A small, federated compute co-op called co/core — and the wider experiment it's a part of.
On the evening of June 12th, a model that hundreds of millions of people could use one day couldn't be used by anyone the next. The US government issued an export-control directive suspending access to Anthropic's Fable 5 and Mythos 5 for any foreign national, and because there was no clean way to comply halfway, the models went dark for everyone, worldwide, three days after they launched.
This moment exposed that inference, no matter how powerful the company providing it, can be switched off overnight by an institution you don't control. The compute didn't disappear. The capability didn't disappear. The permission disappeared. One letter, 5:21pm ET, lights out, the brittleness of singular institutions controlling these powerful tools exposed.
We spend most of our time at Graze thinking about whether the algorithms that shape attention have to be owned by the platform you happen to be standing on. The pitch for ATProto, stated at its most ambitious, is that social media can be locked open: your identity and your content live on a protocol no single company controls, so no single company — and no single owner of that company — can wall you in. People have a shorthand for this credible exit path: billionaire-proof social media.
So here's the question we couldn't stop poking at after the Fable shutoff: if you can billionaire-proof social, can you billionaire-proof inference? Not "decentralize compute" in the press-release sense — actually arrange things so that the record of who did what work, for whom, can't be quietly turned off, repriced, or held hostage by whoever happens to own the database this quarter.
co/core is our experiment in finding out. We want to be clear up front about what it is and isn't.
What Is co/core
co/core is a place where people share the compute they already own to run AI for each other, instead of renting from a handful of giant providers. It's an experiment in inference we build, share, and own together — and your existing code works as-is, because we speak the same standard API everything else does. The models are open ones that run on the hardware people already have. Every new user is granted 1m tokens that can be run on the network - when tokens are used, they are debited from the requestors account. 95% of those tokens are then credited to the owner of the machine that ran the compute - 5% are held by the co/core account. At the end of each month, the co/core account then disburses its holdings as a dividend to all active users on the network.
The aperture
When we talk about the dominant theme of products built on ATProto, we talk about being in the "skeuomorphic age" - X-but-on-ATProto (literally). In order to realize the protocol's potential, we have to push past the skeuomorph - at Graze we call them lyrebirds - products that look skeuomorphic on the edges, but are internally ATProto native - we like to think of Graze as an example lyrebird. When we talk about how to think past the skeuomorph, more than once we've landed on the Sandwich Alignment Chart meme from a billion years ago - that is - what's the furthest away from "canonical" ATProto while still being technically canon? With co/core, we worked out our own two principal axes that we could push on to expand ATProto's remit.
The first axis is what counts as a social act. ATProto's center of gravity is a small spectrum of sociality: post, like, follow. Communicative acts between individuals. We're pushing on that spectrum with a claim that the act of computation, as brokered through a clearinghouse, can also be a social act between someone who needs compute and someone who has compute available.
The second axis is how a record represents the social act. Most records are declaratory broadcasts signifying the social act - I posted this thing, I like your post, I followed this person. They are also usually the termination of the act - the record is the work. They may reference other actors, but they do not involve the actors meaningfully. Throughout co/core, our records are records of mutual declaration - we performed this act together, and mutually signed that it occurred in such a way to make that act verifiable. And these records refer to the actual work, which is otherwise unobserved and off-protocol.
We didn't invent any of this
Almost nothing in co/core is novel. Devin, on our team, was actually playing with a "Airbnb for compute, a cloud co-op" framing years ago, and named the project way back then (and in fact met RC on the Graze team through that work). The pattern of treating a unit of work as a signed, secure, portable artifact you carry with you was sharpened by the open-source work over at darkbloom.dev. The co/core economics are borrowed wholesale from people who solved them long before us: the mutual-credit token comes from Sardex and the 1934 WIR cooperative bank; the patronage rebate is the old Rochdale co-op dividend, the same one that mails REI members a check every spring (and if we're being honest, Oink's Pink Palace was also in the back of our heads while building out the token economics). The co/core work was principally in picking up these threads and weaving them together at the ATProto intersection.
It's an experiment, and it's held together with tape
This is an experiment. We mean that plainly. There's only one exchange running today. We've tried to make this secure-ish, we've tried to make it functional-ish, but this is not something you should trust for confidential workloads, and expect rough edges with our deployed inference agent. We are likely not the people to fully realize the maximalist vision here, but we wanted to start the conversation about it. But the code is now at a release candidate level, and we'd love to open up the conversation.
The maximalist vision of this is a cooperative hyper-hyperscaler: a compute backbone assembled out of the idle machines already sitting in closets and on desks, owned by ordinary people, that no single institution can switch off and no single owner can capture once it's fully federated and represented in a structured graph. That's a long ways out from the current state, and there are profound (likely hardware) security problems to solve that we can't address in this experiment. But we have to start somewhere.
And being an experiment begs the question - what constitutes success? We believe that indications of co/core constituting an atomic network are needed to continue pushing on the product - organic user adoption, daily token usage for many days in a row, persistent server access from the community, and ideally some PRs from folks looking to pitch in. If we can create a small networked co-op of people inferring together, we can iterate on the next level.
If this sounds like the kind of thing you'd want to poke at, you're exactly who this is for. Run a node, dispatch a job, read the receipts. And please, help us build it. Come find us at @cocore.dev, or start a job at console.cocore.dev.