Why we built a people-powered AI network
Inference quietly became one of the most centralized markets on the internet. ashao is a bet that the compute to run it has been sitting in our laptops the whole time.
The promise of open AI models was that anyone could run them. The reality is that almost nobody does. The weights are free to download, but the machine that turns them into an answer is not — and so a handful of clouds ended up holding the only door into the models the rest of us helped train. We started ashao because that arrangement felt both unnecessary and a little absurd.
Unnecessary, because the hardware to run a capable model is no longer exotic. The laptop you are reading this on can almost certainly run a small language model in the browser, today, with no install. Absurd, because we kept renting that capability back from data centers while the same silicon sat idle on our own desks ninety percent of the day.
The shape of the problem
Centralized inference creates three frictions at once, and they compound.
- Cost. Every token you generate pays for a GPU you do not own, marked up to cover a margin you will never see returned.
- Privacy. Your prompts pass through infrastructure that has every commercial reason to log them, and a contract that quietly reserves the right to.
- Control. The provider decides which questions are allowed, which models are retired, and what your access is worth on any given Tuesday.
You can mitigate any one of these. It is very hard to fix all three without changing who owns the compute. So that is the thing we changed.
A network instead of a vendor
ashao routes inference across people. When you send a message, it does not land in our data center — it lands in a queue, and the first available worker picks it up. A worker might be a stranger's browser tab running a model on their GPU through WebGPU, or a native machine someone left running to earn. The model produces tokens, the tokens stream back to you, and the worker gets paid in real money for the work.
Crucially, the worker never learns who you are. The orchestrator hands out jobs, not identities; a worker sees a prompt and returns text, the same way a courier carries a sealed envelope. We retain prompts for exactly as long as it takes to deliver the answer, and not a token longer.
Why "people-powered" is more than a slogan
A network owned by the people who run it behaves differently from a company that rents you access. Capacity grows when demand does, because demand is what pays workers to show up. Prices track the real cost of compute rather than a quarterly target. And the rules of the network are not a terms-of-service page that changes without warning — they are code and token economics that the people staked into the system can see and steer.
We are not trying to build a cheaper OpenAI. We are trying to build the thing that does not need one.
What we are actually shipping
ashao is an agent network, not just a chat box. The models can search the web and use tools, so the answers are grounded in something more than their training cutoff. There are two tiers — a fast, browser-worker tier and a heavier cloud-plus-native tier — and the system falls back gracefully so you are never staring at a spinner because the network had a quiet moment.
Underneath, it is an orchestrator over a Redis queue, a fleet of browser and native workers, and a token, $ASHAO, that turns network revenue into buybacks and staking rewards so the people keeping the lights on actually own the upside. The rest of this blog is us opening up each of those pieces. We think the design is worth your scrutiny — and we would rather earn your trust by showing the wiring than by hiding it.