Four hundred million dollars. That is the total committed to Current AI, a nonprofit barely eighteen months old, whose stated goal is to do for artificial intelligence what the World Wide Web did for information: hand it to the public as a commons instead of selling it back as a product. The French government put in the first 100 million. Ford, MacArthur, DeepMind, and Salesforce followed. Not as investors hunting returns, the organization stresses, but as funders underwriting infrastructure.
The pitch sounds almost quaint in an industry where frontier labs burn billions on compute. Yet the people running Current AI are not naive. Its CEO, Ayah Bdeir, spent years leading AI strategy at Mozilla and previously founded littleBits, a STEM education company. Current AI was created in February 2025 by Martin Tisne, a veteran of philanthropic technology funding. The pedigree matters because the plan is less about building one competing model and more about stitching together a public stack that anyone can use, audit, and extend.
From Mozilla to a Global AI Commons
Bdeir frames the mission in blunt terms. If AI is going to reshape every part of life, she argues, there has to be a public alternative, open the way the early web was open. The comparison is deliberate. The web won not by outspending closed rivals like AOL and CompuServe but by being a standard everyone could build on. Current AI is betting the same dynamic can play out in AI, except this time the data sovereignty questions get designed in from the start rather than bolted on after the fact.
The funding model is a public-private partnership, pulling in governments, companies, and philanthropies. That 400 million is committed, not all deployed, and the organization moves in grants rather than acquisitions. Last month it allocated 3.2 million dollars across four organizations working in Kenya, Lebanon, and the Brazilian Amazon.
The First Real Artifacts: Offline Devices and an Open Chatbot
Talking about infrastructure is cheap. Current AI's early output is more concrete. In India, it partnered with Bhashini, the government's language AI division, to produce Suno Sutra, a pocket-sized device that runs AI in 22 Indian languages with no internet connection. It is open-sourced so developer communities can build on top of it.
In Geneva, at the AI for Good Summit, the organization launched Alpha Chat, an open-source chatbot assembled in seven weeks by a coalition of ten groups including Hugging Face, Mozilla, and MIT Media Lab. Each contributor brought one layer of the stack, from the language model to safety tooling to compute. Separately, Current AI struck a deal with Sakana AI, a Tokyo startup focused on what it calls Sovereign AI, to build a shared open stack serving Japanese language and culture alongside communities across the Global South that dominant systems ignore.
Why "Open" Here Is Really a Data Sovereignty Fight
The deeper argument is not about software licenses. It is about who controls the training data and the consent around it. Bdeir points out that for Indigenous languages, missionary Bible translations have already become training data before the communities involved set any rules. Big tech, she says, builds multilingual models to expand its market "regardless of consent or context."
Current AI's answer is structural. It stores models and data locally, brings in community experts before anything is built, and writes consent protocols into the pipeline so a community can halt the process at any point. None of the grantees have fully solved data ownership, and Bdeir admits that is the point: the work bakes the question into every project rather than accepting the default where a distant company decides for everyone.
| Dimension | Current AI (Public Commons) | Proprietary Frontier Labs |
|---|---|---|
| Funding source | Governments, foundations, philanthropies | Venture capital, corporate revenue |
| Data ownership | Communities retain control locally | Held by the company, used to train |
| Language coverage | Targets ignored languages and dialects | Optimized for large commercial markets |
| Monetization | None, open infrastructure | Subscriptions, API pricing, ads |
| Deployment | Offline, low-cost devices | Cloud, connected, scaled to revenue |
The contrast is not that one model is morally superior. It is that the two serve different populations. A closed lab has no incentive to ship a free offline tool to a farmer with intermittent connectivity, because that user generates no revenue. Current AI exists precisely for the people outside the paying funnel.
What This Means
The 400 million figure looks small next to the hundreds of billions the largest labs will spend on compute this decade. That asymmetry is not a weakness in Current AI's plan, it is the entire logic. The web did not beat proprietary networks by matching their budgets. It won because openness created a coordination point no single company could own. Current AI's real asset is not money but the coalition it assembles: France seeding the endowment, India supplying deployment scale, Japan contributing Sovereign AI research, Kenya and Brazil grounding the work in communities the mainstream ignores. That is geopolitical infrastructure dressed as charity, and it is the kind of multipolar AI order that no individual company can replicate.
There is a hard risk beneath the optimism. Public-interest technology has a long graveyard of well-funded initiatives that collapsed when political winds shifted. A 3.2 million dollar grant split across four organizations is a rounding error against the scale of the problem, and the public-private model is only as durable as the governments backing it. If French priorities change after the next election, the seed funding wobbles. The consortium model also moves slowly precisely because consent and community input take time that venture-backed competitors do not spend.
For builders, the opening is in the gaps the giants will not serve. Offline, sub-fifty-dollar AI devices for the roughly 2.6 billion people with unreliable internet may be the most economically important deployment of the next decade, even though it makes no money directly. And the "consent protocol" Bdeir describes is itself an emerging product category: tooling that lets a community audit, approve, or halt how its data and language enter a model. Founders who build on open stacks like Alpha Chat and the Sakana shared layer get capability without lock-in, while the closed labs keep raising the price of leaving.