A farmer in Rajasthan photographs a dying plant with her phone. She wants to know what to do, but she doesn’t speak English, and the artificial intelligence around her barely understands her language. This isn’t a corner case: over half the world’s population remains excluded from the benefits of LLMs because low-resource languages are almost invisible in training datasets. Current AI, a nonprofit backed by France and DeepMind with a $400 million fund, thinks this is exactly the problem worth solving — and it does so with a radical thesis: artificial intelligence needs a “public option,” open and free, much like the BBC or public postal services.

Behind this rhetoric lies a precise architectural choice. To work in rural India, in areas with intermittent connectivity or where data is too sensitive to leave the device, such a public option must be able to run inference locally, on modest hardware. It’s not just about training multilingual models: what’s at stake is the ability to perform inference on smartphones, edge devices, or small on-premise servers, without relying on hyperscaler APIs. That’s why the game shifts toward aggressive quantization, lean parameter counts, and computational efficiency — everything that lets a model run with limited VRAM, far from data centers.

This move aligns with a broader tension. The current AI oligopoly — a handful of U.S. companies offering models via the cloud — breeds technological dependency and linguistic uniformity. European governments, already sensitive to digital sovereignty, see projects like Current AI as a way to build alternative stacks where training and inference happen on controlled infrastructure, respecting GDPR without negotiating with non-EU providers. France, in particular, has funded similar efforts with Mistral and the Jean Zay supercomputer, and now plants a flag on genuinely accessible AI, nurturing an ecosystem of open models that public organizations can self-host.

Who wins from this pivot? First, the language communities so far ignored, because a model optimized for local inference can become a real agricultural or medical tool, not an academic abstraction. Then, enterprises and public administrations that want to avoid vendor lock-in: having public models that can run on-premise lowers TCO in the mid-term and returns data control. Who loses? The cloud providers selling API tokens by the drink: if a billion people start using on-device LLMs without API calls, the margin on those volumes shrinks dramatically. Even DeepMind, seemingly in conflict with its own business model, appears to understand this — participating in building the public option means securing a layer that could one day become an indispensable infrastructure on which to graft premium services.

Structurally, the announcement signals that AI’s next frontier is not parameter scale but capillarity. We need models that work offline, in low-resource languages, on consumer processors. It’s a challenge that intersects hardware and software: 4-bit quantization, optimized attention, runtimes like ExecuTorch for mobile. For those evaluating on-premise deployment, there are trade-offs between response quality and energy consumption, between the maintainability of a self-hosted model and the convenience of a managed API. But the essence of Current AI is that certain needs — a farmer’s language, a patient’s privacy, the resilience of an emergency service — cannot be met by a prompt sent to a data center thousands of miles away.