AI Race and National Competition

In the last week of December 2024, a Chinese company most Western analysts had never heard of published a research paper and a set of model weights, and quietly upended a year of American strategic planning. DeepSeek, a startup spun out of a Hangzhou quantitative hedge fund, had trained a model that matched the reasoning performance of OpenAI's best systems. It had done so, by its own account, for a headline training cost of under six million dollars, using a cluster of Nvidia chips that were deliberately hobbled versions — the H800s and older H100s that U.S. export controls had permitted precisely because they were supposed to be too weak to matter.

The reaction in Washington was not measured. When DeepSeek's consumer app hit the top of the U.S. App Store in late January 2025, Nvidia's stock lost roughly $600 billion in a single day — the largest one-day loss in market history to that point. The number itself is less important than what it revealed: a strategic bet, worth years of policy and hundreds of billions in market value, resting on an assumption that turned out to be wrong.

The assumption was elegant. The most advanced AI requires the most advanced chips; the most advanced chips are designed by American firms and manufactured on machines only a handful of companies can build; therefore, control the chips and you control the frontier. Deny China the hardware, and you keep it two or three years behind — long enough to matter. It was the closest thing the United States had to a coherent theory of how to win.

DeepSeek did not disprove that chips matter. It proved something subtler and more unsettling: that when you throttle one input, engineers optimize the others. Told they could not have the fastest chips, Chinese researchers rewrote the software to squeeze more out of slower ones. The constraint became a forcing function. And that single fact reframes the entire competition, because it means the race cannot be won the way the United States hoped to win it — by holding one chokepoint shut.

This chapter is about what the race actually looks like once you abandon the fantasy of a single chokepoint and a single winner. It is a marathon on several tracks at once — model quality, industrial deployment, infrastructure, energy, standards, talent — and the leaders on each track are not the same country.

Where the Money Is Going

Start with the crude measure, because it frames everything else: who is spending what.

By the standard tallies of 2025, the United States committed roughly $471 billion to AI — close to 63 percent of all global AI spending (GM Insights, 2025; Spherical Insights, 2025). This is a private-sector story as much as a government one. The hyperscalers — Microsoft, Google, Amazon, Meta — along with OpenAI and Anthropic, are financing the buildout: data centers, chip orders, the training runs that cost more than some countries' science budgets. Washington's role is comparatively light-touch, mostly export policy and defense procurement, with the capital coming from Seattle, Menlo Park, and San Francisco.

China's reported figure, around $119 billion, is roughly a quarter of the American number in absolute terms (GM Insights, 2025). But the comparison flatters the United States more than it should. Chinese state capital coordinates in ways American capital cannot: the National AI Industry Investment Fund and its provincial cousins direct money into smart manufacturing, autonomous driving, and healthcare AI on timelines that answer to industrial policy rather than quarterly earnings. A dollar of directed state investment and a dollar of venture capital chasing a consumer app are not the same instrument, and they do not buy the same things.

The rest of the field is not spending to win. It is spending not to be locked out.

(Figures compiled from GM Insights, Spherical Insights, and national announcements, 2025. They mix public and private commitments and different accounting years; treat them as orders of magnitude, not audited accounts.)

The European Union's €110 billion InvestAI package, unveiled at the Paris AI Action Summit in February 2025, is the clearest tell. Europe is not trying to out-build OpenAI. It is trying to avoid a future in which every foundational layer of its digital economy runs on American or Chinese infrastructure. The smaller committed powers — the UAE, Singapore, Norway, South Korea — punch above their absolute numbers because they invested early in the unglamorous parts: grid capacity, talent pipelines, government adoption. Spending is necessary. It is nowhere near sufficient, and the gap between what money buys on each track is where the real story lives.

Two Different Games, Scored Differently

Ask "who is winning?" and the honest answer is a question back: winning at what?

The United States owns the frontier. The most capable general-purpose models in the world come from OpenAI, Google DeepMind, and Anthropic, and by mid-2025 U.S. models accounted for something like 93 percent of global large-language-model site visits (TIME, 2025). American research still sets the pace others react to. On the narrow, visible metric of "best chatbot," it is not close.

China is playing a different game and keeping score differently. Its strength is not the consumer chatbot but the embedding of AI into physical systems — the factories, the cameras, the vehicles, the ports. Chinese firms lead in AI-driven surveillance and smart-city infrastructure, in industrial robotics, in the deployment of computer vision into the actual machinery of an economy. This is AI as installed capacity rather than AI as a product you visit in a browser.

And on the frontier track where the United States leads, China is closing fast. Traffic to China-based LLMs jumped by around 460 percent over two months in late 2025 (TIME, 2025). DeepSeek was not a one-off; it was the leading edge of a pattern — Alibaba's Qwen, Zhipu, Moonshot, and others releasing capable models in rapid succession. The gap on model quality, once measured in years, is now measured in months.

Read down that column and the multipolar reality is already visible. No single country holds the top row on every line. The competition is fragmenting across tracks faster than any one nation can dominate them all — which is the central claim of this chapter, and the one worth interrogating hardest before the end.

The Containment Premise, After DeepSeek

Return to the chip controls, because they are the policy most directly wounded by 2024, and the debate over what to do next is the sharpest live argument in the field.

The controls were built on a containment logic borrowed from an earlier era: identify the chokepoint, hold it shut, keep the adversary a generation behind. For AI, the chokepoint was compute — specifically the high-end Nvidia GPUs and the extreme-ultraviolet lithography machines from the Dutch firm ASML that no one else can yet build. Deny those, the theory ran, and you deny the frontier.

DeepSeek exposed the theory's flaw. Compute is one input among several — alongside algorithms, data, and engineering talent — and a determined competitor can substitute across them. Restrict the hardware and you raise the price of a given capability; you do not make it unreachable. Efficiency gains can offset hardware disadvantage, and China has a deep bench of exactly the kind of talent that produces efficiency gains.

This does not mean the controls failed outright, and the more careful analysts resist that conclusion. Denying China the best chips still forces it to spend more compute, more engineering effort, and more time to reach a given point — a tax, not a wall. The controls slow China; they do not stop it. The strategic question is whether a tax is worth its costs, and those costs turned out to be real: the restrictions handed Chinese firms a powerful incentive to build a domestic chip industry and to weaponize open-source, both of which now cut against American interests.

The recalibration argument runs roughly like this. If controls can only slow, not stop, then the goal should shift from denial to durable lead — running faster rather than trying to freeze the other runner. That means investing in the American frontier, the grid, and the talent pipeline at least as aggressively as policing exports. And it means being honest that every turn of the export-control ratchet accelerates Chinese self-sufficiency, which is precisely the outcome the controls were meant to prevent. The counterargument — that even a few years of delay is decisive in a fast-moving field, and that abandoning the tax simply hands Beijing time for free — is not obviously wrong either. That the December 2025 decision to let Nvidia's H200 chips flow to China could be read as either shrewd recalibration or strategic surrender tells you the premise is genuinely contested, not settled.

Lock-In Versus Ubiquity

The two superpowers are not just spending differently; they are pursuing structurally opposite theories of how technological advantage converts into power.

The American theory is lock-in. Let U.S. hardware, cloud platforms, and models become the substrate the world builds on, then convert that dependence into leverage. If a Nigerian fintech, a German carmaker, and an Indonesian ministry all run on American chips, American clouds, and American model APIs, then Washington retains influence over the global digital economy even where it controls no end-use application. The H200 export reversal fits this logic: better that the world's AI runs on Nvidia silicon and depends on American upgrade cycles than that it be pushed to build an independent stack. Lock-in is a bet on exclusivity turned into gravity.

The Chinese theory is ubiquity. By releasing capable models as open weights — DeepSeek, Qwen, and the rest — Chinese firms make their technology the cheap default, especially in markets that cannot afford proprietary American systems. Open weights also route around export controls entirely: if the model is a free download, restricting chip sales does little, because anyone can run it on whatever hardware they can find. Where America bets on being indispensable, China bets on being everywhere. A developer in Lagos or São Paulo who fine-tunes a Chinese open model onto local hardware has, without any diplomatic act, entered a Chinese-shaped technical ecosystem — its defaults, its assumptions, its baked-in values.

The logic behind each choice tracks each country's actual position. The United States leads at the frontier and has proprietary products worth protecting, so it monetizes exclusivity. China trails at the very frontier but has scale, manufacturing depth, and a hardware disadvantage to neutralize, so it commoditizes the layer above the hardware and competes where it is strong. Each strategy is a rational response to a different hand of cards — which is exactly why neither is likely to simply capitulate.

Sovereignty as Insurance

For everyone who is not the United States or China, the strategic problem is not how to win but whom to depend on — and increasingly the answer is "neither, if we can help it."

India launched a sovereign large language model at its AI Impact Summit in February 2026, joining a movement that includes Brazil, Indonesia, the Gulf states, and much of Europe. None of these countries expects to out-innovate Silicon Valley or Shenzhen. That is not the point. The point is that dependence on foreign AI infrastructure creates specific, concrete vulnerabilities, and once you name them the investment logic becomes obvious.

Consider what it means to run your economy's cognitive layer on someone else's system. The foreign provider can be ordered by its home government to cut you off — sanctions, export rules, a diplomatic quarrel — and your hospitals, banks, and ministries lose a utility they have come to depend on. The model encodes the provider's language, assumptions, and content rules, so a nation's own tongue, history, and legal norms are underserved or quietly distorted. Your citizens' data flows through infrastructure you do not control and cannot audit. And you are a price-taker, exposed to whatever the provider decides to charge once you are locked in. Sovereign AI is insurance against all of that. It is expensive, often technically inferior to the frontier, and for many states entirely rational anyway — the premium you pay not to have your nervous system hostage to a foreign capital.

This is where the leading powers acquire an obligation, and where the book's distribution question sharpens to a point. Dozens of countries lack the capital, the grid, the chips, and the talent to build even a modest sovereign capability. Their realistic choice is not sovereignty versus dependence but which patron to depend on. If nothing changes, they are simply absorbed into whichever superpower's stack they can afford — a digital-era version of an older subordination, with model weights and cloud contracts in place of infrastructure loans. Whether governance mechanisms — shared public compute, open multilingual models as genuine public goods, pooled regional capacity of the kind the African Union and ASEAN have floated — can give these states real alternatives is one of the central unresolved questions of the decade. Nothing about the current trajectory guarantees a good answer.

The Input Nobody Priced In

For all the attention on chips and models, the constraint most likely to shape the late 2020s is the least glamorous: electricity.

A frontier training run, and then the far larger ongoing job of serving a model to millions of users, consumes power on the scale of a small city. As models grow and deployment widens, the binding constraint shifts from "can we get the chips?" to "can we power them?" — and that turns electricity generation from an infrastructure footnote into a strategic variable.

Here the asymmetry runs the other way. China has generated more electricity than the United States since 2010, and it is adding capacity faster — coal, hydro, nuclear, and a staggering buildout of solar and wind, all at once (Brookings, 2025). The United States, by contrast, is running into its own grid. Data-center demand is already straining networks from Northern Virginia to Phoenix; new generation takes years to permit and build; and the clean capacity the hyperscalers have promised is the slowest of all to come online. On chips the United States holds a design lead, and on talent its universities still draw the world's best. On energy it faces a genuine structural disadvantage, and unlike the others it cannot be closed by a clever algorithm or a better recruiter.

The mechanism by which this becomes decisive is straightforward. Today, restricted chip access caps how much compute China can bring to bear, so its energy advantage is partly latent. But if that cap loosens — through relaxed export policy like the H200 reversal, through a maturing domestic chip industry, or through continued efficiency gains — then abundant, comparatively cheap Chinese power stops being latent and starts being decisive. Chips are a stock you can restrict at the border. Power plants are a flow you have to build at home, over years. If the race runs deep into the 2030s, as it almost certainly will, energy may matter more than any other single input.

How Long Until China Builds Its Own Chips

Which raises the question the energy story hinges on: how fast can China escape its dependence on foreign silicon?

The honest answer is that it is closing the trailing-edge gap quickly and the leading-edge gap slowly. SMIC, China's national champion foundry, has demonstrated 7-nanometer production and is pushing toward 5-nanometer — impressive, but achieved without access to ASML's extreme-ultraviolet lithography machines, which means lower yields, higher costs, and a hard ceiling on how far the older technique can be stretched. Matching the leading edge, where Taiwan's TSMC operates, requires either indigenous EUV — a machine of such complexity that ASML's monopoly has held for two decades — or a fundamentally different manufacturing path. Most credible assessments put genuine leading-edge self-sufficiency somewhere in the late 2020s to mid-2030s, with wide error bars and a real chance the timeline slips.

But — and DeepSeek is the whole lesson here — leading-edge parity may not be necessary for competitiveness. If algorithmic efficiency keeps compensating for hardware, then "good enough" domestic chips produced at scale could carry China further than the node-size gap suggests. The question is not only when China matches TSMC, but whether it needs to. If it does not, the containment premise weakens further still.

What a Multipolar Order Actually Looks Like

Put the tracks together and the shape of the thing emerges. The likeliest future is not one superpower astride the field but a value chain sliced horizontally, with different actors owning different layers.

graph TD
    A["Lithography & fab equipment<br/>ASML (NL), Applied Materials (US), Japan"] --> B["Chip fabrication<br/>TSMC (Taiwan), Samsung, SMIC (China)"]
    B --> C["Chip design<br/>Nvidia, AMD (US); Huawei, Cambricon (China)"]
    C --> D["Cloud & compute<br/>US hyperscalers; Alibaba, Huawei Cloud"]
    D --> E["Foundation models<br/>US frontier; Chinese open-weight"]
    E --> F["Applications & deployment<br/>global, contested layer by layer"]

Whoever controls a layer holds leverage over everyone above it. The most concentrated chokepoint sits at the very bottom — ASML's lithography monopoly and TSMC's fabrication lead give a tiny cluster of firms, mostly in Europe and Taiwan, outsized power over the entire stack. This is why Taiwan is not merely a flashpoint but arguably the single most strategically important place in the AI economy: a disruption there would ripple up through every layer above it.

The geopolitical implication of this fragmentation is that AI power becomes relational rather than absolute. The United States can dominate models and still depend on a Dutch machine and a Taiwanese fab. China can flood the world with open weights and still be throttled by chips it cannot yet make. Because the stack is built on open research, shared datasets, and global talent flows, even adversaries feed each other: Chinese researchers publish papers that improve American models; American open frameworks power Chinese applications; Israeli defense research diffuses into commercial systems worldwide. The boundaries are porous by design, and that porousness is itself an argument against any clean, unipolar outcome.

The values dimension is where fragmentation gets politically heavy. The United States and China are not building the same technology with different logos; they are building ecosystems with different defaults about speech, surveillance, data, and state control, baked in at the level of what the models will and won't do. As countries adopt one stack or the other, they inherit its assumptions. A government that runs on Chinese infrastructure finds surveillance-friendly tooling ready to hand; one that runs on the American and European stack inherits a different, though not value-free, set of norms and a thickening web of regulation. Two spheres are forming — not by explicit alliance, mostly, but by the accumulated gravity of a thousand procurement decisions — and which sphere a country lands in is being decided now, quietly, one contract at a time.

Why 2026 Is a Hinge

Analysts keep calling 2026 a hinge year, and the phrase survives scrutiny because several decisions with decade-long consequences are all falling due at once.

Export policy is at a genuine fork: the H200 reversal could widen into general permissiveness or snap back toward restriction, and each path sends the competition down a different decade. China's semiconductor push is at the point where the trajectory of the next ten years becomes legible — whether SMIC's progress compounds or stalls against the EUV wall. The open-source strategy is reaching the scale where network effects either lock in Chinese models as a global default or don't. And the sovereignty movement is deciding whether it produces real independent capacity or a scatter of underpowered national projects that quietly re-consolidate around the superpowers anyway.

None of these is settled. All of them are being decided right now — in boardrooms, in ministries, in classified briefings like the one that opened this chapter. What makes 2026 pivotal is not that anything is resolved but that the options are still genuinely open, and the choices made while they remain open will foreclose most of the alternatives.

How Much of This Do We Actually Know?

A chapter this confident owes the reader an accounting of its own uncertainty — and this is one domain where the analysts have been humbled recently enough to make humility mandatory.

Start with China's opacity. Much of what matters — real training costs, military applications, the true state of the domestic chip industry — is either classified or wrapped in strategic messaging, and DeepSeek's own headline cost figure is contested precisely because outsiders cannot audit it. The people whose job is to assess Chinese AI capability were blindsided by DeepSeek. That should lower our confidence not just in specific numbers but in the whole enterprise of external assessment. If the analysts were surprised once by a public consumer model, the classified military applications — the domain where surprise matters most — are exactly where our picture is weakest.

The strategic projections are also acutely sensitive to a handful of uncertain variables, and small changes in assumptions swing the conclusions hard. Whether SMIC cracks sub-5-nanometer production; whether the American grid can be scaled fast enough; whether regulatory divergence between the EU, US, and China hardens into incompatible blocs or converges; whether talent keeps flowing to American universities or reverses as visa politics and Chinese domestic opportunity shift — each of these is a genuine unknown, and the plausible answers point at meaningfully different worlds.

Which brings the multipolar thesis itself into the dock. Is fragmentation a confident prediction or one scenario among several? Honestly, the latter — a well-supported base case, not a certainty. It is where the current distribution of tracks points, but specific conditions would break it either way. Meaningful U.S. dominance becomes plausible if a genuine capability discontinuity opens up — a frontier model so far ahead that lead compounds into runaway advantage — while export controls simultaneously bite harder than DeepSeek suggests and the grid problem gets solved. Meaningful Chinese dominance becomes plausible if the chip gap closes faster than expected, the energy advantage converts as the mechanism above predicts, and open-weight ubiquity tips into genuine global lock-in while American infrastructure stalls. The multipolar outcome is the most likely because it requires none of these things to break decisively in one direction — but "most likely" is not "certain," and anyone who tells you they know how this ends is selling something.

Summary

The AI race is not a sprint to one finish line but a marathon on several tracks — model quality, industrial deployment, infrastructure, energy, standards, and talent — with different leaders on each. The United States and China dominate, but asymmetrically: America leads frontier models and holds roughly 93 percent of global LLM traffic, while China leads industrial integration, surveillance infrastructure, and — critically — energy. Investment is dominated by the United States (~$471 billion in 2025) and China (~$119 billion), but the more revealing spending comes from the EU, UK, India, Israel, and the Gulf, most of it aimed not at winning but at avoiding total dependence.

DeepSeek's late-2024 breakthrough is the pivot the chapter turns on. By matching frontier performance on restricted chips at a fraction of the cost, it showed that export controls tax but do not contain — algorithmic efficiency can substitute for hardware. That fact forces a strategic recalibration from denial toward maintaining a durable lead, and it explains the two superpowers' opposite bets: American lock-in through platform exports versus Chinese ubiquity through open weights. Middle powers, unable to win either way, buy sovereignty as insurance against the specific vulnerabilities of running their economies on someone else's infrastructure — leaving open the sharp question of what the leaders owe the states that cannot afford even that.

Energy is the underpriced constraint: China has out-generated the United States since 2010 and builds faster, an advantage that stays latent only as long as chip restrictions cap Chinese compute, and turns decisive the moment they loosen. Domestic Chinese chips may reach leading-edge parity only in the late 2020s to mid-2030s — but DeepSeek suggests parity may not be required. The most likely result is a multipolar order in which power is relational: control is sliced by layer, the deepest chokepoint sits with ASML and TSMC, and two value-laden ecosystems pull nations into competing spheres one procurement decision at a time. That multipolar outcome is a well-founded base case, not a prophecy — and given how badly the analysts were surprised once, honest confidence in any single ending remains low.

Key Takeaways

1. The race is multidimensional, with no single winner in sight. The U.S. leads frontier models and market share (~93% of LLM traffic, 2025); China leads industrial integration, surveillance, and energy. Leadership on one track does not transfer to the others.

2. DeepSeek broke the containment premise. Matching frontier models on restricted chips at a fraction of the cost showed export controls impose a tax, not a wall — algorithmic efficiency substitutes for hardware. The strategic logic must shift from denying compute to sustaining a durable lead.

3. The superpowers bet oppositely. The U.S. pursues lock-in through hardware and platform exports; China pursues ubiquity through open-weight models that both undercut proprietary systems and route around chip controls. Each is a rational response to a different competitive position.

4. Middle powers buy sovereignty as insurance. India, Brazil, Indonesia, the EU, and the Gulf invest not to out-innovate the giants but to avoid the concrete vulnerabilities of foreign dependence — a shutoff risk, encoded foreign values, unauditable data flows, and price-taking.

5. Energy is the underappreciated constraint. China has out-generated the U.S. since 2010 and expands faster; that advantage becomes decisive if chip restrictions loosen, because power plants must be built at home over years while chips can be restricted at a border.

6. A multipolar order is the base case, not a certainty. Power is sliced by value-chain layer, with the tightest chokepoint at ASML and TSMC. U.S. or Chinese dominance remains possible under specific conditions — and given how badly DeepSeek surprised the analysts, confidence in any single ending should stay low.

Sources

• US-China AI Race: 2026 Strategies and Shifts
• How will the United States and China power the AI race? | Brookings
• The Myth of the AI Race: Neither America Nor China Can Achieve True Tech Dominance | Foreign Affairs
• Eight ways AI will shape geopolitics in 2026 | Atlantic Council
• U.S.-China Competition for AI Markets | RAND
• 6 Graphs That Show Who's Really Winning the US–China AI Race | TIME
• The Global AI Race: How Countries Are Competing in 2025 | GM Insights
• Top 10 AI Spending Countries 2025: Statistics and Trends | Spherical Insights
• How 2026 Could Decide the Future of Artificial Intelligence | Council on Foreign Relations
• The Role of Advanced Technology: Reconfiguring the Post-2026 Geopolitical Order | TRENDS Research

Last updated: 2026-08-08