1. Introduction

This report explores how ‘AI middle powers’ — states with significant economic and regulatory capabilities that depend technologically on foreign-owned frontier artificial intelligence (AI) firms — can utilise public procurement as a strategic governance instrument. Leading AI development is concentrated in a small number of private firms headquartered in the United States and China. As middle powers, including the United Kingdom, France, Germany, Canada, South Korea, and Japan, integrate these systems into critical infrastructure, they face operational and strategic dependencies. The paper assesses whether procurement frameworks can transform from transactional purchasing agreements into mechanisms that secure technical and legal guarantees for state buyers.

2. Methodology

The research uses a mixed-methods approach, synthesising more than 150 sources and 15 semi-structured interviews conducted between February and June 2026. The literature review spans AI governance, industry analyses of capabilities, and procurement law. For this report, ‘frontier AI’ is defined relative to the technological state of the art at a given time, referring to general-purpose models at the leading edge of available capabilities. The analysis is forward-looking, noting that while current public sector spending is largely directed at established analytics and machine learning, the strategic significance of frontier AI is growing rapidly.

3. The concentration of leading AI capabilities

Frontier AI capabilities are sharply concentrated geographically and within the private sector. As of mid-2025, the United States accounted for approximately 75 per cent of global AI compute performance, while China hosted 15 per cent. The European Union collectively represented less than 5 per cent. Private sector companies now control 80 per cent of AI supercomputer performance, up from 40 per cent in 2019. US-developed models like GPT-5.4 Pro and Gemini 3.1 Pro Preview lead on graduate-level science reasoning benchmarks like GPQA Diamond, often surpassing specialist human knowledge.

This concentration creates three primary risks for state buyers: vendor lock-in and the threat of sudden service withdrawal; limited strategic visibility into model development; and data sovereignty vulnerabilities. A notable example occurred in June 2026, when a US export control directive suspended all foreign access to Anthropic’s Fable 5 and Mythos 5 models. Such episodes illustrate how a vendor’s home government can abruptly sever access to frontier capabilities, leaving dependent buyers with limited recourse.

4. What can states govern via public procurement?

Public procurement is a transaction-level instrument that can complement statutory regulation. In the EU, procurement represents roughly 15 per cent of gross domestic product; across OECD countries, it accounts for 20–40 per cent of total government expenditure. Large public sector contracts provide commercial leverage through prestige and regulatory validation. For example, Palantir Technologies expanded commercially after validating its technology through years of US government use. However, procurement cannot override the national security laws of a vendor’s home state or eliminate the persistent information asymmetry between state bodies and technology companies.

The report proposes a framework of six contractual guarantees organised around three axes: seeing the system, shaping its development, and securing continuity. These include: model access for evaluation (seeking white-box access to weights and activations); information sharing on R&D and scaling plans; funding for reliability research; the adoption of state-of-the-art risk management frameworks; ‘sunset’ termination clauses requiring extended notice periods; and provisions to build domestic and allied capability.

5. Policy recommendation: exploring shared procurement standards across AI middle powers

The report recommends that middle powers coordinate their frontier AI procurement standards. Coordination increases bargaining leverage by pooling demand and establishes common standards that may become widely diffused through the market, similar to the ‘Brussels Effect’. It also allows states to share scarce evaluation infrastructure and monitor AI-related incidents more effectively. Historical precedents in defence, such as the F-35 Joint Strike Fighter and Eurofighter Typhoon programmes, demonstrate that contractual protections are only valuable when governments retain the institutional capacity to monitor compliance.

Short-term actions (to be initiated within 12 months) include performing a shared assessment of procurement risks, identifying suitable institutional arrangements for coordination (such as the G7 or NATO), and exploring demand-pooling opportunities. Medium-to-long-term actions (1–5 years) include investing in independent infrastructure like domestic or allied compute, strengthening evaluation capacity within state agencies, and establishing cross-border incident monitoring systems.

6. Conclusion and areas for future research

The paper concludes that while public procurement is a useful governance vehicle, it is best understood as a complement to regulation rather than a substitute for it. The effectiveness of procurement as a governance lever will depend on the precision with which middle powers build standardised contractual language and the shared institutions required for AI auditing and evaluation.