClimaX models, along with various academic efforts from MIT, UC Berkeley, and other institutions, round out the US landscape. The common thread: these models are primarily developed by private companies and academic labs, and their primary output is research papers, benchmark scores, and API access.

China: Fengshun, FuXi, and the Operational Deployment Edge

China's approach is led by the China Meteorological Administration (CMA) and national research institutes, with models designed from the start for operational deployment rather than research benchmarks. The "Feng" (Wind) series — Fenglei, Fengqing, and Fengshun — covers the full forecasting spectrum from nowcasting (minutes to hours) to sub-seasonal (3-5 weeks).

Fengshun, the long-range model, ranked first globally in the ECMWF's AI Weather Quest competition's third season — a direct head-to-head comparison with the best models from around the world. Fengqing generates refined global forecasts in just three minutes with 10.5 days of usable forecast range. Fenglei focuses on the hardest problem: severe convective weather like tornadoes and hailstorms, achieving a 25% improvement in strong echo forecasting.

FuXi, developed by the Chinese Academy of Sciences and Fudan University, adds a critical dimension that most AI weather models lack: physics integration. The FuXi-CNOPs variant combines AI with nonlinear dynamics theory to identify error-sensitive regions in typhoon forecasts, reducing path prediction errors by up to 32.33% compared to traditional ensemble methods. It also requires 40% fewer computational samples — a meaningful efficiency gain.

Head-to-Head: Where Each Side Leads

Dimension	USA	China	Edge
Research output	GraphCast in Science, multiple Nature papers	ECMWF competition wins, growing publication record	USA
Operational deployment	Limited; mostly experimental	Full national deployment; 1,909+ offices using AI warnings	China
Global reach	API access, research collaborations	MAZU deployed in 7 countries, cloud services in 40+	China
Model speed	GraphCast: 60 sec for 10-day	Fengqing: 3 min for global; Fengshun: weeks ahead	Tie
Physics integration	Mostly pure data-driven	FuXi-CNOPs: AI + nonlinear dynamics hybrid	China
Open-source	FourCastNet open-sourced; GraphCast code available	Fenghe LLM open-sourced at WAIC 2026	USA (wider)
Computing resources	Google TPU v4/v5, NVIDIA H100 clusters	Domestic AI chips; computing vouchers from government	USA
Developing world support	Limited; mostly commercial	1,000+ professionals trained from 100+ countries	China

Different Games, Different Scoreboards

The most important thing to understand about this competition is that China and the US are measuring success differently. The US approach — led by Google, NVIDIA, and Microsoft — treats AI weather forecasting primarily as a research problem and a commercial opportunity. Success is measured in papers published, benchmarks beaten, and potential enterprise customers. The models are impressive, but they largely exist in the cloud, accessed through APIs, and their deployment in actual emergency operations is limited.

China's approach treats AI weather forecasting as a public service. Success is measured in countries served, forecasters trained, and lives protected. The MAZU platform doesn't just produce forecasts — it integrates satellite data, AI models, warning dissemination, and emergency response coordination into a single operational system. When Cyclone Maila approached Papua New Guinea, a Chinese AI system wasn't just predicting the storm — it was generating preparedness reports with specific recommendations for emergency managers.

The MAZU Factor: When AI Meets Diplomacy

At the 2026 WAIC, Chinese President Xi Jinping announced that MAZU would expand to 30 countries over the next five years. This is not just a technology announcement — it's a diplomatic initiative. China is positioning AI weather forecasting as a global public good, similar to how the US positioned GPS in the 1990s. The difference is that MAZU is a full-stack solution: satellites, models, platform, and training, all provided as a package.

The China Meteorological Administration has established permanent overseas cloud nodes in Singapore and Cairo to improve system accessibility. Fengyun meteorological satellite data now serves users in more than 130 countries and regions. Nearly 1,000 meteorological professionals from more than 100 developing countries have participated in China's early warning training programs. The MAZU Scholarship program, launched in June 2026, supports meteorological professionals from Belt and Road partner countries working on early warning systems.

This is a fundamentally different model from the US approach. American AI weather models are technologically impressive but institutionally fragmented — they live in different companies, use different APIs, and serve different customer bases. There is no US equivalent of MAZU: a unified, government-backed platform designed to be shared with developing countries.

Where the US Maintains Real Advantages

Despite China's deployment lead, the US maintains significant advantages that shouldn't be underestimated.

Research excellence. GraphCast's publication in Science and the broader ecosystem of AI weather research at American universities and companies represent genuine scientific leadership. The conceptual breakthroughs — graph neural networks for weather, Fourier neural operators, diffusion models for ensemble forecasting — largely originated in the US and Europe.

Computing infrastructure. Google's TPU v4 and v5 systems, NVIDIA's H100 and upcoming Blackwell GPUs, and the massive cloud infrastructure of AWS, Google Cloud, and Azure give American AI weather models access to computing resources that Chinese models cannot easily match, especially given export controls on advanced chips.

Private sector dynamism. The American ecosystem of startups, venture capital, and commercial weather services (like The Weather Company, AccuWeather, and Tomorrow.io) creates market pressure for continuous improvement. Chinese weather services remain predominantly state-run, which ensures coordinated deployment but may slow innovation in some areas.

Where China Is Pulling Ahead

China's advantages are structural rather than technological.

Operational scale. The "Zhijingda" AI warning system is deployed across all provincial-level meteorological bureaus and 1,909 city and county-level offices. When a severe storm is approaching, these systems generate automatic alerts faster than human forecasters can. No US AI weather system has anything close to this level of operational integration.

Physics-AI hybrid models. FuXi-CNOPs represents a genuinely novel approach that most Western AI weather models have not attempted: combining data-driven AI predictions with physics-based nonlinear dynamics to identify exactly where forecasts are most uncertain. This hybrid approach addresses the "black box" problem that plagues pure AI weather models.

Deployment in the Global South. MAZU's deployment in Pakistan, Ethiopia, Mongolia, Djibouti, and other developing countries is not just a diplomatic achievement — it generates real-world feedback that improves the models. A system that works in the monsoon systems of South Asia and the sandstorms of Central Asia learns things that a system tested only on European and North American weather cannot.

What the Numbers Don't Capture

Benchmark competitions like the ECMWF AI Weather Quest provide useful comparisons, but they measure a narrow slice of what matters. A model that achieves 99% accuracy on a benchmark but takes 12 hours to run is less useful in an emergency than a model with 95% accuracy that runs in 3 minutes. A model that predicts rainfall perfectly but can't communicate that prediction to the people in harm's way is a scientific achievement, not a life-saving one.

The AI weather forecasting race is often framed as a competition to build the most accurate model. But the real competition is about building the most useful system — and on that metric, the two countries are playing different games entirely. The US is winning the model accuracy game. China is winning the deployment game. Which matters more depends on whether you're a researcher publishing a paper or a meteorologist trying to warn a community about an approaching storm.

Conclusion: A Race Where Both Can Win

Unlike the AI chip race — where US export controls create a zero-sum dynamic — the AI weather forecasting race is one where both countries can advance simultaneously. Better models from the US research community can be incorporated into China's operational platforms. China's deployment experience in diverse climate zones can inform better model design everywhere. The real beneficiaries are not the countries themselves but the communities — particularly in the developing world — that gain access to warning systems they could never afford to build independently.

The question "who is leading" may ultimately be the wrong one. The better question is: are AI weather forecasting systems actually reaching the people who need them? On that measure, China's operational-first approach currently has a clear edge — but the US research engine continues to push the frontier of what's possible. In a world of increasingly extreme weather, both matter more than most people realize.