Most Countries Will Not Win the AI Race, but They Still Need a Strategy

Secretary-general addresses World Artificial Intelligence Conference Meteorological Forum, July 17, 2026. UN Photo.

Artificial intelligence is often described as a global race. In reality, very few countries are racing. In 2025, US institutions produced 59 notable AI models and Chinese institutions 35. More than 90% of prominent frontier models came from industry. The concentration is even starker underneath the models themselves. Nvidia supplies more than 60% of global AI compute capacity. The United States hosts 5,427 data centers, more than 10 times as many as any other country. Almost every leading AI chip is fabricated by a single Taiwanese company, TSMC. US private investment in AI reached $285.9 billion in 2025, while investment in local EU AI companies reached only approximately $6.8 billion.

This is not a simple story of American dominance. China leads in the volume of AI research publications and citations, while Chinese inventors also dominate generative-AI patenting. The money, infrastructure, hardware, and frontier models on which much of the AI economy rests are concentrated in two countries and a small number of companies.

For governments elsewhere, this imbalance creates a strange starting point for national AI policy. They are told that AI will shape economic competitiveness, public administration, and national security, while much of the technology needed to participate is controlled abroad.

Catching up is possible in parts of the stack. Countries can build models in their own languages, develop applications, invest in research and computing capacity, and create domestic AI companies. Open-source development has widened participation, including outside the US, China, and Europe. Yet very few governments can plausibly reproduce the whole chain, from advanced semiconductors and large-scale compute to cloud infrastructure and frontier models. They need another strategy.

Technological dependence is hardly new. Walk through a government ministry or a UN office and there is a good chance that the computers run Microsoft software, the chips were designed abroad, and much of the underlying digital infrastructure comes from foreign companies.

Governments have lived with such dependencies for decades. AI extends the role of foreign technology from the software governments use to the systems they rely on to analyze information and make decisions. Some governments are experimenting with independent AI solutions for public administration, education, health, intelligence, defense, and scientific research, but this is more the exception than the rule on the global scale.

That makes the loose talk about “sovereign AI” worth examining. Several governments, including France, India, Saudi Arabia, and others have announced sovereign models, national computing facilities, and domestic data centers. Some of these investments make sense. The trouble starts when sovereignty becomes shorthand for reproducing the AI stack within national borders. Even a national model may run on Nvidia chips fabricated by TSMC, train on a foreign cloud, and depend on software developed elsewhere.

Full technological self-sufficiency is a poor benchmark for most countries. Dependence itself is also a poor measure of sovereignty. What counts is the character of that dependence: whether governments understand it, whether they have alternatives, and how much leverage they retain over the companies and countries supplying critical technology.

A government that cannot assess the systems it buys, negotiate the terms under which its data is used, or move to another provider has a sovereignty problem, regardless of where its data center happens to sit.

The concentration of AI power has produced an understandable hope that international institutions might provide a counterweight. Proposals for shared public compute, international research facilities, and public-interest AI deserve serious discussion. There are useful precedents for countries pooling expensive scientific infrastructure. Expectations of the UN itself, however, should remain modest.

The organization has just created its most substantial AI institutions to date. The Independent International Scientific Panel on AI brings together 40 experts from across the world’s regions, while the Global Dialogue on AI Governance held its first meeting in Geneva in July 2026. The panel can provide governments with a common body of scientific evidence. The dialogue gives states and other actors somewhere to discuss questions that would otherwise be negotiated largely among governments with advanced AI industries and the companies building the technology. Neither institution regulates AI or provides technological infrastructure.

That division of labor is reasonable. The UN does not have the resources to compete with the investment taking place outside it. Google alone projected $195 billion to $205 billion in capital expenditure in 2026, driven primarily by heavy investments in data centers, custom AI chips (TPUs), servers, and networking infrastructure to scale AI. The UN, meanwhile, is dealing with severe financial pressures across its existing work. Asking it to finance a public alternative to the commercial AI stack would confuse a forum for international cooperation with an industrial policy institution.

The UN can put facts on the table. It can give countries without large AI industries a place to challenge decisions that affect them. It can establish broad principles and expose concentrations of power that companies would rather treat as commercial questions. Those are worthwhile functions. But the harder work falls to states.

Five Elements of an AI Strategy

For countries outside the handful of AI powers, a credible strategy starts with a clear view of where dependence creates risk and where it does not.

First, governments need to know what they depend on. Many national AI strategies begin with ambitions: the number of AI startups, national models, compute targets, and skills programs. A more useful starting point is an inventory. Which cloud providers host sensitive government functions? Which models are used by public agencies? Where is government data stored? Which suppliers provide computing capacity? Which systems could be replaced and at what cost? Where are the country’s technical skills concentrated? What happens if access to a particular service is restricted? Such an exercise will reveal dependencies that are benign and others that deserve attention. It will also stop governments from spending scarce money chasing every part of the AI stack.

Second, states need enough technical capacity to be intelligent customers. The shortage that should worry many governments is not the absence of their own frontier AI company; it is the absence of people inside the state who can independently assess what AI companies are selling them. Governments need engineers and technical officials who can test systems, scrutinize contracts, and advise political leaders. Procurement officials need to understand switching costs and data arrangements. Regulators need access to expertise that does not come solely from the companies they oversee. Universities and public research institutions have a role here too. Without that capacity, governments outsource judgment alongside technology.

Third, governments need to decide what they are unwilling to outsource. This will differ from country to country. Military applications and intelligence systems raise different concerns from translation software. National identity systems deserve different treatment from office productivity tools. Sensitive government data may justify greater domestic control even when the applications using it rely on foreign technology. Public compute can also serve purposes that commercial access cannot. It can give universities, researchers, and local firms room to experiment without depending entirely on foreign platforms.

Governments will have to make choices. Trying to manufacture advanced chips, operate a national cloud, build a frontier model, and fund domestic alternatives to every foreign platform is beyond the fiscal capacity of most states. Spending money on the appearance of sovereignty can come at the expense of capabilities states genuinely need.

Fourth, countries should take AI non-alignment seriously. The analogy with the Non-Aligned Movement is imperfect, though useful. During the Cold War, non-alignment gave states a way to protect room for maneuver amid competition between larger powers. Countries outside the leading AI powers face a related pressure today.Neither Washington nor Beijing presents a single, sealed technological bloc. Yet export controls, semiconductor supply chains, cloud infrastructure, technical standards, and model ecosystems are acquiring geopolitical significance. Countries will face pressure, sometimes explicit and sometimes commercial, to build around one set of technologies.

An AI Non-Aligned Movement could give countries with very different political systems a reason to cooperate around a narrower set of interests: access to compute, interoperability, fair infrastructure deals, protection against vendor lock-in, and bargaining power with large technology companies. It would also face a harder problem than its 20th-century predecessor. A government can change a diplomatic position. Replacing a cloud architecture is harder. Once ministries have migrated data, civil servants have learned particular tools, and domestic companies have built businesses around a particular ecosystem, the cost of changing course rises. Some of the most consequential geopolitical alignments of the AI era may therefore take place through procurement decisions rather than treaties.

Non-alignment cannot guarantee technological independence, but it can give governments a reason to preserve alternatives before those alternatives become prohibitively expensive.

Fifth, countries should pool resources where the national scale is too small. This is where international cooperation has a more practical role than grand plans for a UN-built global model.

Not every country needs its own frontier-scale computing facility. Regional compute facilities could serve universities, public agencies, and companies across several countries. Governments can share model-testing capacity and technical expertise. Common procurement requirements can make it harder for suppliers to impose restrictive contracts. Regional development banks can finance infrastructure on conditions that protect data access, competition, and the ability to switch providers.

Coalitions can also bargain differently than individual governments. A small country negotiating with a massive cloud computing hyperscaler has limited leverage. A regional market representing hundreds of millions of people has more.

The same logic applies when governments negotiate with companies building data centers. Access to land, electricity, water, and tax concessions gives states bargaining power that they can use to secure local capacity, infrastructure investment, and other public benefits rather than treating the presence of a data center as an end in itself.

The African Union, the Association of Southeast Asian Nations (ASEAN), regional development banks, the European Union, and smaller groups of states can all do things that would make little sense at the level of 193 UN members. The UN can still provide political space for these ideas and help countries compare their experiences.

Pooling capacity does require compromises, but that is familiar territory for states. Governments already share infrastructure, establish common standards, and pool bargaining power when acting alone produces worse results.

The Race May Change Before It Ends

There is another reason for governments to avoid tying themselves too tightly to today’s AI order: nobody knows exactly what that order will look like five years from now. The sums being committed are extraordinary. There are also unresolved questions about the returns on this spending. AI companies are generating revenue at a remarkable pace, though infrastructure costs are rising alongside it. The economics of AI models may change, and open models may reduce some dependencies. Companies that look indispensable today may not occupy the same position a decade from now.

Most governments do not have the capital to place bets on that scale, nor do they need to. The decisions they make now on cloud contracts, data-center agreements, government procurement systems, or investment in technical education will still have long lives.

A serious strategy starts with governments knowing where they depend on others, building enough expertise to negotiate from an informed position, protecting the capabilities they cannot afford to lose, retaining alternatives where they can, and working with other states when their individual weight is too small. Most countries will not win the AI race, but they do have a choice over how they run it.

Martin Wählisch is an Associate Professor of International Relations at the University of Birmingham and a member of its Centre for AI in Government. His research focuses on international affairs, AI governance, and multilateral cooperation.