America Outspends China on AI by 23 to 1. The Lead Is Down to 2.7%.

The Stanford Institute for Human-Centered AI released its annual AI Index this week, and the number that will generate the most argument is 2.7. That is the percentage gap between the best American AI models and the best Chinese ones, measured by performance on standardized benchmarks. In 2023, that gap was somewhere between 17 and 31 percentage points. It has largely evaporated.

The comparison the report draws is worth sitting with. The United States spent $285.9 billion on private AI investment in 2025. China spent $12.4 billion. That is a 23-to-1 ratio in favor of the US, and it has produced a lead of less than three percentage points. Whether that represents spectacular progress for China or a poor return on investment for the US is probably both.

Anthropic's Claude Opus 4.6 currently leads the global Arena leaderboard with a score of 1,503. ByteDance's Dola-Seed-2.0-Preview sits at 1,464. The two models have been swapping positions near the top since early 2025, when DeepSeek's R1 briefly matched the best American models and set off a minor panic in Silicon Valley. That pattern has continued. China's open-source models have been particularly competitive, with GLM-5.1 recently topping the global coding leaderboard.

Where China leads

Benchmark performance is not the whole picture. China files 69.7% of global AI patents. Its researchers author 23.2% of the world's AI publications. The country has installed 295,000 industrial robots, versus 34,200 in the United States, a 9x difference. Its electricity grid maintains reserve margins above 80%, which matters a great deal if you are trying to run a data center at scale.

The talent pipeline is also moving in China's direction. The flow of AI researchers to American universities is down 89% since 2017, with 80% of that decline happening in the last year alone. That number deserves more attention than it usually gets. Model performance today reflects investment decisions made years ago. Talent flows now predict the leaderboard in 2030.

The wider picture

Beyond the competitive framing, the report documents how fast the underlying technology is moving. Coding performance on the SWE-bench went from 60% to near 100% in a single year. AI models now answer graduate-level science questions correctly 93% of the time, above the 81% expert baseline. Generative AI reached 53% population adoption in three years, faster than smartphones.

The environmental costs are now large enough to appear in the data. Training xAI's Grok 4 produced 72,816 tonnes of CO2 equivalent, roughly what 17,000 cars emit in a year. Global AI data center capacity has reached 29.6 gigawatts, equivalent to New York State's peak electricity demand.

The report also notes that public trust in government AI regulation sits at 31% in the United States, the lowest figure globally. The world average is 54%. Whatever the US is doing on that front, it is not working.