China in AI research
China writes the largest share of the most cited research in three of the four AI technologies tracked here; the United States leads high-performance computing.
Data: OpenAlex, research published 2021–2025, pulled 23 October 2026. Definition version 0.5.
China's share rose in all four; the US share fell in all four
Each technology's share for China, the United States and the next country, 2021–2023 against 2024–2025. On the right, China's change.
2021–20232024–2025Share of the 10% most cited papers
The numbers behind the chart
| Technology | China | United States | Next country |
|---|---|---|---|
| Large language models | 15.3% → 29.2% | 29.1% → 18.0% | United Kingdom 4.4% → 3.6% |
| AI chips | 42.8% → 49.3% | 15.7% → 7.9% | South Korea 9.5% → 14.4% |
| Advanced robotics | 30.7% → 32.4% | 17.5% → 14.4% | Italy 7.4% → 5.3% |
| High-performance computing | 12.1% → 16.1% | 39.5% → 32.3% | Germany 6.6% → 8.6% |
China leads 20 of 30 sub‑areas
The United States leads nine and Italy leads one; they sit at the foot of the chart. Each paper is sorted into one sub-area by an AI model reading its title and abstract (92% agreement with hand labels on 200 papers).
Share of the sub-area's papers in the top 10%: China against the strongest other country
The other 10 sub-areas, all led by China
Share of the sub-area's papers in the top 10%: China against the strongest other country
The numbers behind the chart
| Sub-area | Technology | Papers | China | Strongest other country |
|---|---|---|---|---|
| Optoelectronic synapses and in-sensor computing | AI chips | 104 | 75.8% (1st) | South Korea 8.6% |
| Ferroelectric memory and synaptic devices | AI chips | 49 | 53.5% (1st) | South Korea 10.2% |
| Tactile sensing and electronic skin | Advanced robotics | 54 | 47.2% (1st) | Singapore 7.0% |
| Multimodal and vision-language models | Large language models | 283 | 49.0% (1st) | United States 10.0% |
| Reservoir and physical computing | AI chips | 30 | 49.2% (1st) | South Korea 14.1% |
| Memristors and resistive-switching synapses | AI chips | 178 | 43.6% (1st) | South Korea 17.2% |
| Synaptic transistors (electrochemical, ionic, 2D, organic) | AI chips | 61 | 52.3% (1st) | South Korea 26.1% |
| Legged locomotion and humanoid robots | Advanced robotics | 66 | 35.5% (1st) | United States 13.3% |
| Retrieval-augmented generation and knowledge graphs | Large language models | 339 | 35.7% (1st) | United States 14.0% |
| Soft robots and artificial muscles | Advanced robotics | 289 | 40.0% (1st) | United States 20.5% |
| Code generation and software engineering | Large language models | 381 | 36.1% (1st) | United States 18.2% |
| Spiking neural networks and neuromorphic systems | AI chips | 66 | 35.0% (1st) | United States 17.9% |
| Motion planning and control | Advanced robotics | 48 | 28.7% (1st) | United States 13.1% |
| Fine-tuning, prompting and reasoning methods | Large language models | 308 | 33.0% (1st) | United States 19.6% |
| LLM agents and tool use | Large language models | 179 | 30.5% (1st) | United States 21.0% |
| Grasping and dexterous manipulation | Advanced robotics | 61 | 23.3% (1st) | United States 13.8% |
| Compute-in-memory and AI accelerator chips | AI chips | 114 | 35.2% (1st) | United States 27.1% |
| Job scheduling, energy efficiency and resource management | High-performance computing | 36 | 22.3% (1st) | United States 16.6% |
| Security, privacy and safety (incl. LLMs for cybersecurity) | Large language models | 191 | 24.3% (1st) | United States 19.1% |
| Model architecture, pre-training and scaling | Large language models | 253 | 26.9% (1st) | United States 23.2% |
| Supercomputer systems, processors, GPUs and interconnects | High-performance computing | 59 | 24.2% (2nd) | United States 25.5% |
| Domain applications (medicine, science, finance, industry) | Large language models | 687 | 22.5% (2nd) | United States 24.0% |
| Photonic AI accelerators | AI chips | 31 | 26.8% (2nd) | United States 28.7% |
| Evaluation, benchmarks and hallucination | Large language models | 473 | 16.8% (2nd) | United States 22.5% |
| ChatGPT overviews, education and societal impact | Large language models | 298 | 9.9% (2nd) | United States 18.1% |
| Robot foundation models and general-purpose robot learning | Advanced robotics | 30 | 16.6% (2nd) | United States 27.1% |
| Human-robot interaction and collaborative robots | Advanced robotics | 94 | 10.4% (3rd) | Italy 28.8% |
| Scientific simulation codes at scale (molecular, fluids, climate) | High-performance computing | 52 | 13.5% (3rd) | United States 37.7% |
| Workflows, data management and AI-HPC convergence | High-performance computing | 40 | 7.5% (3rd) | United States 49.8% |
| Programming models, performance portability and libraries | High-performance computing | 58 | 7.7% (2nd) | United States 52.1% |
Year by year: China overtook the United States in large language models
Share of the 10% most cited papers published in each year.
Fewer than 100 top papers in 2021 and 2022: single years are noisy.
Fewer than 100 top papers in 2021: single years are noisy.
Fewer than 100 top papers in every year: single years are noisy.
The numbers behind the charts
| Country | 2021 | 2022 | 2023 | 2024 | 2025 |
|---|---|---|---|---|---|
| China | 18.5% | 8.5% | 15.8% | 23.2% | 32.6% |
| United States | 17.8% | 35.3% | 28.8% | 22.5% | 15.4% |
| United Kingdom | 1.1% | 2.4% | 4.7% | 3.4% | 3.7% |
| Papers in the top 10% | 9 | 34 | 409 | 1,170 | 2,049 |
| Country | 2021 | 2022 | 2023 | 2024 | 2025 |
|---|---|---|---|---|---|
| China | 37.2% | 42.8% | 46.2% | 46.1% | 51.6% |
| United States | 14.5% | 19.6% | 13.3% | 11.2% | 5.5% |
| South Korea | 8.8% | 10.8% | 8.8% | 16.0% | 13.3% |
| Papers in the top 10% | 81 | 111 | 135 | 152 | 210 |
| Country | 2021 | 2022 | 2023 | 2024 | 2025 |
|---|---|---|---|---|---|
| China | 25.4% | 33.2% | 32.6% | 30.1% | 34.4% |
| United States | 21.5% | 16.6% | 15.1% | 17.0% | 12.1% |
| Italy | 6.8% | 7.0% | 8.3% | 5.7% | 5.0% |
| Papers in the top 10% | 108 | 125 | 135 | 156 | 179 |
| Country | 2021 | 2022 | 2023 | 2024 | 2025 |
|---|---|---|---|---|---|
| China | 10.8% | 19.8% | 5.8% | 15.9% | 16.4% |
| United States | 47.1% | 33.5% | 37.2% | 45.1% | 22.4% |
| Germany | 8.2% | 7.4% | 4.0% | 4.6% | 11.7% |
| Papers in the top 10% | 53 | 48 | 49 | 52 | 67 |
Among the 1% most cited papers, the United States leads advanced robotics
The 1% most cited papers are a tenth as many as the top 10%, so a few papers move them.
Top 10%Top 1%Share of each tier's papers
The numbers behind the chart
| Technology | China, top 10% | China, top 1% | United States, top 10% | United States, top 1% |
|---|---|---|---|---|
| Large language models | 27.4% | 27.8% | 19.4% | 26.1% |
| AI chips | 46.2% | 44.7% | 11.6% | 16.6% |
| Advanced robotics | 31.5% | 24.2% | 16.0% | 27.7% |
| High-performance computing | 13.9% | 6.9% | 36.3% | 37.6% |
Tsinghua University is in three of the four world top tens, more than any other Chinese institution
Chinese institutions in the world's top ten
- Tsinghua University1st in large language models, 6th in AI chips, 7th in advanced robotics
- Zhejiang University2nd in large language models, 7th in AI chips, 3rd in advanced robotics
- Shanghai Jiao Tong University4th in large language models, 1st in advanced robotics
- Peking University3rd in large language models, 4th in AI chips
- Harbin Institute of Technology7th in large language models, 6th in advanced robotics
- Fudan University1st in AI chips
- Chinese Academy of Sciences3rd in AI chips
- National University of Defense Technology4th in high-performance computing
Ranked among all institutions worldwide by share of the top 10%.
Who China writes the top papers with
- CNUSUnited States265 papers
- CNHKHong Kong217 papers
- CNSGSingapore161 papers
- CNGBUnited Kingdom125 papers
- CNAUAustralia76 papers
Shared papers in the top 10%, summed across the technologies.
With excess home citations removed, China's large language models share is 23.6%, from 27.4%, and it still ranks 1st
Citations from a paper's own country, against the share expected from that country's size in the field.
| Technology | Citations from China | Expected from its size | Excess |
|---|---|---|---|
| Large language modelsEvery top-10% paper | 49.4% | 20.8% | +28.6 pts |
| AI chipsUp to 100 single-country papers | 63.9% | 44.6% | +19.3 pts |
| Advanced roboticsUp to 100 single-country papers | 65.6% | 36.3% | +29.3 pts |
| High-performance computing26 single-country papers, a small sample | 62.9% | 13.3% | +49.6 pts |
Home citers include authors citing their own earlier work. Re-ranking with the excess removed is computed so far for large language models (China 27.4% → 23.6%).
Questions
Is China leading in AI?
China writes the largest share of the most cited research in three of the four AI technologies tracked here; the United States leads high-performance computing. Measured by each country's share of the most cited papers published 2021–2025.
Is China ahead of the US in AI research?
China is ahead in large language models, AI chips and advanced robotics (27.4% to 19.4%, 46.2% to 11.6% and 31.5% to 16.0%). The United States is ahead in high-performance computing (36.3% to 13.9%). Among the 1% most cited, the United States leads advanced robotics.
Is China's share growing?
China's share rose in all four; the US share fell in all four, comparing 2021–2023 with 2024–2025: large language models 15.3% to 29.2%, AI chips 42.8% to 49.3%, advanced robotics 30.7% to 32.4% and high-performance computing 12.1% to 16.1%.
Which Chinese institutions lead AI research?
Tsinghua University, Zhejiang University, Shanghai Jiao Tong University and Peking University: each is in the world's top ten by share of the most cited papers in at least one of these technologies.
Which countries come next after China and the US?
In large language models, United Kingdom, Germany and South Korea. In AI chips, South Korea, Singapore and Germany. In advanced robotics, Italy, United Kingdom and Germany. In high-performance computing, Germany, Italy and Spain.
What does this measure, and what does it not?
Each country's share of the most cited research papers published 2021–2025, with each paper's credit split between its authors' countries. Four technologies stand for AI here: large language models, AI chips, advanced robotics and high-performance computing. It does not measure which country builds the most capable models, how much computing power it has, chip production, investment or adoption. Definitions and method: each technology's page and the method page.
The data, and where to go next
- Large language modelsEvery country, institutions, papers
- AI chipsEvery country, institutions, papers
- Advanced roboticsEvery country, institutions, papers
- High-performance computingEvery country, institutions, papers
- China in robotics researchChina 31.5% of the top papers, the United States 16.0%
- China in semiconductor researchChina leads all three
- China in all 45 technologiesWhich way China is moving in each
- Is China ahead of the US in technology?All 45, head to head