Middle Powers and Transformative AI

What’s at stake for middle powers

In a world where AI systems do most cognitive work, reliable access to frontier AI becomes a precondition for having a relevant economy, a functioning security apparatus, and sovereignty in any meaningful sense.

Consider what a middle power that stays on the sidelines may look like in a few years.

As the size of its economy shrinks in relative terms, access to its market matters less to the companies building frontier AI. Its government, in turn, has less influence over the safety standards that shape how the companies develop their models. Frontier capabilities reach its industry and government agencies late or not at all,⁴⁴ For example, Anthropic’s Glasswing program gave a small group of trusted organizations early access to Claude Mythos; the EU’s cybersecurity agency ENISA had to negotiate its way in and, according to reporting, still lacks access to the newest versions. as staged releases become the default and compute scarcity forces providers to aggressively prioritize who gets access. Its national security depends on access to AI systems hosted by others and used on terms set abroad, which can be restricted at any point, even if the intent is not hostile.⁵⁵ One response to the access problem is for middle powers to train frontier systems themselves. We are skeptical that this is the right priority: the frontier is extraordinarily expensive to reach, and if AI research itself accelerates, being a few months behind could compound into a gap that never closes. Reasonable people disagree, and there are versions of the idea we take seriously, such as a capable sub-frontier model for a few critical uses or a multilateral project pooling several countries’ resources. That said, we do not expect to fund direct efforts to build frontier models through this RFP. We are more interested in work that secures reliable access to frontier models, including hosting and deep partnership arrangements with frontier developers, and in analysis that rigorously compares various options middle powers might wish to consider.

At the mild end, this is a country that is less prosperous and consequential than it would otherwise have been. At the severe end, it is one whose government has no real say over the technology underpinning its economy and its defense, which is hard to distinguish from losing its sovereignty altogether.

This matters not just for middle powers: having more countries involved in frontier AI governance increases the chance that safety oversights will be noticed and corrected, and improves democratic accountability for everyone.

The window is closing

Most of the leverage middle powers hold today will quickly fade if action is not taken.

Talent matters less once AI systems do most cognitive work. Supply chain chokepoints erode as the great powers spend heavily to indigenize the steps they still depend on others for. Standards on AI company development practices are only worth complying with as long as a middle power’s market is sufficiently big. And the compute buildout is happening now, mostly without middle powers. Sites, grid capacity and financing are being locked in elsewhere, and a country that is not part of the buildout in the next few years is unlikely to be part of it at all.

Since extraordinarily short timelines to transformative AI are plausible, urgent action is needed.

What we’re interested in funding

Who we fund. We’re looking to fund a wide range of organization types: existing teams looking to deepen and expand their middle power AI policy work, founders looking to enter the space, and individuals doing research and taking action independently.

What kinds of work. We fund both research and agenda-setting work, and practical implementation:

• Research and strategy. This is a new and fast-changing field, with many open research questions.

• Policy engagement. Research must be coupled with changes to policy to have an impact.

• Building institutions. We will need new teams, within and outside government, to do this work.

• Talent and convenings. All of this is enabled by a larger talent pool with the relevant knowledge, skills, and connections to make a difference.

Each focus area below describes what middle powers need to do and why it matters, then lists examples of work we would consider funding.

The lists are illustrative, not exhaustive: feel free to apply with other proposed projects. Naming a project or organization as an example does not mean we endorse everything the organization does. Nor does it mean the area is already covered.

What middle powers should do

There is still time for middle powers to act in ways that keep them relevant, and a world in which they stay relevant is a safer one for everyone.

Concretely, we think middle powers need to:

1. 1.

   Recognize the stakes. Treat transformative AI as the top strategic, economic, and national security priority it is, and build the political will and institutional capacity to act.

2. 2.

   Build leverage. Maintain and strengthen their positions in the data center buildout, semiconductor supply chain, and downstream industrial AI applications.

3. 3.

   Use leverage well. Work out what to ask for (access to frontier systems, transparency into what frontier developers are building, enforceable safety requirements, etc.), coordinate with other middle powers, and use their leverage to partner with great powers on securing these outcomes.

4. 4.

   Help the great powers coordinate. Build the verification technology, institutions, and diplomatic capacity that make agreements between great powers more likely and more robust.

2.1 Recognizing the stakes

To take any substantive actions to build and use leverage, governments must first treat AI as a top strategic priority. Awareness is growing in many capitals, but not fast enough. This area funds the people and institutions that give officials an accurate picture of what is coming, what it means for their country, and what they can do about it.

1. 1.

   Policymaker awareness and political will. Decision makers are busy people. We need timely, accessible materials and briefings in local languages to help policymakers understand key AI issues and what their policy options are. We are especially interested in funding teams that combine frontier AI expertise with local policy experience. Example projects:

   • The UK AI Security Institute’s Strategic Awareness team helps the rest of the government understand frontier developments. Helping other governments set up similar teams could be valuable, e.g., by supporting secondments.

   • Organizations developing and disseminating policy ideas, such as the AI Policy Network and Secure AI Project in the US, and briefing teams such as CivAI in the US and CLTR in the UK. Most middle powers have no equivalents.

   • Crisis simulations and tabletop exercises for officials and national security experts, such as at the Munich Security Conference.

   • Country-specific analyses of what transformative AI would do to an economy and its public finances, and what hedging against it looks like, as an entry point where economic arguments land better than arguments about catastrophic risk.

   • Journalism fellowships covering frontier AI in the outlets policymakers read, in Seoul, Tokyo, Ottawa and Canberra as well as Paris, Berlin and Brussels, modeled on Tarbell’s US program.

   • Translation and national-language briefings of key documents, e.g., the International AI Safety Report’s summary for policymakers is published in all six official UN languages — which between them cover fewer than half the countries this RFP can fund.

2. 2.

   National AI strategies. Each country needs its own plan for the transition; a generic middle power playbook will not do. But a strategy is only useful if someone in government wants it and can act on it, so we will look for a credible route to the officials who would use the work. Narrower strategies on a single question can be as valuable as comprehensive ones. Example projects:

   • Delta Instituut’s National AI Plan for the Netherlands, commissioned by the economics ministry, with 52 recommendations including a national compute plan and a State Secretary for AI. Other countries (Japan, South Korea, Canada, etc.) could learn from this.

     • Similarly, Good Ancestors’ work on Australia’s position in the age of AI. KIRA Center’s compute strategy for Germany is an example of a more narrowly scoped report.

   • Analysis of how governments should use frontier AI themselves, so they can act at the speed the transition demands (e.g., Forethought’s work on the AI adoption gap).

3. 3.

   Public communication. Popular media can be very useful to disseminate these ideas, and targeting them to each country’s context is desirable. Three things need communicating: how fast this is moving, what is at stake, and what governments can do. Example projects:

   • Europe 2031 (along with video adaptations) reached policymakers across Europe in their own languages, spurring renewed attention to Europe’s role in AI development. Similar scenario work could be valuable for other regions, such as the Indo-Pacific or ASEAN.

   • Polling on attitudes to AI and its risks, with policymaker briefings on the results.

   • A dedicated publication venue for middle power research, or a special issue of an existing publication (Asterisk, Works in Progress, ChinaTalk, etc.).

2.2 Building leverage

More countries being involved in the frontier AI value chain reduces the risk of extreme power concentration. It also allows more countries to contribute to frontier safety and security requirements. Even without training their own frontier models, middle powers can supply the compute and energy the models run on, the equipment the chips are made with, the industries that put AI to work in the physical world, and the capital behind the developers.

1. 1.

   Data center buildout. The data centers used to train and run frontier AI models are in short supply, and public opposition is making them harder to build in the US. Countries that succeed at a large-scale data center buildout could gain significant leverage, since they will control a scarce resource AI developers will be competing over. Several middle powers have the electricity generation potential to house large-scale data centers, but permitting and politics can slow this down. Example projects:

   • An IFP for Europe focused on safely scaling AI innovations throughout society.

   • The Center for British Progress’s work on planning reform and the climate impact of data centers and Janet Egan’s case for an ambitious Australian data center strategy. Similar country-specific work is needed elsewhere.

   • The Future Society’s research on what influence hosting a foreign-owned data center provides.

   • Country-level analysis of what actually blocks buildout (grid queues, permitting, energy prices, copyright) and what would unblock it, e.g., for France, the Nordics, Canada, and Australia.

   • Research on the effects on AI timelines from building additional data centers in middle powers.

2. 2.

   Semiconductor supply chain strategy. Upstream of data centers in the AI value chain lies semiconductor production: The Netherlands, Japan, South Korea, Taiwan and Germany are necessary parts of the AI buildout. But they have not yet used their leverage to achieve AI policy objectives. Example projects:

   • Dylan Rogers’ critical evaluation of how little leverage ASML has so far bought the Netherlands.

   • Research on how long supply chain chokepoints last as the US and China indigenize key semiconductor manufacturing steps.

   • Concrete proposals for how chip supply chain countries should cooperate and negotiate with great powers.

3. 3.

   Robotics and the industrial explosion. AI developers will capture only part of the value of the AI transition: downstream industries that make efficient use of AI in the physical world could also benefit greatly. Middle powers are well-placed here, given strong advanced manufacturing sectors in Europe and East Asia. Example projects:

   • Carnegie’s analysis of why Europe is falling behind in general-purpose robotics, and Forethought’s modeling of the “industrial explosion”.

   • Experimenting with regulatory tools to streamline industrial AI applications, while promoting safety, in Germany or Japan, for example.

4. 4.

   Capital: sovereign wealth and pension funds as shareholders in frontier AI. National funds already own pieces of the AI stack, but generally do not use their position as shareholders to promote better safety practices. Example projects:

   • Langsikt’s work on using the Norwegian oil fund to set AI standards.

   • Research and briefings on AI governance policies for Australian superannuation funds and other pension funds worldwide.

   • Mapping which funds hold what kinds of governance oversight over frontier developers and their key suppliers, and an advisory function for funds that want to use their shareholder rights well.

2.3 Using leverage to make AI safer

Leverage is only worth building if it is used for something. Examples of what middle powers could push for include: transparency into the safety profile of frontier AI systems, including enforceable minimum standards; limits on the most dangerous capabilities and deployments; participation in evaluation and incident-reporting regimes; and reliable access to frontier systems for governments and evaluators.

1. 1.

   Compute-for-access deals. Countries that host frontier compute can attach conditions to which companies can buy or rent that capacity. A natural requirement would be that the host government and domestic critical infrastructure providers get frontier model access, e.g., to defend against cyberattacks. Example projects:

   • Policy development fleshing out how these deals could work, and advocacy to make them happen (see existing work from Anton Leicht and Simon Grimm).

   • Research on the security standards that would make AI developers comfortable serving frontier models from middle power soil.

2. 2.

   Using the EU’s market power: the AI Act and Code of Practice. Frontier developers currently have strong incentives to comply with safety requirements set by EU law rather than leave the single market. The AI Act’s obligations on general-purpose models are binding rules aimed specifically at frontier developers, with enforcement powers live since August 2026 and a voluntary Code of Practice that more than 20 companies have signed. Whether they shape practice globally depends on enforcement, on the Code being kept current, and on the scheduled reviews. Example projects:

   • A public watchdog that reads developers’ published safety frameworks, system cards and incident disclosures against the Code and flags where companies may fall short of AI Act obligations, for the AI Office to investigate further.

   • Setting up new independent third-party AI safety evaluators.⁶⁶ Evaluators without a middle powers rationale should look at Tailwind’s Independent AI auditors initiative.

   • Work to make independent third-party evaluation of models, including external evaluators embedded inside AI developers, required.

3. 3.

   Coalitions and coordination among middle powers. Middle powers will fare better if they find like-minded allies to negotiate joint deals together with great powers. Example projects:

   • Research on which forms of coordination have worked for smaller powers before, e.g., the arms-control diplomacy of Ireland and Austria, the Helsinki process, and the climate negotiations.

   • Analysis of what a coalition of chip-supply-chain countries could sensibly coordinate on, and what to avoid: e.g., lessons from OPEC.

2.4 Helping the great powers coordinate

Middle powers can also help make AI development safer by providing third-party evaluations of frontier models, building trusted verification technology, and facilitating AI safety dialogues between the great powers. If the United States and China are to coordinate on AI safety standards and policies, the agreement will need evaluation and verification infrastructure that neither of them can credibly supply alone.

1. 1.

   Independent evaluation and auditing. Middle powers do not develop frontier models of their own, but they can evaluate frontier AI models and companies for safety best practices. The UK’s AI Security Institute has shown what a well-resourced public evaluator can do. Countries such as Australia, Germany and the Netherlands can learn from this as they set up their own AISIs. Example projects:

   • Additional work like the UK AI Security Institute’s incident report on unsanctioned agent behavior found during cyber testing, which gave the whole world better situational awareness.

   • SaferAI’s evaluation of Chinese open-weight models; European and Asian organizations may be better placed to do this than US evaluators.

   • AVERI, which is building the ecosystem of third-party auditors for frontier AI.

   • Shared evaluation methods and standards so that one jurisdiction’s evaluations can be relied on by another.

2. 2.

   Verification technology and institutions. Verification means being able to check claims about AI development, e.g., how much compute was used, what a model can do, or whether a commitment was kept, without taking the developer’s word for it. Any agreement between the great powers will depend on robust verification. Example projects:

   • Work like the Singapore AI Safety Hub’s, and UK-based Amodo Design’s verification technology for US-China agreements.

   • Research papers such as Verification for International AI Governance and Verifying Rules on Large-Scale Neural Network Training via Compute Monitoring, and converting research into technology development and prototyping.

   • A small, neutral policy organization with deep technical verification expertise, modeled on VERTIC or the IAEA’s role in nuclear arms control.

3. 3.

   Convening and diplomacy. Direct US-China talks matter most, but middle powers can sometimes be useful in making these happen. Example projects:

   • The Singapore Consensus on Global AI Safety Research Priorities and the International Scientific Exchange, which brought researchers from the US, China and elsewhere together on a shared agenda, including a proposal for cross-border notification of AI agent incidents. The International Dialogues on AI Safety are also often held in middle powers.

   • Convenings and side events around the 2027 Swiss AI summit and its successors, aimed at concrete cooperation on AI risks.

   • Design and piloting of a cross-border AI incident notification mechanism.

2.5 Growing the field

Everything above needs talent, especially people who combine deep AI knowledge and local credibility. The talent pipeline also needs mid-career and senior people to provide mentorship and onboard new people.

1. 1.

   Fellowship programs. Intensive short-term fellowships have been very successful historically in training up people to do useful AI safety and policy work. Example projects:

   • The Talos Fellowship in Europe, and more middle power variants of various US and UK programs such as GovAI, IAPS, ERA, Pivotal and MATS. Programs in currently neglected countries like South Korea and Japan could be especially valuable.

   • A European research organization that supervises its own fellows in-house, on the GovAI model, so that graduates have somewhere to go.

   • Mid-career and senior programs for people already in ministries, parliaments, industry and the security establishment.

   • Fellowships bringing experienced middle power AI policy and technical talent home from the US, and staffer fellowships placing trained people into parliamentary and ministerial offices, modeled on TechCongress and Horizon.

2. 2.

   Convenings and networks. The people who need to work together on this are spread across ministries, labs, universities and industry, and mostly do not meet. Example projects:

   • IASEAI’s annual conference, held in Paris immediately before the 2025 AI Action Summit (at the OECD) and again in February 2026 (at UNESCO), bringing researchers, policymakers and civil society together. There is no equivalent in the Indo-Pacific.

   • The European Network for AI Safety, which maps and connects AI safety researchers and organizations across European cities. Most regions have no comparable connective tissue.

   • Recurring national roundtables that put a country’s officials and technical experts in one room, in capitals where no such forum exists.

3. 3.

   New organizations and shared spaces. Much of the work in this call will be done by organizations that do not yet exist. We want to make launching new organizations easier. Example projects:

   • An incubation program that helps founders in middle powers start the projects described in this call.⁷⁷ Project Tailwind is also calling for a general AI safety incubator.

   • A multi-organization research center with a shared office and visiting fellows.

Applicants

We think a wide range of people can meaningfully contribute to this space. Here are some stylized applicant profiles we would be particularly excited to hear from:

• AI safety professionals currently working in the US who are interested in moving back to their home countries to found new policy organizations.⁸⁸ Of course, the opportunity cost of leaving valuable jobs in the US may be very high; people should carefully think through the impact case of their current work and future potential in middle powers.

• Policymakers, advisors, and think tankers in middle powers looking to transition into working on AI. Many top AI policy people don’t have technical backgrounds. Upskilling is possible, and people with strong experience in other policy areas could be great fits.

• Founders and entrepreneurs who could partner with subject-matter experts to build new organizations and projects.

• Former national security experts will be well-placed to warn their government about the risks to national power if they miss out on the AI transition, and develop geopolitically realistic responses.

• Writers and communicators can help spread these ideas and grow the field.

• Data center and chip supply chain experts can provide much-needed advice to governments on how to negotiate with foreign companies.

• Talented generalists and early career people can also do great work, particularly when joining forces with more experienced co-founders.

This list is not exhaustive, and we encourage others to apply as well even if you don’t necessarily fit the above profiles.

Applications

We will assess applications based on the following:

• Fit. A strong track record, or other evidence that you and your team have the skills, networks, and judgment to carry out the work you propose.

• Theory of change. A clear and credible account of how the project would improve middle powers’ approach to transformative AI, and thereby make the AI transition safer.

• Understanding of the landscape. Strong familiarity with frontier AI development and its geopolitical implications.

• Ambition and focus. Proposals that are ambitious yet targeted: addressing an important, time-sensitive gap rather than trying to do everything.

• Reasoning transparency. Candor about what you know and how you know it, what you are uncertain about, and what could go wrong, including how you would mitigate the main risks. See here for what we mean by this.

• Budget. A well-considered budget doesn’t necessarily mean asking for less; rather, it means the funding requested should make sense given the proposed plan.

Our advisors

Start here

• The AI Divide — Sam Winter-Levy & Anton Leicht, Foreign Affairs, Feb 2026 (paywalled).

• How Can the Middle Powers Avoid Getting Trounced During the Intelligence Explosion? — Tom Davidson, Forethought, May 2026.

• What Can Middle Powers Do for Frontier AI Governance? — Markus Anderljung & Stephen Clare, Jul 2026.

• The Race Worth Winning: Middle Powers in the Age of Machine Intelligence — Dean Ball & Anton Leicht, FAI, Feb 2026.

• Europe 2031: What getting AI wrong means for us — Daan Juijn et al., Jun 2026.

• Anton Leicht on how middle powers avoid losing everything in a post-AI world — 80,000 Hours Podcast, Jul 2026.

Compute and data centers

• Building Liberal Compute — Simon Grimm, Mar 2026.

• 7 Learnings About Compute Strategy for Middle Powers — Philip Fox, Dec 2025.

• The Compute Coalition: How to Build the Future of AI in the Free World — Alasdair Phillips-Robins, Ryan Tawil & Sam Winter-Levy, Carnegie, Jun 2026.

• How Sovereign Is Sovereign Compute? A Review of 775 Non-U.S. Data Centers — Aris Richardson et al., arXiv, Jul 2025.

• What compute on European soil might buy Europe, and what it depends on — Jonathan Schmidt, Arslan Jurion & Toni Lorente, The Future Society, Aug 2026.

• AI Compute in Germany: Current Capacity, Future Needs — Philip Fox, Monika Schnitzer & Daniel Privitera, KIRA Center, Oct 2025.

Europe

• The World’s First Frontier AI Regulation Is Surprisingly Thoughtful — Miles Kodama, Sep 2025.

• Nineteen Thoughts on AI and Europe — Pieter Garicano & Simon Grimm, Silicon Continent, Jun 2026.

• Preparing Europe for Transformative AI — Arq Foundation, Jun 2026.

• Beware of Geeks Bearing Gifts: Building True EU Frontier AI Sovereignty — Nick Moës et al., The Future Society, Apr 2026.

• Pooling Europe’s Compute: Distributed Training for European Frontier AI — Rafael Andersson Lipcsey & Maximilian Negele, RAND Europe, Jul 2026.

• Delays to Frontier AI in the EU and UK — John Lidiard, Oleksandra Vereschak, Tom Gibbs & Markus Anderljung, GovAI, Jun 2026.

Background on transformative AI

For readers newer to the underlying arguments about where AI is heading and what could go wrong.

• International AI Safety Report — Yoshua Bengio et al., Feb 2026.

• AI 2027 — Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean, Apr 2025.

• Preparing for the Intelligence Explosion — William MacAskill & Fin Moorhouse, Forethought, Mar 2025.

• Extreme power concentration — Rose Hadshar, 80,000 Hours problem profile, Oct 2025.

Writers to follow

• Anton Leicht

• Simon Grimm

• Luis and Pieter Garicano

• Philip Fox

• Dylan Rogers