The US Lead on China in AI Has Nearly Vanished, Stanford Finds

Two years ago, American AI looked like it had a commanding lead. US frontier models outperformed their closest Chinese competitors by 17 to 31 percentage points on standard benchmarks. The 2026 Stanford AI Index, published April 13, reports that gap has collapsed to 2.7 percentage points as of March 2026.

That number deserves careful reading. Benchmark margins at this scale reflect rough parity more than a durable edge. When DeepSeek's models briefly matched the leading US frontier model in early 2025, it was treated as a surprise. Stanford's data suggests it was the beginning of a trend.

Where Each Country Leads

The picture is genuinely mixed. China now leads the world in AI patent grants, holding 69.7% of global grants, and in sheer publication volume. US researchers still produce the most-cited work, but China's output has grown harder to dismiss as derivative.

Private investment remains overwhelmingly American. The US attracted $285.9 billion in AI private investment last year; China attracted $12.4 billion. The US also controls a significant share of high-end GPU compute — a strategic advantage that US export controls are explicitly designed to preserve.

The Coding Benchmark Jump

The report includes one finding that even seasoned AI observers will find striking. On SWE-bench Verified — a benchmark testing AI agents on real software engineering tasks — performance jumped from roughly 60% to near 100% in a single year. That kind of gain, sustained over twelve months, has not been seen on any comparable engineering benchmark before.

Whether that means AI can now reliably write production software, or that the benchmark has become saturated, is an open question. Stanford's authors lean toward the latter. But they note that even if the benchmark ceiling has been hit, the underlying capability gains are real.

What the Report Does Not Settle

The Stanford AI Index has become one of the more reliable annual snapshots of the field, but it has limits. Benchmark performance is a proxy, not a verdict. Several leading Chinese models are not publicly evaluated on the same test sets as their US counterparts, making direct comparisons partially speculative. And the investment gap — more than 23:1 in the US's favor — could compound differently depending on how efficiently each country converts capital into capability.

What the report does establish, convincingly, is that the gap the US held in 2023 no longer exists in any meaningful form. The 2.7-point margin is within the noise of how these benchmarks are constructed. Whatever lead American AI has, it can no longer be taken for granted.

The full 2026 AI Index is available at Stanford HAI. SiliconAngle's coverage of the China-US comparison is here.