Stanford’s AI Index 2025 Tracks China Closing the U.S. Gap

Stanford has just published its AI Index 2025, one of the sector’s most comprehensive assessments. The report portrays an industry in which the United States still leads in investment and frontier models, but where China has rapidly narrowed the quality gap.

The other major force emerging from the data is economic: using advanced models has become dramatically cheaper. That helps explain why generative AI has moved from a technology demonstration into products and processes at companies of all sizes.

China closes in on major benchmarks

U.S. companies and institutions released 40 AI models considered notable in 2024, compared with 15 from China and three from Europe, according to Stanford’s AI Index. The gap in output remains wide, and the United States retains a dominant position in creating the most advanced systems.

But the performance gap has narrowed. Chinese models are now approaching U.S. systems on widely used tests of general knowledge, coding and mathematics. On MMLU — an exam covering questions from multiple subjects — the gap between the top models from the two countries narrowed to 0.3 percentage points. On HumanEval, a code-generation benchmark, it was 1.6 points; on MATH, 2.4 points.

That does not mean every Chinese model is equivalent to the U.S. leaders, or that national origin is enough to judge a system. Benchmarks are limited measures, and models can perform very differently on real-world tasks, in different languages or in professional settings. But it does show that the technical lead the United States held just two years ago can no longer be taken for granted.

This convergence has geopolitical and commercial consequences. A market with competitive models from more than one bloc reduces dependence on a small group of U.S. providers. At the same time, it intensifies competition for chips, talent, data and developers.

Using a model costs a fraction of what it once did

The decline in prices is one of the report’s most striking figures. Inference — the process of asking an already-trained model to respond or generate content — for a system with performance equivalent to GPT-3.5 became more than 280 times cheaper between November 2022 and October 2024.

The cost fell from about $20 per million tokens to $0.07. Tokens are the units of text models process; roughly speaking, they can be words, parts of words or punctuation marks.

The reduction does not eliminate the costs of deploying AI at scale, especially if a company handles confidential documents, needs highly reliable answers or uses more expensive reasoning models. But it does change which projects are viable. An assistant that analyzes thousands of queries, translates documentation or classifies incidents may no longer be an experiment reserved for large tech companies.

The report also documents efficiency gains in smaller models. Systems with far fewer parameters — the internal values they learn during training — are achieving results that required much larger models just a few years ago. That is a significant trend for phones, personal computers and organizations that want to run AI at lower cost or within their own infrastructure.

Investment remains concentrated in the United States

The U.S. advantage remains highly visible in private capital. In 2024, private investment in generative AI in the United States reached $33.9 billion, up 18.7% from the previous year. China received $9.3 billion and the United Kingdom $4.5 billion.

Across private AI investment as a whole, the United States amassed $109.1 billion in 2024. The concentration of financial resources helps explain why U.S. labs can take on increasingly expensive training runs: Stanford estimates that training GPT-4 cost about $79 million in compute, while Gemini 1.0 Ultra reached roughly $192 million.

Those figures also point to a central tension in the sector. Inference is getting cheaper for users, but building frontier models requires resources that very few companies can assemble. AI is becoming more accessible to use while becoming more concentrated to produce.

Business adoption accelerates alongside regulation

Some 78% of organizations surveyed by McKinsey said they used AI in 2024, up from 55% the previous year. The figure does not mean all of them have transformed their businesses or achieved clear returns: adopting a tool and genuinely redesigning a process are two different things. Even so, it confirms that AI has entered the operational agenda of most of the companies surveyed.

The regulatory response is growing as well. U.S. federal agencies introduced 59 AI-related regulations in 2024, more than twice as many as in 2023. Worldwide, 42 countries passed 131 laws related to the technology during the year.

The AI Index also recorded 233 AI-related incidents in 2024, a 56.4% increase from 2023. The figure covers problems such as fake content, bias, system failures and misuse. The technology’s expansion is pushing the discussion beyond model capabilities: who controls them, how they are evaluated and what mechanisms are in place when they fail all matter.

Stanford’s snapshot does not point to an immediate change in U.S. leadership. It shows something more precise: the race no longer depends solely on who unveils the most capable model, but on who can make it affordable, useful, reliable and available to millions of people and businesses.