Artificial Intelligence (AI) Concentration and Governance.

 

As AI capabilities have increased rapidly, investment has also risen, with both compute and capital increasingly concentrated in a handful of countries. In 2024, US private AI investment reached US$109 billion, against China’s US$9.3 billion and the United Kingdom’s US$4.5 billion, and the US accounted for 40 notable frontier models against 15 from China and three from Europe. At the same time, the cost of running these models has dropped dramatically, as shown in Figure 5.5. Inference for GPT-3.5-equivalent performance fell from about US$20 per million tokens in November 2022 to US$0.07 by October 2024, a 280-fold drop in less than two years. Cheap inference combined with expensive training favours a small set of firms and states: those with the capital to build frontier models, and the customer base to spread the cost of inference across millions of users

This concentration carries a direct energy cost. The International Energy Agency projects global data centre electricity consumption rising to around 945 terawatt-hours by 2030, more than double 2024 levels and exceeding Japan’s current total demand. The strain is already visible at country level. Data centres in Ireland accounted for 22 per cent of all metered electricity in 2024, a fourfold increase from five per cent in 2015. At the firm level, the two largest hyperscale data-centre operators have reported sharp emissions growth driven by AI infrastructure: one disclosed that its total emissions across all categories rose 29.1 per cent from a 2020 baseline, and the other reported total greenhouse-gas emissions for 2023 of 14.3 million tonnes of carbon dioxide equivalent, a 48 per cent increase from 2019. Countries that host frontier-model training enjoy the upside of the AI economy, while those that only import AI services bear the rising cost of inference and the opportunity cost of foregone investment in their own energy infrastructure. Compute is no longer just a technical input; it is becoming a geopolitical category, a strategic resource over which states are willing to compete in the same way they once competed over oil refining capacity or semiconductor fabrication.

 The international response to AI’s emerging warfare applications has not yet settled into a coherent legal framework.Governance has emerged instead as a series of overlapping summits, declarations, voluntary commitments, regional regulations, and United Nations Security Council resolutions, each addressing a different slice of the problem and each with a different set of participating states. The trajectory is one of accelerating activity rather than coordinated control, and the gap between deployment and rule-making has widened, not closed, since 2023.

The most visible track is the series of AI Safety Summits. The first, held at Bletchley Park, England, in November 2023, produced the Bletchley Declaration, endorsed by 28 countries and the European Union (EU). Six months later, the Seoul Summit of May 2024 added detail, producing the Seoul Declaration, a Statement of Intent on AI Safety Science, the Frontier AI Safety Commitments from 16 leading developers, and a Ministerial Statement endorsed by 27 countries plus the EU. However, this trajectory of increasing cooperation broke at the Paris AI Action Summit in February 2025. The closing Statement on Inclusive and Sustainable AI for People and the Planet was signed by 61 countries, but the United States and the United Kingdom declined to sign, marking the first major split between the original Bletchley signatories. Figure 5.6 shows the number of major AI governance instruments that countries worldwide are part of. There were 118 countries that were not parties to any of the seven major governance initiatives.

At the global level, the General Assembly of the United Nations took its first major step on AI governance through a resolution in August 2025, which set up an Independent International Scientific Panel onAI of 40 members and a Global Dialogue on AI Governance, both anchored in the UN’s Pact for the Future and the Global Digital Compact. The first session of the Global Dialogue ran in September 2025; the second is scheduled for the International Telecommunication Union’s AI for Good Summit in Geneva in 2026. The United Nations also adopted a draft resolution on lethal autonomous weapons in November 2025, and the General Assembly plenary endorsed it a month later. The same plenary adopted a separateresolution on the risks of integrating AI into nuclear command,control, and communications systems, the first dedicated General Assembly action on AI in nuclear command and control. Together, the two resolutions establish a near-universal political baseline for the principle that humans should retain meaningful control over lethal force and over nuclear release, even though both resolutions remain non-binding.

At the regional level, the Council of Europe Framework Convention on AI and Human Rights, Democracy and the Ruleof Law opened for signature in Vilnius in September 2024, with the US, the UK, the EU, and several other Council member states among the initial signatories.