The pitch is nostalgic on purpose. Current AI, a nonprofit founded last year in Paris, has announced a project it calls the World Wide Web of AI, an attempt to build an open, free layer of AI infrastructure in the spirit of the early internet. The project’s backers include several Turing Award winners and leading European research institutions, according to the announcement.
The components are threefold. Open-weight models that anyone can download and inspect. Shared training datasets, assembled and maintained in public. And public inference application programming interfaces, so that researchers and small developers can run models without owning the hardware. Together, the project’s organizers argue, these pieces form the equivalent of the web’s early protocols: a common substrate that no single company controls.
The philosophy is explicit. The early internet was built on open standards, TCP/IP, HTTP, HTML, and that openness is what allowed anyone to build on it. The organizers argue that AI is becoming the next general-purpose technology and that allowing a handful of companies to control its most important layers would repeat the mistakes of the past. The project’s name is the argument: a World Wide Web of AI should work the way the web works, with open protocols underneath and innovation on top.
Current AI is a natural home for the effort. The organization was created to fund and coordinate public-interest AI work, with support from European governments and philanthropists, and it has said its mission is to make sure AI’s benefits are not captured entirely by private firms. The new project gives that mission a concrete architecture, and its European anchoring is deliberate. The region has spent years regulating AI while worrying about its dependence on American and Chinese platforms, and a European-anchored, open infrastructure is a way to turn that anxiety into an asset.
The hardest parts are the ones the announcement describes last. Open-weight models are plentiful, since Meta’s Llama and others already publish weights, but models are only as good as the data they are trained on, and assembling shared datasets that are legal, clean, and large enough is slow, expensive work. Public inference application programming interfaces cost money to run; accelerators do not become cheaper because the operator is a nonprofit. The project will need sustained funding to keep the free layer actually free.
Skeptics note that open-weight models already exist and that the missing ingredient was never willingness but resources. The big labs spend billions on data, compute, and talent, and a nonprofit that wants to match them needs a comparable budget or a different theory of how to win. The organizers say they are not trying to match the frontier labs model-for-model. They are trying to build the shared layer underneath, data, benchmarks, and access, that makes the field more competitive, the way open standards once made the internet more competitive.
The project also gives European researchers a reason to stay in Europe instead of moving to the labs where the money is. One of the region’s quiet problems is that its best AI talent trains at European universities and then joins American companies; an open infrastructure that researchers can build on, and even govern, changes that calculation. The involvement of Turing Award winners and major research institutions is meant to signal that this is a scientific project, not a political stunt.
Governance will be the test. An open project with models, data, and compute needs rules about who can use it, how changes are made, and what happens when someone abuses it. The early internet solved these problems badly at first and painfully later, and the AI project will have to solve them in public, with regulators watching and with the lessons of the past two decades of platform governance available. The organizers have said they intend to build governance into the project from the start, though the details have not been published.
The project will also face competition from the private sector’s generosity. The big labs give away substantial free tiers of their own, and open-weight models from major companies are available to anyone. The difference is control: the free tiers belong to companies that can change the terms, while the project’s promise is that its layer belongs to everyone. Whether that promise holds will be decided by the funding, the governance, and the users.
The project’s success will be measured in adoption, not announcements. If researchers build on the shared models and developers call the public application programming interfaces, the World Wide Web of AI will become what its name promises. If it stays a statement of intent, it will join a long list of open projects that wanted to change the industry and did not. For now, the founders have done the easy part: they have put the idea in the open, where anyone can argue with it.
