Two years ago, the consensus in the AI industry was that frontier model capability was the exclusive preserve of a small number of well-capitalised American laboratories. Building a competitive large language model required enormous GPU clusters, teams of hundreds of researchers, and training runs that cost tens or hundreds of millions of dollars. The barrier to entry appeared structural — a function of compute availability and capital concentration that would keep meaningful AI development within a narrow corridor for years.

That consensus has been disrupted. Not gradually, but in distinct shocks that have forced a reassessment of who can build competitive AI, at what cost, and under whose control.

To understand the stakes, start with the technical distinction. A closed model is accessible only through a remote API. You send a request to a server owned by the provider — OpenAI, Anthropic, Google — the model runs on their infrastructure, and you receive a response. You never see the model's weights: the billions of numerical parameters that encode its capabilities. You have no access to its training data, its fine-tuning methodology, or its safety constraints. You trust the provider on all of these. And you pay per token.

An open-weights model is different in kind. The weights are downloadable. You can run them on your own infrastructure — including, for smaller models, commodity hardware. You can fine-tune them on your own data, modify their behaviour, integrate them into applications you fully control. Nothing leaves your systems. You are not dependent on the provider's uptime, pricing decisions, or policy changes for continued access.

This distinction has geopolitical weight. When a European hospital, a national government agency, or a financial institution uses a closed model to process sensitive data, that data passes through American servers, subject to American law — including the CLOUD Act, which grants US authorities access to data held by American companies regardless of where it is physically stored. This is not a theoretical risk; it is a policy constraint that procurement offices at regulated institutions are already grappling with.

Into this gap stepped Mistral AI. Founded in 2023 by researchers from Google DeepMind and Meta, the Paris-based startup has demonstrated that competitive model performance does not require American-scale resources. Mistral 7B outperformed models twice its size on standard benchmarks. Mixtral 8x7B competed with GPT-3.5 across a wide range of tasks. All with fully open weights, a fraction of the headcount, and infrastructure that would fit in a medium-sized data centre.

The more dramatic disruption came from DeepSeek. In January 2025, DeepSeek R1 demonstrated reasoning performance competitive with GPT-4, while claiming training costs an order of magnitude lower than US frontier labs. The implications were significant: if a Chinese laboratory, operating under US chip export controls and with limited access to the latest Nvidia hardware, can produce a model of this capability at this efficiency — the moat that Silicon Valley built on compute access is narrower than it appeared.

The open source question in AI is, at its root, a question about who controls the infrastructure of cognition. The applications we use daily to access information, assist our decisions, and process our communications are increasingly mediated by AI systems. If those systems are black boxes hosted on foreign servers, operating under terms we did not negotiate and cannot audit, we are not merely in a state of economic dependence. We are allowing our cognitive environment to be shaped by systems over which we have no meaningful control.

Owning the weights of the AI you deploy is not just a technical preference. It is, increasingly, a prerequisite for institutional autonomy — in healthcare, in education, in government, and in the organisations that produce knowledge. The open source movement in AI is the mechanism by which that autonomy remains achievable.