building publicly owned AI

This edition examines whether an AI model can have an economy without a single owner, and what distributed training, onchain coordination and tokenised rights would need to make that possible.

AI is becoming easier to use and harder to own.

A developer can rent a model powerful enough to write software, review contracts or work through scientific problems. The intelligence underneath the product, however, remains someone else’s property. The provider controls the price, the capabilities and the terms of access.

Open-weight models changed part of that relationship. Developers can download, adapt and run them without routing every request through one company.

But access is not ownership.

The people using the model do not participate automatically in its economics. The model has no built-in way to finance its next generation. Its future still depends on the company, government or laboratory willing to pay for it.

Decentralised AI is beginning to test a different structure: intelligence that can earn, improve and remain available without becoming the permanent property of one company.

Frontier AI rewards scale. Training a large model requires specialised chips, vast amounts of power and data, and teams capable of operating the entire system together.

The companies that control those resources also have the clearest business model. They pay for training, retain the model weights and charge users for access. Revenue from the current system helps finance the next one.

This structure has produced rapid progress. It has also made intelligence something most businesses rent. The concern is that a small number of private companies are setting the rules for infrastructure that may eventually sit beneath much of the economy.

Here’s Jack Dorsey’s take (co-founder of Twitter):

Open weights offer the clearest counterweight. They make powerful systems available beyond the companies that trained them.

They still leave the economics unresolved.

An open-weight release is not always open source in the full sense. A lab may publish the model weights while keeping the training data, code, cleaning pipeline or training recipe private. The licence may also restrict commercial use or redistribution.

Where self-hosting is permitted, developers can run the model without paying the original lab for every request. That improves access and competition, but it also means the released model does not capture revenue each time someone uses it.

The lab can still earn through hosting, enterprise services and applications built around the model. The model itself does not have an economy that pays for its continued development.

In a recent un# conversation, CoinFund founder Jake Brukhman described a structure between proprietary and open-weight AI: a model that is both valuable and public.

Users would still pay for its output. The difference is where that money goes. Instead of flowing entirely to one company, part of it could return to the network that supplied the compute, research and operation behind the model.

Until recently, this idea failed at the technical starting line. Training a serious model required machines packed inside one data centre and connected fast enough to exchange enormous amounts of information continuously.

That is no longer an absolute constraint.

Covenant-72B completed the pre-training of a 72-billion-parameter model through a permissionless network. Participants communicated over the public internet, and the system allowed peers to join without approval from a central gatekeeper.

The run still required powerful hardware. Each participant needed eight B200 GPUs. Covenant did not show that spare laptops can replace a frontier data centre.

It demonstrated something narrower and more important: independent participants can contribute to one serious training process without placing every machine under one owner.

The advance came from changing how the model trains. Machines on the open internet cannot remain in constant lockstep like GPUs connected inside a data centre. Distributed methods allow each participant to work for longer before sharing a compressed update with the rest of the network.

The internet did not become faster. The training process became less dependent on continuous communication.

The model calculations still happen off-chain. A blockchain can sit around that process, registering participants, tracking contributions and coordinating rewards when no one wants a single operator controlling the record.

This makes distributed training technically credible. It does not decide who owns the model when the run ends.

A distributed training run can end exactly like a centralised one. Hundreds of participants may contribute compute, yet one company receives the final weights and captures the resulting value.

Production has been distributed. Ownership has not.

Jake’s proposal connects the two layers. Compute providers, researchers, data contributors and operators could receive rights linked to the model they help create.

Users would pay the network to access the model, and the protocol could distribute part of that revenue among participants.

“If you want to get the output of this model, you’ve got to pay the network.”

That would place the model somewhere between a proprietary service and an open-weight release. It could generate revenue without one company owning and operating the entire system.

The design is harder than issuing a token.

Technical protection is only one part. A token may allocate governance or revenue rights, but it does not automatically establish legal ownership. Licences, contracts and jurisdiction still determine what those rights mean.

A publicly owned model therefore needs three things to align: technical control, economic rights and legal enforceability.

The first publicly owned model is unlikely to compete directly with the largest general-purpose chatbots.

The added coordination becomes worthwhile when no participant can assemble the full system alone.

A medical model may depend on expertise and data held across hospitals that cannot place everything in one database. A financial model may require input from institutions that cannot expose client information or proprietary records. Scientific models may draw on researchers, laboratories and specialised agents contributing different parts of the work.

In these cases, the network is doing more than collecting GPUs. It allows valuable inputs to meet without requiring every contributor to surrender them to one company.

The same structure can coordinate work after the model has been trained.

Jake cited an Eigen Labs experiment based on a quantum-cryptography result published by Google without its underlying circuit. Researchers and AI-assisted contributors used a public repository to inspect and improve candidate circuits together.

The agents did not replace the researchers, and the work did not depend entirely on a blockchain. It showed how people and specialised agents could build on one another’s progress around a visible, shared problem.

That is a more plausible starting point for publicly owned AI: specialised intelligence built by participants who need attribution, payment and a common record of what each contributed.

> in practice

Any system that allows software to act on someone else’s capital faces a related question: can the user verify what it did without relying on the operator’s account?

At aarnâ, âTARS analyses opportunities while smart-contract and governance controls limit what it can execute. Allocations and rebalances operate within defined policies, and completed transactions remain visible onchain.

Compute may still gather around a few well-funded providers. Early investors or validators may accumulate most of the governance power. Participants may manipulate rewards, collude to extract the model or spread responsibility so widely that no one remains accountable.

A system can decentralise one layer while concentrating another.

Public ownership exists only when no participant can quietly capture the model, its revenue or its governance.

That standard is considerably harder than launching a token. It is also what separates a new ownership structure from a conventional company with a distributed supplier base.

Closed labs will continue building frontier models. Open-weight releases will continue making powerful systems cheaper to use and easier to adapt.

Publicly owned models do not need to replace either one.

They need to prove that a network can produce intelligence worth paying for, keep it available, reward the people who sustain it and prevent one participant from taking control of the entire system.

Open-source software allowed code to outlive the company that wrote it. Public blockchains allowed networks to operate beyond the companies that launched them.

Publicly owned AI attempts something harder: intelligence that can earn, improve and remain useful without becoming the permanent property of one firm.

The model will still need capital, builders, governance and rules.

It just may not need a single proprietor.

This issue grew out of a conversation with Jake Brukhman, founder of CoinFund, on distributed training, publicly owned AI models and agent networks. The full episode is available on un#.

reflections-

Gain an edge in DeFi alpha with aarnâ’s AI-driven insights and DeFi vaults. Try the dApp now.