Access to compute, which encompasses the specialised semiconductors, data centres, software and operational capacity that underpin AI systems, is concentrated at nearly every layer of the AI stack, and this concentration is producing system risks and inequalities, resulting in a new form of digital divide. Most countries are unable to build or govern compute capacity to support public priorities in AI development, and must rely on capacity that is allocated according to commercial priorities, with little leverage over pricing, availability, or allocation of compute infrastructure. National responses have not closed this gap, despite the emergence of ‘sovereign’ AI strategies. This brief argues that equitable compute cooperation is rooted in three coordination problems which require international cooperation to solve: First, current approaches prioritise autonomy and capacity, without addressing usability or control. Second, the lack of interoperability, governance frameworks and shared standards leads to overspend on, and underutilisation of, compute. Third, there is a lack of standardised information to evaluate and forecast compute supply and demand across countries. The brief argues that the G20 is well placed to address these concerns through three lines of action: First, developing a common evidence base, working with international organisations such as the OECD, that measures both compute availability within a jurisdiction and the capacity subject to public allocation rights, beginning with voluntary self-reporting. Second, coordinating public financing and procurement so that public contributions across the supply chain include public access and benefit considerations, including through benefit-sharing principles, a toolkit of model contractual provisions, and the integration of public compute into multilateral development bank appraisal and financing of digital infrastructure. Third, establishing voluntary arrangements for pooled access and interoperability between public compute programmes, so that public-interest workloads can draw on capacity across participating jurisdictions and resilience is improved through portability.
