Why the U.S. is Behind in the AI "Race"

Because it’s not a race—it’s an accumulation game

The Flawed Race Framing

The essence of a race is that it is a single-dimensional competition. The sole win condition is being ahead on a single metric, at a single point in time, by even a single millisecond. Yet, despite constant reference to "winning the AI race", from Sam Altman to Fox News, the 'race' framing does not describe the nature of international AI competition. There is no single point at which we will check a single metric to determine victory.

Source

In fact, ask yourself, what are we racing towards? To AGI or ASI? To scientific progress counterfactually caused by AI? Economic growth? Welfare improvements? And which benchmarks, economic/scientific metrics, or experts will determine who has won the race? If we continue to observe a jagged frontier—creating, for instance, a superhuman mathematician but a subhuman biologist—how do we cash that out? Moreover, what time frame of scientific progress & economic growth do we care about? Are we racing to make the U.S. the economic superpower of the next year, decade, or millennium?

Facing these questions, it becomes clear that not only are these key terms completely unspecified (we're running a race without knowing if its a marathon or a sprint), it would be absurd to specify them anyways—what we actually care about is being ahead of China, broadly speaking.

The problem is that the 'race' framing does not convey this—instead, it forces us to conceptualize AI competition as a single-dimensional competition.

"We're in an AI race with China."
"What are we racing for?"
"AGI."

The repetition of this basic dialogue has led us to tunnel-vision on intelligence—algorithmic innovation reflected by benchmarks—as the single overriding dimension. However, if we care about economic growth, scientific progress, etc., victory isn't determined by benchmark performance alone, but how this performance maps on to economic growth, etc. Here, I argue that while intelligence certainly contributes to these domains, there are other complementary factors that may prove equally decisive in the long-run, but are ignored under this misleading framework.

The Problem with the Threshold View

One might defend the 'race' framing by arguing that intelligence is indeed the singularly important competitive dimension. The reasoning here is that beyond a certain threshold of intelligence, all other disadvantages become irrelevant as the AI's intelligence will solve these. This is roughly the AI Futures Project view—once AGI is achieved, ASI is quick to follow, and from there, ASI will use its intelligence to overcome any non-cognitive disadvantages, leading to rapid transformative impacts. Under this view, the U.S.'s lead on compute & algorithms puts it at a significant advantage, as these are key inputs to intelligence. I'll refer to this argument as the "threshold view."

My problem with the threshold view is this: while it’s a convincing response to those skeptical that AI can achieve superhuman performance at all (or within a relatively short timeframe) due to certain bottlenecks, it is less persuasive in a competitive context that is sensitive to speed.

Intuitively: while it's a mistake to bet against a brilliant engineer solving a task requiring him to create a robot arm, it seems unlikely that he'd be faster at this task than a less intelligent peer who happens to already just have a robot arm. While intelligence can undoubtedly be applied to eventually overcome some e.g. material disadvantage, it's not necessarily that this process will be faster than the less intelligent party's immediate leveraging of their e.g. material advantage (particularly if this application of intelligence entails a complex or challenging process—like engineering a functional robot arm).

Let's apply this to one of the aforementioned domains: scientific progress.

On scientific progress, it's argued that AI may be limited by its physical capacity to run experiments and collect data. In response, one might say that AI will use its intelligence to accelerate the experimentation and data collection processes. However, accelerating these processes is, in and of itself, a scientific & engineering challenge requiring data and experimentation. For example, one limitation is that collecting scientific data is highly labor-intensive. If human lab technicians are a limitation, AI could apply its intelligence towards creating robot technicians, allowing it to run more experiments. But this entails solving a notorious engineering challenge requiring not just intelligence, but the time and resources to design, prototype, manufacture, verify, etc.—physical processes limited themselves by the speed of experimentation, data collection, and industrial production. Since realistic computer simulations of complex physical systems remain elusive, the vast majority of science and engineering problems don't reduce to problems that can bypass physical experimentation. To create these simulations, this would require experiments & data to verify that the simulation's predictions are accurate to reality, the production of which again entails the same physical bottlenecks. Ultimately, bringing back the competitive context, it seems dubious that "Step 1: create a hyperrealistic physics simulator. Step 2: use the simulation to invent & manufacture humanoid robot technicians" is faster than "just start with a big surplus of human technicians."

I don't pretend that this is a bulletproof case against the threshold view, but I find it at least more convincing than blanket assertions that intelligence solves anything, faster. If you put some weight on these responses, the following arguments might be quite important.

An Accumulation Game: Gains & Capabilities

Rather than framing AI competition as a race, we should frame it as an accumulation game. First, this would enrich discourse by prompting us to ask questions relevant to our actual priorities.

"AI competition with China is an accumulation game."
"What are we accumulating?"
"Gains from AI"
"How might we accumulate each gain? What's the role of AI?"
"[…]"

Second, this framing acknowledges that national interests lie not in acquiring AI for its own sake, but in accumulating gains from AI—its fruits. We want AI to help us invent the future: cure disease, extend lifespans, bring clean & abundant energy, raise fertility, and more. And we want AI to make us all wealthier: more goods, more leisure, more surplus, less inequality. Technological innovation and a wealthy & healthy population are major assets in any world, and map on to two gains: scientific power and economic power.

Capabilities are what enable us to accumulate gains. Cognitive capabilities (i.e. intelligence) are important, surely, but non-cognitive capabilities also contribute significantly to overall ability. For scientific power, this includes the capability to scale scientific infrastructure, which enables AI to collect data and conduct experiments. For economic power, this includes diffusion capability, which has historically enabled innovation to cash out in growth, as well as political capability, which allows a nation to undergo large-scale, rapid societal transformation without falling to chaos or backlash.

Under this framing, the U.S. is certainly still ahead on algorithmic innovation, and compute especially. But because we lag behind China on these key non-cognitive capabilities, we risk losing out on scientific and economic gains, putting us at a long-term disadvantage in the accumulation game.

Scientific Gains

Capability: Scaling Scientific Infrastructure

When it comes to AI for science, data and experimentation are crucial inputs to actual progress. This does not bode well for the U.S., as China has demonstrated a remarkable capacity for scaling up scientific infrastructure.

Although the United States' recent Gross Domestic Expenditures on R&D (GERD) remains higher in absolute terms ($923B vs. $812B), China's GERD has averaged double-digit growth since 2000 while the U.S. has had sub-10% growth in the same timeframe. In 2023, growth in R&D expenditure in China was over five times greater than in the U.S. (8.7% vs. 1.7%), further reflecting a marked investment in catching up. In fact, adjusting for cost-effectiveness, China may have already caught up—for every American R&D worker supported by $100K of spending, a Chinese lab can employ 2.3 workers due to lower labor costs.

Beyond spending, China has also systematically expanded its State Key Laboratory (SKL) network. These SKLs receive funding, administrative support, and developmental guidance from China's central government, and serve as key drivers of China's innovation agenda. China deliberately modeled this system after the U.S. national laboratory network, but has scaled it far more aggressively: from its inception in 1984, the SKL system has grown to 469 labs by 2019. For comparison, the U.S. operates 17 national laboratories, 42 federally-funded research and development centers (FFRDCs), and 15 university-affiliated research centers (UARCs)—though note that these tend to be more specialized and individually better-funded than SKLs.

On human capital, since the mid-2000s, China has consistently awarded more STEM PhDs than the United States, with 45% of doctorates graduating from Double First Class (A) universities: the country's most elite educational institutions. Overall, China now produces 3.57 million STEM graduates annually—more than 4x as many as the U.S.

This massive investment has translated into demonstrated innovative capacity. In 2022, China overtook the U.S. as top contributor to the 82 leading natural science journals, and while the U.S. still leads on top 1% cited articles, its proportion (1.7%) is at the lowest level since 2006, while China's proportion has risen steadily from 0.4% in 2006 to 1.3% in 2022.

When AI systems require vast amounts of experimental data and iterative testing, having an enormous pool of trained researchers and distributed laboratory infrastructure (and demonstrated capacity for further rapid expansion) may be a significant advantage that algorithmic superiority alone may struggle to overcome quickly.

Economic Gains

Capability: Diffusion

Source

Jeffrey Ding, assistant professor of political science at GWU, argues that analyses of a nation's scientific & technological capabilities often prioritize innovative capacity over diffusion capacity. However, history shows that countries with a diffusion surplus—diffusion capacity that outpaces innovative capacity—are more likely to sustain their rise to power than typical analyses would suggest.

The United States itself exemplifies this principle. During the late 19th century, the U.S. rose to power in the 'Second Industrial Revolution', surpassing the UK in GDP per capita and labor productivity around 1900. From a purely innovative perspective, the United States' steady rise to economic preeminence was rather surprising, as American scientific research lagged far behind that of the UK, France, and Germany throughout the decades of its growth.

But for what it lacked in innovative capacity, the U.S. made up for in diffusion capacity, excelling at adapting breakthroughs for various economically valuable applications. For example, despite trailing far behind Germany in chemical innovation, the U.S. was the first to institutionalize chemical engineering. Industry and academia (most notably MIT) collaborated to establish practical methods such as unit operations, which enabled the widespread diffusion of chemical processing, transforming industries ranging from glass production to petroleum refining.

Today, China appears to be in a similar position. The first industrial robotics patent was filed by an American, but China now accounts for over half of new industrial robotics installations globally, while the U.S. trails at less than 10%. General Motors produced the first modern electric vehicle in 1996, but EVs comprise only 8% of the American market compared to over 50% of new Chinese vehicle sales today. The first experimental and energy-generating nuclear reactors were American-built, but China currently has over ten times as many prospective nuclear facilities planned as the U.S.

Moreover, China has been implementing this same diffusion-focused playbook specifically for AI. The CCP's "AI+ Plan" emphasizes incorporation of AI across society, encouraging "expansive experimentations with industrial and social applications", with ambitious targets for nation-wide AI adoption: 70% adoption of AI-powered terminals, devices, and agents by 2027, scaling to 90% by 2030. There has been extensive coordination between the public sector, private firms, and academia, with many leading generative AI companies emerging directly from university research labs—mirroring precisely America's historical advantage in chemical engineering. Diffusion policies have also tightened integration between deployment and development, demonstrated by projects like the world's first AI hospital, launched by Tsinghua University.

When AI's economic value depends on rapid deployment across industries—as it has historically—having superior diffusion capabilities may prove more decisive than maintaining a temporary lead in innovation.

Capability: Political Response

While AI could cause explosive economic growth through broad labor automation, this would be a massive societal transformation requiring unprecedented coordination and adaptation.

China has demonstrated significant capacity for large-scale labor reallocation. Between 1995 and 2001, China laid off approximately 34 million state-owned enterprise employees—roughly one-third of the total—during economic reform. While this generated significant unrest and was economically devastating for certain regions, the government persisted with the policy. More recently, targeted investments from the "Go West" campaign and Belt and Road Initiative raised urban employment in western provinces by 10-15%. This capacity for coordinated labor reallocation could be crucial if markets fail to redirect displaced workers toward productive sectors. For instance, if firms focus on competing over AI rents rather than developing new major industries, or if investments that complement AI in the long-term (like expanded scientific infrastructure) & could absorb significant labor supply are neglected due to poor short-term commercial prospects.

Relatedly, China's dibao cash transfer program maintains minimum living standards for rural residents and contributed to lifting millions out of poverty, providing a foundation for more extensive support programs that might be necessary due to job displacement. There's also little cultural stigma around such programs, unlike in the U.S. where anything resembling socialism (apart from entrenched programs like Social Security) may face significant resistance.

But most importantly, the PRC's authoritarianism means that it faces fewer constraints from public opinion during jarring economic transitions. Meanwhile, the U.S. will have to worry about backlash, with survey data indicating that barely 30% of Americans trust AI, compared to over 70% of Chinese. And this disparity exists before AI has caused any significant job displacement. If the broad automation necessary for explosive growth occurs, American resistance will likely intensify. It's telling that hundreds of AI regulation bills have already flooded U.S. state legislatures, reflecting Americans' concerns about harms ranging from healthcare errors to discrimination to human extinction.

While these concerns aren't necessarily misguided, the underlying dynamic is clear: Americans seem unlikely to readily accept widespread job displacement from AI when public trust in the technology is already so low—and in a democracy, this popular resistance directly constrains policy in ways that China's system simply does not face.

Conclusion

The threshold view—the view intelligence alone outpaces any non-cognitive advantage—reminds me of Robin Hanson's satirical "betterness explosion", and seems similarly dubious. It seems quite dubious to claim that the advantages China has in scientific, economic, and political infrastructure will be completely negated by a few years or even mere months of algorithmic lag.

While China is still behind on compute, it has also leveraged these precise advantages to catch up rapidly. The gap in computational power between Chinese and non-Chinese chips has shrunk from 10x in 2018 to 3x today, with Chinese chips even briefly holding the lead with the BR100 datacenter GPU, allegedly the fastest chip available at time of release. Its software ecosystem is comparatively quite underdeveloped, but I haven't seen analyses that put China more than a decade from reaching performance parity. Meanwhile, DeepSeek has already demonstrated China's ability to innovate under compute constraints.

So let’s fast forward a decade: China with an independent chip ecosystem, massive scientific buildout, surplus of STEM graduates, and a political system capable of unilaterally imposing order during large-scale labor displacement/reallocation. Contrast this with a U.S. potentially embroiled in chaos—white-collar workers in life-or-death political fights over automation, Congress gridlocked on UBI, scientific funding relegated to the background. Which scenario seems more conducive to accumulating gains from AI?

Yet, our politics have over-indexed on the flawed race framing, tunnel-visioning on benchmarks, compute, and "winning the race to AGI", rather than building the comprehensive capabilities required to accumulate long-term gains.

But perhaps this avoidance is understandable—acquiring these capabilities won't be easy. One suggestion comes from RAND for an Apollo-style program for AI development focused on "whole-of-society" education and targeted industrial policy. While better than a narrow compute focus, it's still a limited suggestion. First, AGI-driven transformation isn't the same as civilian scientists working together to reach the moon—it's full-scale labor automation, likely resembling a few entrepreneurs founding companies with countless dirt-cheap AI employees that displace everyone else. Second, while such programs would certainly help, China enjoys clear structural advantages in industrial policy implementation. The U.S. is already embroiled in political battles over some of the report's suggested policies while China rolls full-steam ahead.

Fundamentally, we have to grapple with the fact that some of our most cherished institutions may disadvantage us in rapid, coordinated societal transformation. We must ask: how can we preserve our values while securing future prosperity?

The race metaphor obscures this nuanced reality. While the U.S. maintains leads in cognitive capabilities, China has built major advantages in the non-cognitive capabilities that may prove decisive for translating AI advances into long-term scientific and economic gains. Recognizing AI competition as an accumulation game rather than a race is the first step toward a more realistic—and potentially more effective—strategic response.