Stanford: China Nearly Matches the US in AI Models

Title: Stanford: China Nearly Matches the US in AI Models

Excerpt: Stanford’s AI Index 2025 documents the narrowing gap between Chinese and US models, a 280-fold drop in inference costs, and record investment and adoption.

The AI Index 2025, the annual report from Stanford’s Institute for Human-Centered Artificial Intelligence (HAI), depicts an AI industry that is more capable, affordable and widespread—but also less dominated by the United States. The gap between the best US and Chinese models has narrowed to the point of nearly disappearing, while the cost of using models with performance comparable to GPT-3.5 has fallen 280-fold in two years.

Published on April 7, the report also records a sharp rise in investment, business adoption and regulation. The conclusion is not that all countries are competing on equal terms: the United States retains a clear lead in private capital and the production of leading models. But China is no longer a secondary player when it comes to performance.

The US lead has narrowed to a razor-thin margin

According to Stanford’s measurements, in March 2025 the best US model outperformed the best Chinese model by just 1.7% across the main performance benchmarks. That puts the two countries at virtually the same level at the technological frontier, although the rankings can change with every new release.

The United States still leads in the number of notable models. In 2024, it developed 40 models considered leading, compared with 15 in China and three in Europe. It also accounts for a disproportionate share of the private capital and infrastructure needed to train advanced systems.

The difference, then, depends on what is being measured. The United States produces more frontier models and attracts far more money; China is increasingly competitive on quality and has a large-scale industrial and scientific base. The progress of models such as DeepSeek has intensified that competitive pressure and challenged the notion that Chinese systems need to lag far behind to be useful.

Advanced models are much cheaper to use

One of the report’s most consequential practical findings is the drop in inference costs. Inference is the process by which a trained model responds to a prompt. The AI Index estimates that the cost of achieving performance comparable to GPT-3.5 fell from $20 to $0.07 per million tokens between November 2022 and October 2024—a 280-fold reduction.

Tokens are the units models use to process text. In practice, this drop allows a company to run many more queries on the same budget and makes applications viable that would have been too expensive two years ago.

The reduction is also changing the competitive landscape. Value no longer lies solely in access to the most powerful model, but in integrating it well, controlling consumption, protecting data and turning its responses into reliable processes. A model that is somewhat less capable but cheaper and faster may be more useful to a company than a frontier model with a high price tag.

The cost of training the largest systems, however, continues to rise. Stanford estimates that training GPT-4 required about $79 million in computing capacity, while Gemini Ultra came in at nearly $192 million. These estimates do not represent the companies’ total budgets, but they show the concentration of resources required at the industry’s frontier.

Business adoption reaches a new high

The report says that 78% of organizations reported using AI in 2024, up from 55% the previous year. The figure includes use in at least one business function and does not amount to full automation: in many cases, AI is being used as an assistant for drafting, analyzing information, coding or handling queries.

The expansion of generative AI has been especially rapid. Its use has moved from internal experiments to production tasks, although problems with accuracy, security and measuring return on investment persist. Reported adoption also does not guarantee that tools are integrated into core processes or that they deliver sustained cost reductions.

The AI Index finds mixed signals in the labor market. Companies are deploying systems capable of handling increasingly complex tasks, but performance remains uneven when an activity requires multiple steps, extended planning or interaction with changing information. Improvements on coding and reasoning tests do not eliminate the gap between solving an isolated task and working reliably for hours.

More investment, and more rules

The United States remains the financial center of AI. In 2024, it received $109.1 billion in private investment, nearly 12 times more than China, with $9.3 billion, and 24 times more than the United Kingdom, with $4.5 billion, according to the Stanford report. Generative AI attracted $33.9 billion in private investment worldwide.

Regulation has also gained weight. The United States passed 59 AI-related regulations in 2024, compared with one in 2016. Around the world, references to artificial intelligence in legislative initiatives rose sharply, signaling that the technology has moved from being an issue reserved for laboratories and innovation departments to taking up space on government agendas.

That regulatory growth is arriving as capabilities advance faster than evaluation systems. Models now outperform humans on some specific tests, but they still fail at tasks requiring common sense, error control or consistency. Stanford thus points to a paradox: AI is more capable on established benchmarks, but measuring its reliability in real-world settings remains difficult.

The 2025 picture shows a less lopsided race, with lower barriers to entry for everyday use. The United States retains its advantage in capital, leading models and computing capacity; China has nearly closed the performance gap; and falling prices are giving smaller companies access to capabilities once reserved for major labs. The next battle will increasingly shift from model launches to cost, reliability and effective adoption.