Why the Future of Artificial Intelligence May Be Decided Outside America

While Silicon Valley argued over whether artificial intelligence should be open or closed, something more consequential happened in Washington. OpenAI and Anthropic, two laboratories competing for engineers, customers, capital and prestige, found themselves standing on much the same ground. Both wanted the American government to take a closer look at the most powerful models before they reached the wider world.

Their argument was not frivolous. A machine capable of discovering vulnerabilities in computer networks, designing biological experiments or carrying out complex operations with little human supervision might justify scrutiny before release.

Yet every safety barrier creates a boundary. Those inside it help determine how high it should be. Those outside must find a way across.

Then Jensen Huang intervened.

The Nvidia chief executive argued that open models were not a dangerous residue of the artificial intelligence revolution, but one of its essential foundations. Governments, universities and companies needed systems they could inspect, adapt and operate themselves. Cyber defenders could not be expected to confront attackers armed with powerful AI while relying entirely on restricted services controlled by a handful of private companies.

At almost the same moment, China supplied the demonstration. Moonshot AI released the weights of Kimi K3, a vast Chinese model designed for reasoning, coding, vision and long-running autonomous work. It was not necessarily the most intelligent model in the world. That was not the point. It could be downloaded, altered and installed beyond the reach of any American laboratory.

While Washington debated the conditions under which intelligence should be released, Beijing was constructing the institutions through which it could be distributed.

The new AI cold war is not principally a competition between Kimi and Claude, DeepSeek and GPT, or one benchmark score and another. Models change too quickly. Today’s leader becomes tomorrow’s ordinary tool.

The deeper contest concerns who owns the infrastructure, writes the standards, trains the engineers, controls the data, supplies the cloud and determines the conditions under which intelligence may be used.

It is a struggle over the operating system of the twenty-first century.

Nvidia’s declaration

Huang has spent much of the past year presenting open artificial intelligence as a matter of economic participation and national sovereignty.

In March, Nvidia established the Nemotron Coalition, bringing together model builders and research laboratories to develop advanced open models. Huang described such systems as the lifeblood of innovation, allowing students, scientists, start-ups and industries to participate in the AI revolution. Nvidia later joined the Open Secure AI Alliance, which argues that open technology is necessary for security as well as innovation.

There is conviction in this position, but also commercial logic.

Nvidia does not need to own the model at the top of the system. It sells the machinery beneath it: processors, networking equipment, software and increasingly complete AI factories. Every new laboratory, national model, corporate installation and university research programme creates further demand for computation.

A world dominated by two or three closed laboratories would concentrate enormous power in those companies. They could eventually design their own chips, bargain down infrastructure costs or direct customers towards their private clouds.

A world of thousands of models is more favourable to Nvidia. Competition may make intelligence cheaper, but running that intelligence still requires processors, memory, electricity and software.

Huang’s defence of openness is therefore neither purely philanthropic nor necessarily cynical. His commercial interest and his strategic argument point in the same direction.

Countries do not want their public administration, defence systems, hospitals and industrial secrets permanently dependent upon a foreign subscription service. Businesses do not want a model provider to change prices, withdraw access or alter acceptable-use policies after an entire operation has been built around it.

Nvidia offers a different promise. Build your own system. Keep your own data. Adapt the model to your own language and institutions.

But run it on Nvidia.

The company does not need to possess the world’s intelligence. It needs the world’s intelligence to require its machinery.

Washington writes the threshold

The American government is not yet operating a compulsory licensing system for artificial intelligence.

An executive order signed in June created a framework through which developers may provide designated frontier models to the government before wider deployment. The arrangement is officially voluntary and initially concentrated on advanced cyber capabilities. The order expressly rejects a general system of government approval for new models.

Yet voluntary systems can become powerful.

A company seeking government contracts, access to critical industries or the confidence of major corporate customers may find participation difficult to avoid. A test designed for national security can gradually become a market standard.

Anthropic insists that it has never advocated a general prohibition on open models. It describes less capable open systems as a public good. Its anxiety concerns the small number of models that might provide dangerous biological or cyber capabilities or allow foreign competitors to reproduce American advances through industrial-scale distillation.

That case deserves to be treated seriously. A downloadable system cannot monitor its users or revoke access after release. If an AI model materially lowers the expertise required to conduct a biological attack or penetrate critical infrastructure, openness becomes more than a philosophical virtue.

But safety rules also have economic consequences.

OpenAI and Anthropic possess secure facilities, specialist testing teams, government relationships and large legal departments. Smaller laboratories and university groups do not. A compliance regime that appears modest to an incumbent can become prohibitive to a challenger.

This does not require conspiracy. Governments naturally consult the companies with the greatest technical knowledge. Those companies explain what responsible development requires. Their practices become the standard. Their costs become the price of entry.

Washington may believe it is constructing a safety barrier. Those outside may discover that it is also a wall.

China builds the alternative

China has chosen a different starting point.

It does not deny the dangers of artificial intelligence. Nor has it abandoned regulation at home. Beijing controls online services, demands security assessments and imposes political requirements that no American laboratory would accept.

Internationally, however, China speaks a different language.

It talks about access, development, sovereignty and the danger that poorer countries will become tenants in an intelligence economy owned elsewhere.

In July, representatives gathered in Shanghai for the establishment of the World Artificial Intelligence Cooperation Organization. WAICO is intended to become the first intergovernmental organisation devoted specifically to AI, with its headquarters in Shanghai.

Chinese officials describe it as an answer to the demands of the Global South. They say it will promote capacity building, shared governance and equal participation by countries that have had little influence over the rules now being written in Washington, Brussels and California.

Xi Jinping presented the organisation as a mechanism for preventing a new global intelligence divide. China offered thousands of training places, cooperation centres linked to regional organisations and assistance with practical applications, including meteorological warning systems. Beijing has invited further countries to join and portrays WAICO as a forum for building trust and reaching international consensus.

The language is diplomatic. The strategy is structural.

China is assembling the pieces of an export system.

The first piece is the model itself.

DeepSeek showed that a Chinese laboratory could produce a highly capable system at a cost far below the vast sums associated with the American frontier. Kimi K3 extends the argument. A government or company need not wait for China to overtake every American benchmark. A model becomes strategically useful when it is capable enough, affordable enough and available without permanent dependence upon a distant provider.

The second piece is infrastructure.

For more than a decade, Chinese companies have built telecommunications systems, cloud services, data centres, surveillance platforms and digital payment networks across Asia, Africa, Latin America and the Middle East. AI can now be placed on top of this existing technological geography.

A country already using Chinese networks, servers or cloud systems faces lower costs when adopting Chinese models. What appears to be a choice between two pieces of software is often the final stage of a much older infrastructure decision.

The third piece is training.

Engineers, civil servants and students invited to Chinese universities or technical programmes learn particular systems, methods and assumptions. When they return home, they do not merely possess knowledge. They carry professional relationships and familiarity with Chinese platforms.

The fourth piece is language.

Most of humanity does not conduct its daily business in English. The decisive model for an African ministry, an Arab university or a Southeast Asian company may not be the system that wins an American reasoning test. It may be the one that understands local law, administrative vocabulary, accents, documents and cultural references.

China can offer to adapt models alongside national governments rather than asking those governments to adapt themselves to a service designed in California.

The fifth piece is the institution.

WAICO gives these activities a political centre. Training programmes become capacity building. Chinese models become tools against technological inequality. Infrastructure contracts become exercises in national sovereignty. Technical standards become the product of international cooperation.

The Belt and Road Initiative financed the ports, railways, energy systems and communications networks through which economies move. China’s AI diplomacy seeks influence over the systems through which governments interpret information and make decisions.

The offer will be attractive.

Use a model that can be operated inside the country. Keep national data on national servers. Train local engineers. Build applications for agriculture, education, health, translation and public administration. Avoid the subscription costs and political conditions of Western providers.

Yet Chinese sovereignty can also create Chinese dependence.

The model may be local while the hardware, cloud architecture, updates, technical expertise and standards remain Chinese. A government may escape reliance upon an American corporation only to enter a wider technological sphere centred on Beijing.

China is not giving intelligence away. It is building an ecosystem in which Chinese technology becomes the easiest, cheapest and most politically acceptable choice.

The real battlefield

The United States still holds formidable advantages. It possesses the leading chip designers, the largest private laboratories, deep capital markets, world-class universities and much of the cloud infrastructure on which advanced AI depends.

But technological leadership and ecosystem control are not identical.

The strongest model may be too expensive for most countries. The safest service may require data to cross borders. The most sophisticated laboratory may refuse politically sensitive applications or withdraw access during a diplomatic dispute.

The real battlefield is therefore larger than the laboratory.

It includes the chips and electricity beneath the model, the languages in which it speaks, the universities that teach it, the applications developers build around it and the rules governments adopt for its use.

It concerns where national data are stored and whether a foreign company can inspect them. It concerns whether a ministry can modify its own system or must request permission. It concerns which technical standards become so ordinary that later alternatives appear inconvenient or incompatible.

American laboratories are trying to protect the frontier.

Nvidia is trying to supply the entire field.

China is trying to organise the countries beyond it.

No single decision will announce the winner. A university will choose a model for its students. A ministry will install a translation system. A national cloud will be built around one technical architecture. Developers will begin writing applications for it. Its standards will enter contracts and legislation.

Gradually, the infrastructure will disappear from view.

The country that shapes the world’s artificial intelligence ecosystem may not be the one that builds the single smartest machine.

It may be the country whose models, institutions and standards become so deeply embedded that the rest of the world no longer notices they are there.

It will have become the operating system of the twenty-first century.