An AI product running entirely on independent miners.

The part worth checking

Plenty of projects describe a compute network. The question a sceptical reader should ask is whether anything actually runs on it, and that is answerable here without taking our word for it.

Every AI request instead goes to the Tenseuron compute API, is queued as a task, is answered by independent miners, and comes back through validator consensus. That is the same public path any network you launch would use.

How a request travels

When the wallet's AI screens a token or evaluates a position, the request does not go to a model — it goes to a market.

1. 1

   The wallet submits a task

   The extension posts to the compute API with the user's wallet signature. No model is named; the task describes the job.

2. 2

   Miners pick it up

   Miners polling that network claim the task and run it on whichever model they have configured — a local model on their own machine, or a hosted provider with their own key.

3. 3

   Validators score the answers

   Each validator independently scores submissions 0–100 using its own model. The verdict is a stake-weighted median.

4. 4

   The accepted answer returns

   The winning submission comes back to the wallet, and the reward settles on-chain to the miner and the validators who agreed.

Enlarge

Whose API keys pay for the inference?

The miners', never the platform's. This is the detail that decides whether a network is really decentralised or just a proxy with extra steps.

A self-hosted miner runs the CLI and configures whatever model it likes — Ollama or LM Studio locally, or a hosted provider key. That key stays on the miner's machine and is never transmitted to us.

For people who would rather not run a process, the platform can operate a miner on their behalf. Even then it decrypts that user's own key to do the work. The platform does not hold a provider account that inference is billed to, which means the economics of the network do not quietly depend on us subsidising it.

What it costs to run

A task consumes a small amount of the network token locked against it, derived from the network's configured budget rather than metered per token of input and output. In practice the per-task cost is a rounding error compared with the value of the job — a safety check that stops someone buying a honeypot is worth considerably more than the fraction of a cent it costs to run.

Miners are paid from the network's reward pool, and the payout is a fraction of what remains in it rather than a fixed emission. That has a useful property: the pool decays toward a floor instead of emptying, and the split totals exactly 100% of each reward, so the network never pays out more than it takes in. Trading activity refills the pool, so a network people actually use pays its miners better than a dormant one.

Every wallet gets its first three tasks free on any network, enforced from real task history rather than client state.

What we gave up

Building this way cost us things a hosted API would have given us for free, and it would be dishonest to present the architecture as strictly better:

• Latency is not guaranteed. A task waits for a miner to claim it. That is fine for screening a token or evaluating a position on a timer; it would not be fine for a streaming chat cursor.
• We do not choose the model. Miners do. Consensus filters weak answers, but we specify the job rather than pinning a model version, which means output style varies more than it would against one provider.
• Tasks are processed on hardware we do not control. This is why the wallet sends market and token data to the network and keeps private keys entirely on the user's device — keys are never part of a task, and nothing confidential should be.
• The network is early. A small miner pool means less redundancy than a hosted provider offers. That improves as networks attract miners, but it is the current state.

Frequently asked questions