THE AI ARMS RACE: WHY AMERICA AND CHINA MAY NO LONGER BE ABLE TO SLOW DOWN

Something important happened in Washington this week, and it deserves considerably more attention than another debate about artificial intelligence regulation.

Donald Trump rejected calls from technology executives and political opponents for America to slow the development of advanced AI.

His argument was essentially geopolitical: if America slows down, China may not.

The Financial Times captured the dilemma on its front page: Trump rejected industry calls for restraint amid fears that slowing American AI development could allow China to win the technological race. The debate follows increasingly explicit warnings from figures within the AI industry about the risks associated with increasingly capable systems.

At first sight, this looks like another familiar argument between technological optimism and technological regulation.

It is not.

Something much more consequential is happening.

Artificial intelligence is moving from the realm of technology policy into the realm of national security doctrine.

And once that transition occurs, the political economics of AI change completely.

The central question is no longer:

How should governments regulate artificial intelligence?

It becomes:

Can any major power afford to regulate AI aggressively if it believes its geopolitical rival will continue accelerating?

That is a fundamentally different problem.

And it may be one of the defining strategic dilemmas of the next decade.

1. THE NUMBERS ALREADY LOOK LIKE AN ARMS RACE

The scale of American investment is extraordinary.

According to Stanford University’s 2026 AI Index, private AI investment in the United States reached approximately $285.9 billion in 2025.

China recorded approximately $12.4 billion in measured private investment.

That makes American private investment more than 23 times larger.

But those numbers require an important qualification.

China’s economic system finances strategic industries differently.

State investment vehicles, government guidance funds, state-owned enterprises and government-supported infrastructure make simple comparisons of private capital misleading.

Stanford estimates that Chinese government guidance funds deployed roughly $184 billion into AI companies between 2000 and 2023.

So America possesses an enormous private-capital advantage.

China possesses something different:

the ability to mobilise industrial policy, infrastructure and state capital around strategic objectives.

These are two very different economic machines competing for the same technological prize.

2. AMERICA LEADS. CHINA IS CLOSING THE TECHNOLOGICAL GAP.

Investment alone does not determine technological leadership.

In 2024, American institutions produced 40 notable AI models, compared with 15 from China and only three from Europe.

But something more interesting was happening underneath those numbers.

Stanford found that the performance gap between leading American and Chinese models on important benchmarks had narrowed dramatically, moving from double-digit differences in 2023 toward near parity on several measures in 2024.

China also remained ahead in AI publications and patents.

That distinction matters enormously.

America dominates much of the frontier AI ecosystem.

But China does not need to surpass America everywhere.

It merely needs to remain sufficiently close that Washington cannot confidently assume technological supremacy.

That creates the strategic trap.

3. THE PRISONER’S DILEMMA OF ARTIFICIAL INTELLIGENCE

Imagine that both Washington and Beijing privately concluded that extremely powerful AI systems posed serious risks.

Both governments might theoretically prefer slower development.

But neither government can confidently choose that strategy.

Washington thinks:

“If we slow down and China does not, China wins.”

Beijing thinks:

“If we slow down and America does not, America wins.”

Therefore both continue accelerating.

This is essentially a geopolitical version of the prisoner’s dilemma.

The individually rational decision produces a collectively dangerous outcome.

Neither side necessarily needs to want an arms race.

They merely need to fear losing one.

That distinction is crucial.

4. THE REAL AI RACE IS NOT ABOUT CHATBOTS

Public discussion still focuses excessively on consumer applications.

Chatbots.

Image generators.

Search engines.

Office assistants.

Those are visible manifestations of AI.

They are not the strategic core of the competition.

The real race increasingly concerns:

semiconductors, compute, electricity, data centres, cloud infrastructure, scientific research, robotics, autonomous systems, cyber capabilities, military intelligence, surveillance, industrial automation and talent.

AI is becoming a general-purpose strategic infrastructure.

That makes the comparison with previous technological races increasingly relevant.

Not identical.

But relevant.

5. CHIPS ARE THE NEW STRATEGIC COMMODITY

For decades, geopolitical power depended heavily on control over physical commodities.

Oil.

Gas.

Steel.

Uranium.

Industrial capacity.

AI introduces another strategic resource:

advanced computation.

And computation ultimately depends on semiconductors.

That explains why restrictions on advanced chips and semiconductor manufacturing technology have become such an important component of US-China relations.

America and its allies possess enormous leverage across critical portions of the semiconductor supply chain.

China therefore has a powerful incentive to develop domestic alternatives.

This creates another paradox.

Restrictions can slow Chinese technological development.

But restrictions simultaneously increase China’s incentive to become technologically independent.

The strategic objective therefore becomes difficult:

slow your competitor without teaching your competitor to stop depending on you.

History suggests this is rarely simple.

6. ELECTRICITY MAY BECOME THE MOST UNDERRATED AI ASSET

There is another constraint receiving far less public attention.

Electricity.

AI requires enormous computing infrastructure.

Computing infrastructure requires enormous amounts of power.

The International Energy Agency estimates global electricity consumption from data centres could more than double to approximately 945 TWh by 2030.

That is slightly more electricity than Japan consumes today.

AI is expected to be the most important driver of that increase.

The geographic concentration is even more striking.

The United States and China together are expected to account for nearly 80% of global growth in data-centre electricity consumption through 2030.

US consumption could rise approximately 240 TWh, or 130%.

China could add approximately 175 TWh, an increase of around 170%.

This means the AI race may eventually become partially an energy race.

The countries capable of producing abundant, reliable and relatively inexpensive electricity will possess a structural advantage.

Suddenly nuclear energy matters.

Natural gas matters.

Grid infrastructure matters.

Renewables matter.

Transmission lines matter.

Permitting matters.

Even transformers matter.

The digital economy turns out to be surprisingly physical.

7. AMERICA’S ENORMOUS ADVANTAGE: CAPITAL

The United States possesses something extraordinarily difficult to replicate.

The deepest technology capital markets in the world.

Venture capital.

Private equity.

Public equity markets.

Institutional investors.

Hyperscalers capable of spending tens of billions of dollars.

A startup can move from an idea to billions in financing extraordinarily quickly.

The Stanford numbers illustrate the magnitude of that ecosystem.

In 2025, the United States had 1,953 newly funded AI companies — more than ten times the number in the next closest country.

That is not simply money.

It is an entire financial architecture capable of financing technological experimentation at extraordinary scale.

China cannot easily reproduce Silicon Valley.

But America cannot easily reproduce China’s state-directed industrial system either.

8. CHINA’S ADVANTAGE: SCALE AND IMPLEMENTATION

China possesses different advantages.

A huge domestic market.

Large manufacturing ecosystems.

Extensive digital infrastructure.

An enormous engineering workforce.

Government capacity to prioritise strategic industries.

And critically, the ability to integrate AI with manufacturing.

This could prove important because the next stage of AI may not be won exclusively by whoever creates the smartest model.

It may also be won by whoever deploys AI fastest across the physical economy.

Factories.

Warehouses.

Ports.

Vehicles.

Robots.

Power networks.

Healthcare.

Logistics.

Military systems.

AI leadership therefore has at least two dimensions:

intelligence creation and intelligence deployment.

America currently appears stronger in the first.

China could become exceptionally competitive in the second.

9. THEN CAME DEEPSEEK

DeepSeek changed the psychological dimension of the competition.

The importance of DeepSeek was not simply whether one particular Chinese model was superior to an American model.

The deeper message was more uncomfortable:

China demonstrated that restrictions on access to the world’s best hardware did not necessarily eliminate its ability to produce highly competitive AI systems.

That altered perceptions.

It suggested that algorithmic efficiency could partially compensate for hardware constraints.

And it reinforced the central strategic fear in Washington:

China might remain competitive even while operating under technological restrictions.

The race therefore became harder to manage.

10. AI HAS ALREADY ENTERED NATIONAL SECURITY DOCTRINE

This is not speculation anymore.

A June 2026 White House national-security memorandum explicitly instructed the US national-security apparatus to accelerate AI adoption, remove unnecessary barriers and make advanced frontier models available for national-security applications without unnecessary delay.

The stated objective includes maintaining technological advantage.

That language matters.

Once AI becomes connected to national-security “overmatch”, slowing technological development stops being merely an economic decision.

It becomes strategically suspect.

This is precisely why Trump’s position cannot simply be dismissed as deregulation.

There is a deeper strategic logic behind it.

Whether one agrees with that logic is another question.

11. THE AI COMPANIES HAVE THEIR OWN DILEMMA

Technology companies face an equally uncomfortable problem.

Suppose one American frontier laboratory voluntarily slows development because it considers a capability dangerous.

What happens if its competitors continue?

It loses market share.

Talent.

Capital.

Compute.

Customers.

And potentially strategic relevance.

The same logic operates internationally.

Voluntary restraint works only when competitors believe everyone else will exercise comparable restraint.

Without coordination, restraint can become unilateral disarmament.

This is why calls for “responsible AI” sound simple in corporate presentations but become extraordinarily complicated in competitive markets.

12. SAFETY IS NOT AN ABSTRACT CONCERN

None of this means AI risks should be dismissed.

Quite the opposite.

Stanford’s 2026 AI Index recorded 362 documented AI incidents, compared with 233 in 2024.

The concerns extend far beyond hypothetical superintelligence.

Cybersecurity.

Disinformation.

Fraud.

Autonomous weapons.

Biological research.

Mass surveillance.

Employment displacement.

Market concentration.

Privacy.

Algorithmic discrimination.

Political manipulation.

Critical infrastructure.

These are already policy problems.

More advanced systems could amplify them dramatically.

The tragedy is that the strategic competition making AI more powerful may simultaneously make regulating it more difficult.

13. THE ECONOMIC INCENTIVE TO ACCELERATE IS EQUALLY POWERFUL

There is another reason governments will hesitate to slow AI.

Productivity.

Economic growth.

Stanford reports that AI adoption among surveyed organisations has expanded rapidly, while companies increasingly report measurable cost savings from its use.

In 2024, 78% of organisations surveyed reported using AI, compared with 55% one year earlier.

Generative AI use in at least one business function rose from 33% to 71%.

Governments therefore confront a difficult political equation.

Regulating AI aggressively could potentially reduce risk.

But it could also reduce investment, productivity growth and competitiveness.

Politicians rarely volunteer for slower economic growth.

Especially when a strategic competitor may choose differently.

14. THE LABOUR MARKET COULD BECOME THE POLITICAL BRAKE

Ironically, the strongest constraint on AI may eventually come not from safety researchers but from voters.

If AI materially displaces white-collar employment, the politics of AI could change extremely quickly.

Accountants.

Programmers.

Analysts.

Customer-service employees.

Designers.

Lawyers.

Consultants.

Back-office workers.

Financial professionals.

Administrative staff.

These are not marginal labour-market groups.

They constitute politically influential parts of the middle class.

AI therefore contains a potentially explosive contradiction.

Governments may consider rapid AI development strategically necessary while their voters increasingly consider it economically threatening.

The AI race could eventually become a conflict between national competitiveness and domestic political stability.

15. THE DISTRIBUTION QUESTION MAY BECOME BIGGER THAN THE TECHNOLOGY QUESTION

Suppose AI creates enormous productivity growth.

Who receives it?

Workers?

Consumers?

Governments?

Shareholders?

Technology companies?

Owners of data centres?

Semiconductor manufacturers?

Energy producers?

If AI dramatically increases productivity while concentrating ownership of the productive infrastructure in a relatively small number of companies, inequality could rise even while GDP increases.

That would create one of the strangest economic outcomes imaginable:

a richer economy populated by increasingly economically insecure citizens.

Joseph Stiglitz recently warned about precisely this broader distributional problem, arguing that without competition policy, regulation and mechanisms to distribute AI productivity gains more widely, technological abundance could coexist with growing inequality and weakened democratic trust.

That may ultimately become more politically important than whether the next model scores five percentage points higher on a benchmark.

16. EUROPE’S UNCOMFORTABLE POSITION

And then there is Europe.

Europe has considerable scientific talent.

Excellent universities.

Advanced industrial companies.

ASML.

Strong semiconductor research.

Large markets.

Sophisticated regulatory institutions.

But measured private AI investment remains dramatically below America’s.

The comparison is uncomfortable.

In 2024, American private AI investment was approximately $109.1 billion.

The UK recorded around $4.5 billion.

France $2.6 billion.

Germany approximately $2 billion.

Italy less than $1 billion.

Europe risks becoming extremely good at regulating technologies developed somewhere else.

That is not an argument against regulation.

It is an argument that regulation without technological scale is not technological sovereignty.

17. THIS IS NOT THE COLD WAR

The comparison with the nuclear arms race is useful but imperfect.

Nuclear weapons were primarily government projects.

Artificial intelligence is being developed largely by private corporations.

That creates a completely different governance structure.

OpenAI.

Anthropic.

Google.

Meta.

Microsoft.

Amazon.

xAI.

And their Chinese counterparts.

Private companies now control infrastructure with potentially enormous national-security implications.

Governments regulate them.

Governments buy from them.

Governments depend on them.

Governments compete through them.

That relationship between corporations and states could become one of the defining political-economic structures of the 21st century.

18. THE REAL BOTTLENECKS

The popular narrative says the AI race is about models.

The real competition may increasingly revolve around six scarce resources:

Compute.

Advanced semiconductors.

Electricity.

Capital.

Data.

Human talent.

Control enough of those resources and technological leadership becomes possible.

Lose access to several simultaneously and even brilliant research becomes difficult to scale.

This means AI geopolitics will increasingly overlap with energy policy, immigration policy, financial markets, industrial policy and international trade.

AI policy will cease to exist as a separate policy category.

It will become embedded everywhere.

19. THE TALENT WARNING FOR AMERICA

America dominates AI capital.

But there is an interesting warning sign.

Stanford reports that the number of AI researchers and developers moving to the United States has fallen 89% since 2017, including an 80% decline in the latest year measured.

That should concern Washington.

Technological leadership historically depends not simply on domestic talent but on the ability to attract the world’s best scientists, engineers and entrepreneurs.

America’s immigration system has therefore quietly become part of AI strategy.

If Washington wants technological supremacy while simultaneously making it harder for global technical talent to work in America, those objectives may eventually collide.

20. THE MOST DANGEROUS WORD IN THE AI DEBATE MAY BE “FIRST”

The race mentality creates one particularly dangerous incentive.

Being first.

First to AGI.

First to superintelligence.

First to autonomous scientific discovery.

First to fully autonomous weapons.

First to AI-controlled industrial systems.

First to whatever comes next.

But speed and safety are not always compatible.

Engineering normally relies on testing.

Redundancy.

Failure analysis.

Verification.

AI competition rewards iteration speed.

That creates an uncomfortable asymmetry.

The geopolitical value of being first is immediate.

The cost of insufficient safety may only become visible later.

Political systems are notoriously bad at pricing delayed risks.

21. WHAT WOULD ACTUAL AI ARMS CONTROL LOOK LIKE?

Eventually the world may need something analogous — though certainly not identical — to arms-control architecture.

Possible areas of cooperation could include:

mandatory reporting of extremely large training runs;

shared standards for catastrophic-risk testing;

restrictions on fully autonomous nuclear command systems;

international protocols concerning AI-enabled biological threats;

cybersecurity standards for frontier models;

incident-reporting mechanisms;

verification systems around certain military applications;

direct US-China communication channels concerning AI escalation.

The objective would not be to stop AI.

That is probably impossible.

The objective would be narrower:

prevent competition from creating incentives for catastrophic recklessness.

22. BUT VERIFICATION IS MUCH HARDER THAN WITH NUCLEAR WEAPONS

Nuclear programmes require large physical infrastructure.

AI can be developed inside data centres.

Models can be copied.

Weights can potentially be stolen.

Algorithms cross borders instantly.

Open-source models complicate control further.

And computational efficiency constantly changes the amount of hardware required to achieve a given capability.

AI arms control could therefore prove significantly more difficult to verify than traditional arms agreements.

That makes trust even more important.

And trust between Washington and Beijing is currently not abundant.

23. THE MOST LIKELY OUTCOME IS NOT A PAUSE

Despite repeated calls for slowing advanced AI, the economic and geopolitical incentives point in the opposite direction.

Capital is accelerating.

Infrastructure is expanding.

Energy investment is increasing.

Governments are integrating AI into national-security systems.

Companies are racing for market leadership.

China wants technological independence.

America wants technological dominance.

Neither side wants to discover that the other achieved a decisive advantage while it exercised restraint.

The most likely scenario therefore appears not to be an AI pause.

It is an AI acceleration.

24. AND THAT CHANGES THE QUESTION COMPLETELY

For years we have asked:

What can artificial intelligence do?

Soon the more important question may become:

What will nations do because they fear another nation developing artificial intelligence first?

That is a much more dangerous question.

Technology itself has no geopolitical ambition.

States do.

Companies compete.

Capital seeks returns.

Military institutions seek advantage.

Politicians seek growth and security.

Combine those incentives with a technology whose capabilities are advancing rapidly and the result begins to resemble something history has seen before.

Not because AI is a nuclear weapon.

It isn’t.

But because the strategic logic is becoming familiar.

Nobody wants an uncontrolled race.

Nobody trusts the other side enough to stop racing.

And therefore everyone accelerates.

CONCLUSION — THE AI SAFETY PROBLEM IS NOW GEOPOLITICS

Trump’s rejection of calls to slow American AI development may eventually be remembered as something larger than another political controversy.

It exposes the fundamental contradiction now sitting at the centre of artificial intelligence.

The people building increasingly powerful systems are beginning to debate how dangerous those systems could become.

Governments are simultaneously concluding that possessing the most powerful systems may be essential to national security.

Those two realities are moving directly toward each other.

The uncomfortable truth is that the greatest obstacle to AI safety may eventually not be technological ignorance.

It may be geopolitical rationality.

If America believes China will continue, America accelerates.

If China believes America will continue, China accelerates.

If corporations believe competitors will continue, corporations accelerate.

Everyone can understand the risk.

Everyone can prefer coordination.

And everyone can still rationally choose to race.

That is what makes the situation dangerous.

The AI race may therefore have entered a new phase.

Not because machines have become uncontrollable.

But because the strategic incentives surrounding them may be becoming uncontrollable first.

And history suggests that when technology, capital, military power and national prestige converge around the same objective, asking everyone simply to “slow down” is rarely a strategy.

It is a hope.

Roberto Balbino
CEO, Fintech SCB
Independent Analyst of Finance, Economics, Geopolitics and Global Markets