Stanford AI Index Report: US-China Gap Narrows to 2.7%—What Does This Really Mean?

Stanford’s Human-Centered AI Institute (HAI) publishes an AI Index Report every year. This is the ninth edition—423 pages covering every aspect of global AI development. It’s become a “must-read” in the industry, widely cited by governments, research institutions, and media worldwide.

But one number this year caught everyone’s attention: the US-China AI capability gap has narrowed to 2.7%.

Let me first explain how this number was derived. HAI used a comprehensive evaluation framework covering model performance, research output, patent counts, investment scale, talent pool, and other dimensions. The final calculation shows the US scoring 100, China scoring 97.3—a gap of 2.7 percentage points.

After this number came out, I saw two very different reactions: excitement—“finally caught up!”—and skepticism—“how can it be this small? Is there something wrong with the data?”

Honestly, I think both reactions are too extreme. Let’s calmly analyze what this number actually means.

First, this “2.7%” is a composite metric. If you break down individual dimensions, the picture gets more nuanced. In some areas—like large model applications, computer vision, speech recognition—China has indeed achieved “parity” or even “localized leadership.” But in other areas—fundamental research, high-end chips, original algorithm frameworks—gaps persist, sometimes significantly.

Second, this evaluation framework itself has limitations. HAI’s methodology is based on “quantifiable indicators,” but some important capabilities are hard to quantify—like innovation quality, technical influence, ecosystem-building capacity. These “soft indicators” don’t show up in composite scores, but they may determine long-term competitive dynamics.

Here’s another easily overlooked detail: the report’s statistical period covers all of 2025. In other words, it reflects the state “over the past year.” But AI moves so fast that this year’s situation might already be different.

Okay, after all this “cold water pouring,” I have to admit: this 2.7% number does reflect an objective trend—the US-China AI capability gap is narrowing. This isn’t an illusion or data manipulation, but genuine progress.

But I want to emphasize: “narrowing gap” does not equal “caught up,” much less “comprehensive leadership.”

I’ve been looking at comparative tests of open-source models recently. Chinese models perform well in some scenarios, but there’s still room for improvement in stability, generalization capability, and handling long-tail cases. Users might not encounter these details often, but when they do, the experience gap becomes very noticeable.

There’s an even more important question: how will this “2.7%” gap evolve going forward?

I have two different hypotheses.

The first is “continued narrowing.” Because China’s AI industry has formed a complete chain—from data, computing, algorithms to applications—every link is iterating rapidly. This “full-stack capability” will drive sustained catching up.

The second is “gap stabilization or even reversal.” Several reasons: one, the US has deeper accumulation in fundamental research and core technologies, potentially more staying power; two, geopolitical factors leading to technology restrictions might create bottlenecks at critical nodes; three, AI competition is entering a new phase—shifting from “chasing models” to “chasing ecosystems” and “chasing applications,” requiring different capabilities.

I personally lean toward the second possibility. Not out of pessimism, but because I think the next few years will be the critical watershed for AI competition. Whether we can go from “catching up to running alongside” to “running alongside to leading” depends on many non-technical factors—policy environment, capital efficiency, talent density, openness to international cooperation.

Finally, let’s talk about what this 2.7% number means for us as developers and practitioners.

I think the most important thing is: stop holding onto the stereotype that “domestic is inferior to foreign.” In many practical application scenarios, Chinese models are perfectly capable. Rather than blindly trusting “foreign big tech,” spend time understanding different models’ characteristics and choose the one that best fits your needs.

At the same time, stay clear-headed: narrowing gap doesn’t mean gap eliminated. We still have a lot of fundamentals to practice, many foundational capabilities to build. These two things aren’t contradictory.

What do you think about this 2.7%? Feel free to share your thoughts.