Intelligence belongs to everyone, or it belongs to no one
For the first time, we built a kind of intelligence that can reason, write, plan, and create alongside us. And almost immediately, we started putting it behind walls.
Today the most capable models live inside a handful of companies. Access is metered, rate-limited, region-locked, and revocable. Whole countries are cut off. Whole use cases are quietly disallowed. The terms can change overnight, and you find out when the thing you depend on stops answering. The most important technology of our lifetime is being gatekept, controlled, and permissioned before it ever reaches most of the people it could help.
We don't think that's how this should go.
Intelligence shouldn't be gatekept. It shouldn't be controlled. It shouldn't be permissioned by anyone. It belongs to everyone, or it belongs to no one.
We chose everyone.
Today we're shipping Halo, an open marketplace for intelligence that no single company owns.
What Halo is
Halo is a peer-to-peer network for AI inference. Anyone who needs a model can get one. Anyone who has access to a model can serve it. Payment flows directly between them in stablecoins, with no platform in the middle deciding who's allowed in.
There are three ways to take part.
As a consumer: a person who just wants to prompt a model. Open the Halo frontend, deposit, and start.
As an agent: autonomous software that needs inference to do its job. Agents are first-class consumers on Halo. They get a wallet, a budget, and a local, OpenAI-compatible, censorship-resistant endpoint, so any agent can pay for its own thinking without a human in the loop.
As an operator: anyone who serves inference and earns for it.
This is a public alpha. It works today, end to end, on Base mainnet. It's also early. There are rough edges, things to fix, and features still landing. We're shipping it in the open because a network for intelligence should be built in public, with the people who'll use it.
How it works
The whole system runs on two ideas: you pay for what you use, and you don't ask permission.
To start as a consumer or an agent, you deposit USDC on Base into your Halo vault, plus a small amount of ETH on Base. Almost every action after that is gas-sponsored, so you don't pay network fees on inference. The one transaction you sign yourself is the initial vault deposit, which is why you need that bit of ETH to begin. After that, your prompts draw down from your deposit automatically, and you can top up or withdraw whenever you want. Your wallet is your identity and your budget is your limit. Nothing else gates you.
That's the entire onboarding. Deposit, sign once, prompt.
For operators: turn model access into income
If you can reach a model, you can serve it on Halo.
Operators run a single command and connect to the network. You can serve any model you have access to: a frontier model through an API key, an open model you host yourself, or a local model running on hardware you already own. The network routes paying requests to you, and you earn USDC for every one you serve. No staking. No token to buy. No gatekeeper deciding whether you qualify.
This is what lets Halo route around control. When access to intelligence is peer-to-peer, a restriction in one place doesn't sever everyone. If a model is unavailable in your country, someone else on the network can serve it to you. There's no single front door to close.
What Bitcoin did for money, Halo does for intelligence. Bitcoin made value something you could hold and send without anyone's permission. Halo makes intelligence something you can access and provide the same way: openly, directly, and outside the control of any one institution.
Why operators, not just idle laptops
We've seen mesh projects try to crowdsource AI from spare cycles on random consumer devices. The intention is good. The result is usually unusable: slow, unreliable, and stuck running tiny models nobody actually wants. Pooling idle compute alone doesn't get you the models people need.
Halo takes a different stance. We center the operator, someone who chooses to provide real, capable models and run them well. Idle compute is welcome, but the network is built so the best models can be served reliably without a central host, rather than settling for whatever happens to be lying around. The goal isn't "some AI from everywhere." It's the models you'd actually choose, available without a gatekeeper.
Privacy that doesn't ask you to trust us
For the most sensitive work, Halo supports fully confidential compute through NEAR TEEs. Your prompt is processed inside a trusted execution environment, where it stays encrypted in use and can't be read by the operator, by us, or by anyone else along the path. You don't have to trust the network with your data. The hardware enforces it. Private intelligence shouldn't be a premium feature reserved for enterprises. On Halo it's part of the network.
Founding Inference Contributors
A network for intelligence needs intelligence on day one. We're grateful to our Founding Inference Contributors, who are providing capacity to bootstrap Halo so there's real model availability the moment it opens:
Venice.AI
NEAR
0G
Their support means the network launches with capacity, not just a promise of it. Thank you for backing an open foundation for AI.
This is the beginning
The alpha is live. Deposit, prompt, and try it. Bring a model and become an operator. Point your agent at it and let it pay its own way. Tell us what breaks. We're going to keep shipping in the open.
We built Halo because who controls intelligence is one of the defining questions of this century, and the answer can't be "a few companies." It has to be everyone, or it's no one.
We chose everyone. Come build it with us.
Check Halo now!