61% vs. 26%: The number behind the real US-China AI divide

Editor's note: Tian Feng is the dean of Thinking Fast and Slow Research Institute, an independent AI think tank. The article reflects the author's opinions and not necessarily the views of CGTN.

In a factory outside Shenzhen, a two-legged robot leans over an assembly line and picks up a car door hinge, its knee joints whining faintly with every squat. Halfway around the world, in a converted warehouse south of San Francisco's Market Street, a dozen engineers stare silently at a screen full of falling numbers – a "loss curve," the holy chart of machine learning.
Both scenes are called artificial intelligence (AI). Neither looks much like the other.

That gap has just been measured, not guessed at. In August, the US-based think tank RAND Corporation's Center on AI, Security, and Technology published a detailed headcount of the US and Chinese AI industries: 1,181 companies – 743 American and 438 Chinese – sorted by what each firm actually builds, sells and patents. The headline number is simple. Sixty-one percent of American AI firms are pure software shops. Only 26% of Chinese firms are. The rest are busy building bodies – robots, cars, drones – for their algorithms to live in.

Twins at birth, strangers by adulthood

Here's the twist RAND's researchers did not expect: on almost everything except that one number, the two industries look like siblings. Roughly 28% of US firms and 27% of Chinese firms build on the transformer architecture, the same family of neural networks behind ChatGPT and DeepSeek alike. About nine in ten American AI companies and three in four Chinese ones sell "applied AI products" – an app, a diagnostic tool, a logistics dashboard – rather than raw model access. Even the fabled "AI arms race" between frontier labs turns out to be a small neighborhood: only about 19% of firms in either country are foundation-model developers – those that pretrain general-purpose models from scratch, the way OpenAI or Alibaba do. Everyone else is quietly writing software for hospitals, factories and insurance companies.

Where the two markets genuinely part ways is in what their AI is for. Chinese firms cluster in manufacturing, transportation and energy – the physical economy. American firms cluster in healthcare, scientific research, and cybersecurity – the knowledge economy. Proportionally, China's sample includes more than five times as many humanoid-robot companies as America's – 12% of Chinese AI firms versus just 2% of US ones; America builds far more diagnostic-software startups. Same toolbox, different job site.

Markets are shaped by what they're fed

Why the split? The Scottish economist and philosopher Adam Smith gave the short answer nearly 250 years ago, long before anyone imagined a chip: the division of labor is limited by the extent of the market. China still runs the world's densest manufacturing supply chain – motors, batteries, actuators, sensors, all a phone call away – so its AI naturally crawls into machines. It also has an industrial policy that says so out loud: China named "embodied intelligence" a national priority in 2025 and published a national standards framework for humanoid robots the following year. On top of that, well over half of China's AI firms in the RAND sample sit inside larger conglomerates – Alibaba, Tencent, and their peers – giving them a built-in factory floor or delivery fleet to deploy on.

America's advantage runs the other way. Its capital markets can throw tens of billions of dollars at a single language model in the hope of a breakthrough, a bet that Wall Street, not the state, is willing to underwrite. Its deep bench of medical, scientific, and security expertise gives software AI rich, high-value problems to chew on. Call it comparative advantage, the old David Ricardo idea updated for the algorithmic age: nobody is smarter, they're just specializing in what their own economy hands them.

Cheap is its own kind of smart

There is a second race running quietly underneath the robots-versus-chatbots story, and it is about price. RAND tracked real developer traffic on OpenRouter, a gateway that routes application programming interface (API) calls to hundreds of AI models. In late 2024, American models carried roughly nine-tenths of that traffic; Chinese models were nearly invisible. By April 2026, the two sides were dead even – 49% to 47% – propelled by open Chinese models like DeepSeek-V3 and the Qwen series, which score close to top US systems on independent benchmarks while costing a quarter to a fifth as much per token. It is not that Chinese labs suddenly out-thought their rivals at the very top of the scale; American systems still lead on raw intelligence scores. It is that "good enough, four times cheaper" is exactly the kind of quiet, compounding advantage that reshapes a market from the bottom up – the same brutal logic economist Joseph Schumpeter once dubbed "creative destruction."

Competing on two tracks, talking on a third

None of this describes a winner-take-all sprint. It describes two economies solving different halves of the same puzzle, and the gap between them is exactly where cooperation is cheapest and most useful: safety testing for humanoid robots working alongside people; shared standards that allow a sensor built in Shenzhen and software written in Seattle to actually talk to each other; joint research on AI for drug discovery and disease diagnosis. Geoffrey Hinton, who shared the 2024 Nobel Prize in Physics for laying the mathematical groundwork of today's neural networks, has said plainly that the technology is moving faster than most people – himself included – expected, and that its risks are too large to be managed by rivalry alone. It is an engineering fact.

Plenty left to do

British mathematician and logician Alan Turing, writing in 1950 at the very start of this story, put it simply: "We can only see a short distance ahead, but there is plenty there that needs to be done." Seventy-five years later, one country has taught its machines to talk and reason; the other has taught them to walk and lift. Sooner or later, something is going to need to do both – and when it does, the two halves of this story, brain and body, will finally have to meet.