What can middle powers do for frontier AI governance?

Markus Anderljung & Stephen Clare, Jul 07, 2026

Premise

• Common view: only the US and China matter for AI — frontier models come from a handful of American companies, China is the only challenger, and Washington increasingly controls access. The authors argue this is a mistake.
• “Middle AI powers” = countries with strong political institutions and economic power but no frontier AI companies — roughly the G20 minus the US and China, plus AI-specific players like Taiwan, the Netherlands, Norway, and the UAE.
• The piece makes six claims about what these countries can do.

1. Middle powers have already shaped AI development

• Until at least 2026, middle powers arguably influenced frontier company behaviour more than the US government did.
• AISI model: the UK’s AI Security Institute became the world’s leading evaluator of dangerous capabilities within two years; copied (imperfectly) by the US, Singapore, Japan, Korea, Canada, and the EU.
• Summit series: Bletchley, Seoul, and New Delhi produced concrete company commitments — first to test models for dangerous capabilities, then to publish safety frameworks — which made it easier for US state legislators and EU regulators to legislate similar provisions.
• Regulation: the EU AI Act’s general-purpose AI provisions plus its Code of Practice may be among the most important developments of the past two years; the authors bet Chinese frontier developers will publish safety frameworks and model specs sooner as a result.
• The policy space outside the US is less crowded, so individuals can have outsized impact: South Korea may have 1000x less leverage than the US, but it may be more than 1000x easier for the right person to improve its AI policy.

2. Middle powers have assets that give them leverage

• On classic power indicators they do well: collectively more GDP than the US and China combined; individually ahead on military spending, R&D spending, and total investment.
• On AI-specific metrics the US dominates (domestic high-end compute, private AI investment, frontier-lab valuations, chip design) — but middle powers aren’t out of the game.
• They generate lots of electricity, which could support a much more aggressive datacentre buildout.
• They hold critical chip supply chain links: ASML (Netherlands) ~80% of lithography and 100% of EUV; Samsung and SK Hynix (Korea) ~three-quarters of HBM; ASE (Taiwan) the largest OSAT firm; TSMC leading fabrication.
• Ukraine’s drone-warfare leadership shows a middle power leveraging niche expertise for support from Europe and the Gulf.

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

• The US has ~25% of global GDP but hosts ~75% of world compute performance. Hosting datacentres confers geopolitical power: the host can control, restrict, or monitor foreign access to models running on those chips.
• AI companies want to diversify away from the US (political toxicity risk, potential US restrictions on foreign access). The binding constraint is build speed, so winners will have spare grid capacity or allow behind-the-meter generation.
• Climate bargain: behind-the-meter power today mostly means gas, but states could trade short-term gas permits for commitments to grid renewal and an eventual switch to renewables.
• Beyond compute, middle powers could:
  • Build domestic datacentre infrastructure companies (hyperscalers increasingly use specialists like Nebius and Nscale).
  • Poach frontier talent to found domestic champions — top researchers leaving labs often reach $1bn+ valuations quickly.
  • Deepen existing strengths (Dutch chip equipment, Japanese semiconductor materials) and push AI into exposed sectors (European deep tech and pharma, UK finance and professional services).
  • Embrace creative destruction: much of AI’s value accrues to adopters (e.g. AI-enabled pharmaceuticals), so let laggard firms be superseded.
  • Invest in the frontier directly: Japan and the Gulf are overrepresented among OpenAI/Anthropic investors; SoftBank alone holds ~13% of OpenAI.

4. Middle power influence doesn’t need hard power

• Superpowers usually have escalation dominance, so leverage is most useful held in reserve as a deterrent rather than used coercively.
• Market access is influence: only ~20% of ChatGPT users are in the US, so conditions on market entry shape company behaviour globally (the Brussels Effect) — e.g. pre-deployment evaluations run for one market inform deployment everywhere.
• Middle powers can produce public goods: published AISI evaluations, funded safety research, and accessible risk-management tools that lower mitigation costs worldwide.
• Soft power matters: South Korea’s biggest governance impact came not from its HBM leverage but from convening the 2024 Seoul Summit and securing frontier safety framework commitments.

5. Middle powers can directly mitigate some risks

• By controlling how AI is used within their borders, they can reduce cyber and bio misuse, surveillance, authoritarian applications, and democratic backsliding.
• Four risk categories by which actors must be involved:
  • Great-power problems: only Washington and Beijing move the needle (e.g. preventing egregiously misaligned frontier systems); middle powers help only indirectly.
  • Coverage problems: solutions scale with number of countries acting (e.g. cyber resilience and infrastructure hardening).
  • Weakest-link problems: one lax jurisdiction creates system-wide vulnerability, so near-universal participation is needed (e.g. AI-enabled biorisk monitoring and enforcement).
  • Best-shot problems: one success suffices for everyone (e.g. an interpretability breakthrough).
• On coverage, weakest-link, and best-shot problems, middle powers “can just do things.”

6. Middle powers can work together

• The old assumption — the US writes global rules, allies adopt them in exchange for compute access — now looks less likely; US allies may need to proceed on their own.
• Options for coordination:
  • Unilateral recognition of or convergence on standards: accept EU GPAI compliance as sufficient for market entry, or design requirements that track the EU’s substance.
  • Invest in the Summit series: make the 2027 Geneva AI Summit a success with a new round of frontier-company commitments, especially transparency about how companies use internal models to accelerate their own R&D and manage related risks.
  • Build fora that don’t depend on the US (or China) for trade, national security, and frontier AI coordination.
  • Pool bargaining power (hardest, highest impact): align export controls, jointly secure early model access, harmonise procurement, form a frontier-model buyers’ club, or co-invest in near-frontier developers.

What this means for AI governance work

• Make middle-power policymakers feel AI is a big deal — belief, not proposals, is usually the blocker.
• Build expertise, networks, and talent; the sparse policy field means one capable person can make a big difference.
• Make the Geneva 2027 AI Summit great, pushing for new company commitments.
• Support middle-power coordination: flesh out EU GPAI recognition, joint early model access (mirroring the recent US executive order), and a buyers’ club.
• Help middle powers seize AI opportunities: accelerate responsible adoption; race to be first to approve and commercialise what AI invents.
• Build societal resilience via coverage problems and public goods.
• Engage sovereign wealth funds and large investors for direct stakes in the frontier.
• Spread useful norms and start policy thinking on near-empty questions like legal personhood for AI.
• Caveat: the authors remain uncertain how much middle-power action will shape the trajectory, and challenges will grow — but middle powers can do more than spectate.