Why the AI Race Has No Winner

Washington keeps talking about artificial intelligence the way it once talked about the moon landing: a single track, a single finish line, a single flag planted at the end.

That’s the frame behind the Trump administration’s AI action plan, released in July 2025 under the headline “Winning the AI Race.” (The White House)

But the “race” metaphor is doing real damage. It suggests there’s one kind of AI leadership, one scoreboard, and one champion—either America or China. In reality, the competition is splintering into multiple arenas that don’t automatically convert into each other. You can win one event and still lose the decathlon.

The more realistic end state isn’t “U.S. dominance” or “China dominance.” It’s something messier—and more stable than people admit:

asymmetric AI bipolarity, where each side leads in different layers of the ecosystem.

And that matters, because the strategy for a long, uneven bipolar era is totally different from the strategy for a sprint.

Contents

The race is actually four races

If you want to understand where this is going, stop asking “Who wins AI?” and start asking: wins what part of AI?

At minimum, the competition is breaking into four tracks:

Frontier models
The most advanced large language and multimodal systems—who builds the smartest models, who pushes reasoning and tool-use forward, who gets the breakthrough architectures first.
Compute and infrastructure
Chips, data centers, energy, supply chains, and the ability to scale training and inference at massive volumes.
Diffusion and standards
Whose software stacks, frameworks, APIs, and model formats become the default across the world—especially in emerging markets that don’t want to be trapped inside one superpower’s cloud.
Embodied AI
The part that doesn’t live in chat windows: robots, factories, vehicles, drones, and military platforms. This is where “AI” turns into industrial productivity and physical capability.

Here’s the key: being #1 in one track doesn’t guarantee you win the others. The future can easily look like this:

the U.S. stays ahead on frontier models and premium AI services
China becomes the dominant supplier of “good enough” AI—cheap, deployable, adaptable—embedded into machines and infrastructure worldwide

That’s not a clean win for either side. It’s a split world.

America’s real edge isn’t just brains—it’s horsepower

Right now, the United States still has the clearest advantage at the frontier: the most capable models and the most commercially valuable AI services tend to cluster around U.S. firms.

But the deeper moat is compute—the ability to train huge models and serve billions of requests without collapsing under cost.

One way to see this gap: by AI supercomputer performance, the U.S. hosts roughly three-quarters of global GPU-cluster capacity, with China a distant second. (Epoch AI)

That compute dominance doesn’t just produce better models. It produces something more important:

faster iteration cycles
more experimentation
more “failures” tolerated at scale
and enough inference capacity to turn raw model strength into real products people use every day

Which is why export controls became the cornerstone of U.S. AI strategy: slow China’s access to top-tier chips and the tools needed to manufacture them, and you slow the whole machine.

Then came the H200 decision—and the game got stranger

In December 2025, Trump announced a major reversal: the U.S. would allow exports of Nvidia’s H200 (its second most powerful AI chip line) to approved customers in China, reportedly with a government fee on the sales. (Reuters)

The logic is seductive:

“Let them buy good-enough chips.”
“Our companies make money.”
“Our software ecosystem stays the standard.”
“We keep the true cutting edge.”

But there’s a danger baked into this logic that the “race” framing hides.

If you sell high-end compute to your main competitor, you might not be helping them beat your best models tomorrow—but you could be helping them build an entirely different kind of advantage: scale, diffusion, and embodied deployment.

And once China has “enough” compute, the competition becomes less about who has the single smartest model and more about who can roll out AI across:

factories
logistics networks
public services
ports
energy grids
transportation systems
and military platforms

That’s where “good enough” beats “best in the lab.”

The world doesn’t buy “the best.” It buys “the usable.”

Here’s the part American policymakers often forget: most countries are not shopping for the world’s most advanced model. They’re shopping for AI that is:

affordable
customizable
deployable on local infrastructure
compliant with local data rules
and not locked behind a foreign cloud they don’t control

This is where China has a structural play.

U.S. labs tend to dominate with closed, proprietary models offered as services. They’re powerful and convenient—but tightly controlled and harder to tailor deeply.

Chinese firms, by contrast, have leaned hard into open-weight models and flexible deployments: cheaper, modifiable, easier to run on local or regional clouds, and easier to adapt to specific languages, sectors, and constraints.

If Chinese firms can pair that software posture with abundant compute—whether built domestically or deployed abroad—the result is a very particular kind of influence:

Not “China makes the best AI,” but “China becomes the default AI supplier.”

And if those deployments run on advanced U.S.-designed chips (thanks to relaxed export rules), you get an even weirder outcome:

a U.S.-enabled Chinese AI stack spreading across the Global South.

That’s not dominance. That’s a split ecosystem.

The embodied AI angle is where the race metaphor really breaks

People love measuring model benchmarks. But the largest economic impact of AI may come less from smarter text generators and more from putting intelligence into physical systems.

Embodied AI is where bits become atoms.

And China has a serious advantage here: it can pilot AI in the real economy at breathtaking speed, because it has the manufacturing depth, the industrial clustering, and the policy willingness to subsidize deployment at scale.

Look at industrial robotics as a crude proxy for this momentum. According to the International Federation of Robotics, China’s operational stock of industrial robots surpassed two million in 2024. (IFR International Federation of Robotics)

That matters because embodied AI is not just software—it’s:

sensors
motors
production lines
supply chains
maintenance networks
procurement incentives
and a million “boring” deployment decisions

This is exactly where a country can be “behind” on frontier models and still pull ahead in industrial productivity and physical capability.

In other words: the AI race is not a sprint. It’s a grueling decathlon.

What “AI bipolarity” actually means

Put the pieces together and the most plausible future looks like this:

The U.S. leads on frontier models and high-margin AI services delivered through major cloud platforms.
China leads in wide deployment of “good enough” AI, especially open-weight ecosystems, infrastructure bundles, and embodied systems tied into manufacturing.

That’s asymmetric bipolarity: not equal, not symmetrical, but persistent and hard to “finish.”

And in that world, Washington has to stop acting like one dramatic breakthrough will settle the contest.

It won’t.

The strategy shift America needs

If the competition is long and multidimensional, U.S. strategy has to change in four ways:

1) Treat compute like the bedrock it is

If Washington wants to preserve a real moat, it can’t casually bleed away compute advantage through broad high-end exports. If the H200 policy stays, then licenses should move slowly, with aggressive scrutiny of end users and routing—especially where military or dual-use ties are obvious.

2) Compete on diffusion, not just invention

If China is building the default stack for emerging markets, the U.S. needs a serious, funded diffusion strategy—financing, infrastructure partnerships, and credible alternatives that don’t force countries into permanent dependence.

3) Prepare for an uneven domestic shock

If the U.S. dominates AI services, it may also experience sharper disruption in white-collar work—especially entry-level jobs. Meanwhile, if China dominates industrial AI, U.S. manufacturing competitiveness could keep eroding. The right answer is not nostalgia. It’s upgrading training pipelines, mid-career retraining, and building AI-enabled production capacity at home.

4) Compete hard—while talking constantly

Bipolar AI doesn’t remove shared risk; it multiplies it. Neither side can fully insulate itself from the other’s failures, misuse, or escalation spirals. That makes ongoing dialogue not a gift to Beijing, but a form of self-preservation—especially around catastrophic cyber misuse, biosecurity risks, and loss-of-control dynamics.

The point

The “AI race” story is emotionally satisfying because it’s simple: pick a lane, run faster, win.

But the real world is not waiting at a finish line.

It’s building an AI ecosystem—chips, clouds, standards, factories, robots, and influence networks—that can split into two partially compatible worlds.

In that world, the question isn’t “How do we win the AI race?”

It’s:

How do we stay strong across multiple arenas, keep allies close, prevent an ecosystem split from turning into a security spiral—and avoid sleepwalking into a race to the bottom?