A CERN for AI-assisted science?

The article proposes that science needs publicly funded frontier AI systems, including simple conversational and agentic interfaces with substantial public compute behind them, with free baseline access for researchers and competitive applications for multi-agent systems and large-scale compute for projects needing more substantial resources. Relative to the earlier use of the CERN analogy by the European Commission in describing its AI infrastructure plans, the author offers a more specific version centred on AI as public infrastructure for scientific research, explicitly separating free baseline access from competitive allocation of large compute. This is an opinion and policy proposal; its argument rests on four problems the author enumerates (sensitive or classified data, the sensitivity of unpublished research, misalignment between private-company incentives and the scientific community, and highly unequal access to the strongest AI systems) plus public facts such as existing Canadian and EU investments and cooperation agreements.

The article identifies four problems with leaving frontier AI access concentrated in a handful of private companies: research can involve sensitive, proprietary or even classified data requiring strong confidentiality and clear accountability; unpublished research itself is sensitive, since prompts, uploaded documents and agent interactions can contain new ideas and partial results that must not become available to systems that might reproduce or build upon them; private-company incentives differ from those of the scientific community, since competition encourages rapid demonstrations of new capabilities and spectacular results while science also needs verification, attribution, careful exposition and human understanding; and access to the strongest AI systems is currently highly unequal, risking a two-tier scientific system. It shifts the discussion of AI in science from capability demonstrations to institutions and governance, treating confidentiality, data reuse, incentive misalignment and access inequality as structural problems that public infrastructure should address. It draws on the Navier-Stokes result being produced by an internal OpenAI model the company describes as significantly more capable than its publicly available frontier model, on several other recent mathematical advances obtained using unreleased models, and on Julia Stadlmann's experience of two years of work being surpassed by AI-assisted efforts within days to illustrate the effect of speed on early-career researchers.

The article argues the timing may be unusually favourable: the EU and Canada are rapidly deepening their strategic relationship, both sides are making major public investments in AI and computing infrastructure, and they have explicitly agreed to explore cooperation on fundamental AI research, agentic systems for scientific discovery, access to advanced AI infrastructure, and co-development of advanced AI models for the public good. It connects the institutional proposal to a concrete policy window, pointing to arrangements such as Canada's commitment of C$2 billion over five years to sovereign AI compute, including up to C$1 billion for public supercomputing infrastructure, and the EU's development of RAISE and a network of AI Factories plus a call for up to seven AI Gigafactories backed by up to €10 billion in public funding and expected to unlock at least €20 billion in additional private investment. It relies on publicly stated policy commitments and cooperation agreements cited in the text, describing the policy environment rather than reporting empirical research results.

The article stresses that such a system cannot be merely another supercomputer centre: the user-facing layer is essential, with every researcher able to access frontier AI models through a simple chatbox without becoming an expert in model deployment, cloud infrastructure or GPU computing, while more ambitious projects could decompose a difficult research programme into many interacting subproblems explored semi-autonomously by coordinated AI agents, with researchers directing overall strategy and interpreting results. It places usability and human-machine division of labour at the centre of the design of public AI research infrastructure, and proposes competitive allocation of large compute and multi-agent systems, citing the Digital Research Alliance of Canada's peer-reviewed Resource Allocation Competition as an existing mechanism. This is a design proposal and analogy; the author also lists open questions, including whether publicly funded models could realistically remain near the frontier, whether to train from scratch or combine public infrastructure with commercial and open-weight models, how such an institution should be governed, what scale of funding would be required, how access to very large compute should be allocated, and how to prevent the public infrastructure itself from becoming bureaucratic, technologically stagnant or captured by particular national or commercial interests.

The article addresses research governance and infrastructure policy: it argues for free baseline access for every researcher to a world-class conversational and agentic AI research environment, with projects requiring exceptional resources applying competitively for large compute allocations and sophisticated multi-agent systems, and it is meant for partners willing to contribute resources and subscribe to common standards of scientific governance, confidentiality, access and accountability. The author states that his motivation is not primarily to save academics the cost of AI subscriptions but the long-term independence of research, and he wants the initiative open to other research partners.

The author himself lists several open questions: whether publicly funded models could realistically remain near the frontier; whether to train models from scratch or combine public infrastructure with commercial and open-weight models; how such an institution should be governed; what scale of funding would be required; how access to very large compute resources should be allocated; and how to prevent the public infrastructure itself from becoming bureaucratic, technologically stagnant or captured by particular national or commercial interests. In addition, this is a blog opinion piece without experimental data or quantitative evaluation, so readers seeking concrete feasibility evidence would need to follow subsequent policy discussion and pilot arrangements.