Stanford puts the US lead over Chinese models at 2.7 per cent but does not say which benchmarks that is

The single most quoted number from Stanford's 2026 AI Index is that the American lead over Chinese models has narrowed to 2.7 per cent. It is being used to argue about export controls, chip diplomacy and national AI strategy, all of which are premised on the existence of a measurable American lead. The number is genuine and it is published. What is missing is the thing that would let anyone check it.

The report's own framing, on its landing page and in its technical performance chapter, is that the US and China AI model performance gap has effectively closed, with the top US model leading by just 2.7 per cent as of March 2026. The landing page attributes that leading model to Anthropic. The chapter, which is where the measurement sits, does not name the company at all. It adds that US and Chinese models have traded places at the top of performance rankings multiple times since early 2025, that DeepSeek R1 briefly matched the top US model in February 2025, and that the gap has fluctuated over the past year while remaining in single digits. That last point matters more than the headline: a 2.7 per cent gap measured on a single March date, in a series that has changed hands repeatedly, is a snapshot rather than a trend.

The chapter as published does not specify which benchmarks were used to produce the 2.7 per cent, nor which Chinese model it is measured against, nor whether the figure is an average across evaluations or a single leaderboard margin. Without that, the number cannot be reproduced. It can be cited, and it is being cited heavily, but a reader cannot open the underlying evaluation and confirm it. For a figure now load bearing in policy argument, that is a real limitation, and it is a limitation of the presentation rather than of the underlying research, which may well be specified in the full report volume.

The surrounding measures point in a less ambiguous direction, and most of them favour the United States by wide margins. The report counts 5,427 data centres in the United States, more than ten times any other country. Its research and development chapter records 59 notable models from the United States in 2025 against 35 from China, and states that industry produced over 90 per cent of notable AI models that year, while noting that training code, parameter counts, dataset sizes and training duration are no longer disclosed for several of the most resource intensive systems, including those from OpenAI, Anthropic and Google. Global AI compute capacity has expanded to 17.1 million H100 equivalents since 2022, with Nvidia controlling more than 60 per cent of it. China leads on a different measure: Chinese representation among the 100 most cited AI papers rose from 33 in 2021 to 41 by 2024.

The investment figures are frequently reported as though they conflict, and they do not. Stanford's summary article of 13 April 2026, listing twelve takeaways from the report, states that global corporate AI investment hit 581.7 billion dollars in 2025, up 130 per cent on the prior year, and that private investment within that total reached 344.7 billion dollars, up 127.5 per cent. The 344.7 billion is a global figure, not an American one. The same passage puts United States investment at 285.9 billion dollars, or 23.1 times the 12.4 billion invested in China, which is the number the report's landing page carries. The economy chapter is consistent with both, recording private investment growing fastest at 127.5 per cent and now accounting for 60 per cent of the corporate total. Stanford itself adds the caveat most often dropped in circulation: comparisons based on private investment alone probably understate Chinese capital, because the Chinese government also channels money through state initiated guidance funds, an estimated 912 billion dollars of which were deployed across all industries between 2000 and 2023.

Independent verification of the model gap is harder to obtain than it looks. Epoch AI maintains a database of more than 3,500 models with training compute, parameters, cost and power consumption, selecting notable models by criteria including state of the art performance, more than 1,000 citations, or over a million monthly active users, but its public landing page does not publish an aggregated country comparison. Artificial Analysis runs an Intelligence Index, currently at version 4.1.1, combining nine benchmarks across a field it lists as 608 models, and its roster includes Alibaba's Qwen alongside DeepSeek, Kimi, OpenAI, Anthropic, Google, Meta and Nvidia. Both rank individual models. Neither publishes the aggregated country level comparison that the 2.7 per cent asserts, which means that on the published material the figure currently stands largely on its own authority.

The labour market claims from the same report have travelled just as far. Its economy chapter states that employment for software developers aged 22 to 25 fell nearly 20 per cent from 2024, while employment among older colleagues grew. Lightcast, which has contributed job posting data to the Index since 2021, reports that AI skills now appear in 2.5 per cent of US job postings, up 55 per cent on the previous year and 297 per cent on a decade earlier, with agentic AI the fastest growing cluster, rising from 0.06 per cent of postings in 2024 to 0.23 per cent in 2025.

What is not known is the composition of the 2.7 per cent: which evaluations, which models, and whether the margin has held since March 2026.