Copyright Project Syndicate
WASHINGTON, DC – An international coalition of AI labs and cloud providers just did something refreshingly practical: they pooled their compute resources to make Apertus, a Swiss-built open-source large language model (LLM), freely accessible to users around the world. The queries Apertus receives might be served by Amazon Web Services in Switzerland, Exoscale in Austria, AI Singapore, Cudo Compute in Norway, the Swiss National Supercomputing Centre, or Australia’s National Computational Infrastructure. Could this project point the way forward for international cooperation? In the 20th century, international cooperation became practically synonymous with the rules-based multilateral order, underpinned by treaty-based institutions such as the United Nations, the World Bank, and the World Trade Organization. But great‑power rivalries and structural inequities have eroded the functioning of these institutions, entrenching paralysis and facilitating coercion of the weak by the strong. Development finance and humanitarian aid are declining as basic principles like compromise, reciprocity, and the pursuit of mutually beneficial outcomes are called into question. The retreat from cooperation by national governments has increased the space for other actors – including cities, firms, philanthropies, and standards bodies – to shape outcomes. In the AI sector, a handful of private companies in Shenzhen and Silicon Valley are racing to consolidate their dominance over the infrastructure and operating systems that will form the foundations of tomorrow’s economy. If these firms are allowed to succeed unchecked, virtually everyone else will be left to choose between dependency and irrelevance. Governments and others working in the public interest will not only be highly vulnerable to geopolitical bullying and vendor lock-in; they will also have few options for capturing and redistributing AI’s benefits, or for managing the technology’s negative environmental and social externalities. But as the coalition behind Apertus showed, a new kind of international cooperation is possible, grounded not in painstaking negotiations and intricate treaties, but in shared infrastructure for problem-solving. Regardless of which AI scenario unfolds in the coming years – technological plateau, slow diffusion, artificial general intelligence, or a collapsing bubble – middle powers’ best chance of keeping pace with the United States and China, and increasing their autonomy and resilience, lies in collaboration. Improving the distribution of AI products is essential. To this end, middle powers, and their AI labs and firms, should scale up initiatives like the Public AI Inference Utility, the nonprofit responsible for the provision of global, web-based access to Apertus and other open-source models. But these countries will also have to close the capability gap with frontier models like GPT-5 or DeepSeek-V3.1 – and this will require bolder action. Only by coordinating energy, compute, data pipelines, and talent can middle powers co-develop a world-class AI stack. There is some precedent for this type of cooperation. In the 1970s, European governments pooled their capital and talent, and coordinated their industrial policies, to create an aircraft manufacturer capable of competing with America’s Boeing. An “Airbus for AI” strategy would entail the creation of an international, public-private frontier lab dedicated to pre-training a family of open-source base models and making them freely available as utility-grade infrastructure. The result would not be another monolithic AI titan, but rather open infrastructure on which many actors could build. This approach would drive innovation by allowing participating national labs, universities, and firms near the frontier (such as Mistral and Cohere) to reallocate up to 70% of their model pre-training funding to post-training (specialized or inference models), distribution, and demand-driven use cases. Moreover, it would enable governments and firms to take control of the AI ecosystems on which they increasingly rely, rather than being held hostage by geopolitical uncertainty and corporate decisions, including those that lead to “enshittification.” But the potential benefits extend even further. This open infrastructure – and the data pipelines on which it is built – could be repurposed to meet other shared challenges, such as lowering the transaction costs of global trade in green energy or developing an international collective-bargaining framework for gig workers. To showcase the full potential of this new collaborative framework, middle powers should target problems for which mature data ecosystems and technologies already exist; participants’ self-interest outweighs the transaction costs of cooperation; and the value of shared action is apparent to citizens and political leaders. In a few years, when the current AI innovation and capital cycle has run its course, middle powers can either be lamenting the demise of the rules‑based order and watching AI giants ossify geopolitical fault lines, or they can be reaping the benefits of innovative new frameworks for cooperation. The case for public AI is clear.
