What can middle powers do for frontier AI governance?

6/7/2026

Written by:
Markus Anderljung and Stephen Clare

Sometimes people round the influence of countries other than the US and China on the AI trajectory down to zero. The most capable models are built by a handful of American companies; China is the only serious challenger; and the US government is increasingly interested in controlling who – including which states – can access frontier systems. It looks like everyone else will have to be content watching from the sidelines.

That seems like a mistake. In this piece, we give the other side of the argument, offering reasons why states other than China and the US can play important roles. The “middle AI powers” – countries which have strong political institutions and economic power, but lack frontier AI companies – have already shaped frontier AI development, and in fact could do more to shape the AI trajectory than they currently are.

These AI middle powers are, roughly, the G20 (minus the US and China), and a few AI-specific players like Taiwan, the Netherlands, Norway, and the UAE.

Here are six claims about what they can do, and why it matters.

1. Middle powers have already shaped AI development

Several of the most consequential AI governance moves since ChatGPT were made by middle powers. In fact, at least until 2026, middle powers arguably had more influence on frontier company behaviour than the US government, including by:

Pioneering the AISI model: Within two years of its founding, the UK’s AI Security Institute became the world’s leading evaluator of new AI models for dangerous capabilities and risks. It’s since been copied (albeit imperfectly) by the US, Singapore, Japan, Korea, Canada, and the EU, among others.
Launching the Summit series: Companies made concrete commitments in Bletchley, Seoul, and New Delhi; first, committing to test their models for dangerous capabilities, then to publish safety frameworks. This, in turn, made it far easier for US state legislators and EU regulators to include such provisions in legislation.
Leading on regulation: The EU AI Act’s provisions for general-purpose AI, plus its Code of Practice, may prove one of the most important developments of the past two years; we’d bet Chinese frontier developers start producing safety frameworks and model specs significantly sooner as a result.

So middle powers can influence frontier AI development. And since the policy research and advocacy space in these countries tends to be less crowded than it is in the US, it’s also plausible that individuals seeking to shape frontier AI governance can have more impact by working in middle powers. For example, even if a country like South Korea has 1000x less leverage over AI than the US, it may well be more than 1000x easier for an individual with the right background to improve South Korean AI policy.

2. Middle powers have assets that give them leverage

People often say middle-power strategy means first building leverage: investing in capacities that allow them to potentially impose costs on frontier AI countries, and using this power to make deals or extract concessions. For example, this might involve developing AI inputs that they want to access, datacentres they want to use, or a market they want to sell to.

How do middle AI powers stack up against the US and China on classic indicators of power? They’re actually in pretty good shape. Collectively, they account for more of the world’s GDP than the US and China combined, and beat them individually across military spending, R&D spending, and total investment.

Things look rougher when you narrow to AI. The US dominates domestically-housed high-end compute, private investment in AI, frontier-lab valuations, and chip design. But middle powers aren’t totally out of the game. For example, they generate lots of electricity, which could support a far more aggressive datacentre buildout. They also occupy critical links in the AI chip supply chain: lithography equipment (the Netherlands’ ASML holds ~80% of the lithography market and 100% of EUV), high-bandwidth memory (Korea’s Samsung and SK Hynix supply around three-quarters of all HBM), advanced packaging (Taiwan’s ASE is the largest OSAT firm) and fabrication itself (famously led by TSMC). Further, Ukraine has quickly become a leader in drone warfare, e.g. leveraging its expertise for military support from Europe and the Gulf.

3. Middle powers could grow their leverage by investing in AI

Middle powers can increase their leverage – and better position themselves to share in the benefits of AI – by growing their participation in the AI stack.

Start with compute. The US represents about 25% of global GDP but hosts roughly 75% of the world’s compute performance. Domestic compute capacity might not be a prerequisite for capturing economic benefits from AI, but will very likely confer geopolitical power: countries that host datacentres could control, restrict, or monitor foreign access to AI models running on those chips, a considerable source of leverage.

Middle powers could probably do more to build datacentres. AI companies want to diversify their footprint. Concentrating their compute in the US will become a business risk if datacentres become politically toxic there, or if Washington restricts foreign access to models hosted on American servers. The binding constraint for building more datacentres is usually speed: how fast a site can come online. So the winners will be the countries that can clear the way for construction: those with spare grid capacity, or which allow developers to generate their own power (“behind-the-meter generation”).
That last point will worry climate advocates, since behind-the-meter power today usually means gas. But there may be a bargain to strike here. The current data center buildout is one of the largest capital expenditures in a century; a middle power could plausibly trade permits for short-term gas generation in exchange for commitments to invest in grid renewal and eventually switch to renewables.

All that said, participation in the AI stack doesn’t just mean building compute. Middle powers could also:

Build domestic datacentre infrastructure companies. Hyperscalers increasingly rely on specialist providers like Nebius and Nscale rather than building everything themselves, which creates opportunities for new entrants, including in Europe.
Poach frontier talent. Very few middle-power leaders have seriously tried to lure senior people away from leading labs to found domestic champions. They should. In recent years, top researchers leaving frontier labs can often start companies that quickly reach $1+bn valuations (sometimes without much of an obvious product or plan).
Build on existing strengths. Some middle powers already own critical links (think chipmaking equipment in the Netherlands and semiconductor materials and equipment in Japan, for instance) and can deepen those positions while pushing hard to integrate AI into the parts of their economy with the most exposure: deep tech and pharma in Europe, finance and professional services in the UK.
Embrace creative destruction. Much of AI’s value will likely accrue to those who put it to work: raising productivity in existing industries, and developing and adopting the wave of new technologies it enables. It will likely be difficult for frontier AI companies, for example, to capture the majority of profits from AI-enabled pharmaceuticals, without developing those drugs themselves. To capture this value, states will likely need to embrace creative destruction, allow for old businesses that don’t adopt AI fast or competently enough to be superseded.
Invest in the frontier directly. To hold a stake in a post-AGI economy, own a piece of it. Japan and the Gulf are already heavily overrepresented among investors in OpenAI and Anthropic – SoftBank alone holds around 13% of OpenAI.

4. Middle power influence doesn’t need hard power

Building leverage is one thing; actually using it coercively against the world’s most powerful countries is another. In the context of trade wars, analysts often talk about escalation dominance: if two countries are trading punches, the one who can inflict more costs as things escalate has an advantage. In many cases the superpowers will hit harder, making middle powers reluctant to start a fight even when they have some leverage to use. For this reason, leverage might be most useful as a deterrent: something the middle powers can hold in reserve to discourage aggressive moves by the superpowers.

Fortunately for the middle powers, leverage is not the only way they can affect AI governance. They can also influence AI companies directly. Collectively, their markets matter a lot to frontier developers – only around 20% of ChatGPT users are in the US (though the enterprise market is more US-weighted). So when middle powers make market access conditional, the steps companies take to comply with those conditions can have much broader effects. A company might run pre-deployment evaluations to satisfy one country’s requirement, but what those tests reveal about a model's capabilities and risks can be used wherever the model is deployed. This is called the Brussels Effect; the EU is such a large market that many EU requirements have global impacts.
In addition to shaping incentives for companies, middle powers can produce public goods for responsible AI development. AISI evaluations are expensive to run, but if their findings are published they can be used by policymakers, deployers, and researchers around the world. Middle powers can also fund safety research and build accessible risk-management tools to make risk mitigation cheaper for AI developers and users everywhere.

Finally, there are more subtle forms of influence: spreading useful information, improving international coordination, and establishing or strengthening norms. Their impacts are hard to quantify, but can be significant. Recall that South Korea hosts the firms that make three-quarters of the world’s high-bandwidth memory, a critical input to every AI accelerator. Yet its biggest impact on frontier AI governance to date came not from using that leverage, but from convening countries at the 2024 AI Seoul Summit and encouraging developers to commit to publishing frontier safety frameworks. Sometimes soft power is preferable to hard power.

5. Middle powers can directly mitigate some risks

We’ve discussed how middle powers can shape the way frontier models are developed. But by controlling how AI is used within their borders, they can also act on (some of) the risks those models pose. They can reduce cyber and bio misuse, surveillance, authoritarian applications, democratic backsliding, and more.

To help middle powers focus these efforts, we can divide AI risks into four different categories according to which actors need to be involved to manage them:

Great-power problems, where only Washington and Beijing can move the needle. Preventing frontier labs from building egregiously misaligned AI systems that might seek to undermine control measures and disempower humans is usually thought to look like this. Middle powers have little direct influence on this risk, except insofar as they can shape how the frontier AI countries act.
Coverage problems, where the solution scales with the number of countries acting. Building resilience to cyberattacks might look like this: the more countries harden their infrastructure and implement effective monitoring systems, the harder attackers will find it to identify targets and execute attacks.
Weakest-link problems, where one lax jurisdiction introduces system-wide vulnerabilities, so you need near-universal participation. Broad monitoring and enforcement might be needed to reduce AI-enabled biorisks, for example, because without it, attackers could operate out of poorly-defended jurisdictions.
Best-shot problems, where only the best attempt has to work, so one state’s success suffices for everyone. A cure for a disease discovered by one lab can be used worldwide, for example. In AI, a single team could make a breakthrough in interpretability – tools that help researchers understand an AI model’s behaviour – that can be widely adopted to make development safer.

Middle powers can only help with great-power problems indirectly. But on coverage, weakest-link, and best-shot problems, they can just do things.

6. Middle powers can work together

So far we’ve established that middle powers have a range of options for influencing AI governance. Many of those options will be even more powerful if they’re exercised by multiple countries in concert. And crucially, they don’t necessarily need the US to lead these efforts.

The natural assumption used to be that the US would set global frontier AI standards. America would write the rules and allies would adopt them, getting access to American compute in return. Right now, that path looks less likely than it used to. But this doesn’t mean that coordination is impossible. Instead, it means US allies might have to go ahead on their own.

What would that look like in practice? It could take several forms:

Unilateral recognition of (or convergence on) standards. A country could unilaterally recognise EU compliance for general-purpose AI models, telling frontier developers that meeting the EU’s bar is sufficient for entering its market. Alternatively, a country could design its requirements to track the substance of the EU’s without formally tying its policy to Brussels. Either way, the country inherits a workable standard rather than building one from scratch, and developers already meeting EU rules face little additional burden.
Investing in the Summit series. Middle powers can help make the 2027 Geneva AI Summit a success. In particular, they could coordinate to push for a new round of frontier-company commitments, strengthening the series’ legacy of pushing companies to strengthen their risk management practices. As AI development is increasingly reliant on the use of internal models, new commitments related to transparency would be welcome: getting companies to provide more information about how they are using AI to accelerate their research, and how they’re managing the related risks, could prove highly valuable.
Building fora that don’t depend on the US. Middle powers could also stand up new venues to coordinate on trade, national security, and frontier AI that don’t rely on American or Chinese participation.
Pooling bargaining power. The hardest but highest-impact option is for middle powers to coordinate, pooling their leverage and market power so that they can make stronger asks of frontier companies and states. For example, middle powers could align more closely on export controls, jointly secure early model access, harmonise procurement requirements, form a frontier model buyers’ club, or co-invest in developers sitting just behind the frontier.

What this means

This piece has made the case for the importance of middle powers. (1) They have already shaped how frontier AI is built; (2) they hold important leverage over the AI supply chain; (3) they can grow that leverage – and their share of AI's upside – by investing more in AI; (4) yet much of their influence doesn’t depend on leverage at all; (5) they can directly mitigate many of the risks that matter most; and (6) they can coordinate to multiply all of this, with or without Washington.

So what does this mean for those of us working on AI governance? It expands both the set of countries to consider and the set of available levers. All things considered, we’re still uncertain about how much middle power actions will shape the broader AI trajectory. But it seems worth investigating how to:

Make middle-power policymakers feel that AI is a big deal. Belief, not proposals, is usually the blocker to action.
Build expertise, networks, and talent in middle powers. Mentor people, build the field. The policy space outside the US is sparse enough that a single capable person can make a big difference.
Make Geneva 2027 AI Summit great. Push for a new round of frontier company commitments, especially about how companies are using AI to accelerate their own R&D.
Support AI middle power coordination. Flesh out what unilateral recognition of EU GPAI compliance, joint asks for early model access mirroring the recent executive order, and a frontier-model buyers’ club might look like.
Help middle powers seize AI opportunities. Accelerate (responsible) adoption, and race to be the first country to approve and commercialise what AI invents.
Build societal resilience. Countries can act on coverage problems, and their efforts protect their own people regardless of what others do. They can also create public goods that help everyone.
Engage sovereign wealth funds and other large investors. This is how some middle powers can take a direct stake in the frontier.
Spread useful norms. Lay the groundwork for international agreements, and the first serious policy thinking on questions like legal personhood for AI systems, where the field is almost empty.

Of course, middle powers also face major challenges in shaping the trajectory of AI. As the technology’s social, political, economic, and security implications become clearer, those challenges will probably grow even more difficult. But we’re confident that they can do more than spectate.