China nearly closes AI performance gap with US: Report

China Economic Net

Beijing: China has nearly closed the artificial intelligence (AI) performance gap with the United States, with its leading models achieving near parity with American counterparts on key benchmarks, a Stanford University report released April 7 showed.

Top U.S. and Chinese AI models achieved near parity on these performance benchmarks by the end of 2024, Stanford’s latest AI Index report found. The U.S. lead over China shrank to 0.3% on MMLU, 8.1% on MMMU, 1.6% on MATH, and 3.7% on HumanEval tests, the data showed. That contrasts sharply with U.S. advantages of 17.5%, 13.5%, 24.3%, and 31.6% respectively on the same benchmarks at the end of 2023, according to the annual analysis.

These benchmarks test distinct AI capabilities. MMLU assesses broad knowledge; MMMU tests combined text and image understanding; MATH measures complex mathematical problem-solving; HumanEval evaluates code generation, CEN reported on Friday.

Notably, top-tier U.S. models have generally required significantly more training compute than their Chinese counterparts, the analysis found. Since late 2021, the compute used for the top 10 Chinese language models scaled about three times per year, slower than the five-times-per-year rate seen elsewhere since 2018, according to Epoch AI data cited in the report.

This compute difference is reflected in estimated costs, with some prominent Chinese models like DeepSeek-V3 reportedly trained for millions of dollars, compared to figures exceeding $100 million or even approaching $1 billion cited for some leading U.S. models.

China leads globally in AI publications and patents filed, the report said. Chinese institutions produced 15 ‘notable’ AI models in 2024, second only to the U.S. (40 models) and ahead of Europe (3 models).

The report cited DeepSeek’s R1 model from China ranking among top global AI systems. Separately, the performance lead held by the best U.S. model over the best Chinese model narrowed from 9.26% in January 2024 to 1.70% by February 2025, the data indicated.

This near parity was reached despite U.S. private AI investment ($109.1 billion in 2024) significantly exceeding China’s ($9.3 billion), the Stanford report noted.

Published annually by Stanford’s Institute for Human-Centered Artificial Intelligence (HAI), the AI Index report is a comprehensive resource tracking global AI trends.