Symbiosis: a decentralized workforce of home GPUs

Whitepaper, draft 0.1, October 2026. Adrian Nunez Gonzalez. This document describes a testnet. It will change as the software changes, and each change will be dated here.

Abstract

AI compute is concentrated in a few datacenters that cannot be built fast enough, burn new power, and belong to a handful of companies. At the same time, millions of capable graphics cards sit in homes, switched on and mostly idle. Symbiosis is a network that lets ordinary people lend the graphics card they already own to do real AI work, on their own terms, and be credited for it in the open. This paper describes the problem, the design of the first version, the credit that records contributed work, the trust model, the security promises the software keeps, and exactly where the project stands today: an invite-only testnet with about ten helpers.

1. The problem

Three problems are growing at once. We name them Speed, Efficiency and Symbiosis, which is where the credit's name comes from.

Speed: the datacenter crisis. AI needs more computers than anyone can build. New datacenters take years, swallow land and water, and fight over the same power lines.

Efficiency: the energy bill of AI. Every new giant building burns new power, while the hardware that could do much of the work already exists and is already switched on, in homes.

Symbiosis: the missing people. All of AI is concentrated in a handful of companies. The rest of us are users, never builders, never credited for the part we could play.

2. The idea

One answer covers all three: a workforce. People lend the graphics card they already own to do real AI work, when they choose, and are credited for it. No new building. No new power plant. No gatekeeper deciding who gets to take part.

This is not a protest against the big companies. It is a demonstration: that a network of small, independently owned machines can carry real load, count the work honestly, and show every participant what their machine did.

3. How the network is built

3.1 The helper

The helper is the software on a participant's computer. It takes short text jobs from a coordinator, hands each one to Ollama, the open AI engine the participant installed themselves, and sends the finished text back. Ollama writes the answer on the participant's own graphics card. The helper only ever moves text. It consists of two programs: the app with its status page, which listens only on the participant's own computer, and the worker that handles jobs.

3.2 Coordinators

Coordinators route jobs to helpers, count the work each helper does, and sign payout tickets. The helper finds coordinators from a list kept on a test blockchain, which it reads every few minutes at no cost. No single server decides everything, and a participant can see which coordinator they are working with.

3.3 Models

The helper runs only a short list of approved open models. Today that list is qwen2.5 at 3 billion or 7 billion parameters, about 1.9 GB, pulled by the participant from Ollama. The helper never downloads a model or any other file because a server asked.

3.4 Where the work comes from today

The jobs today are the builder's own research tasks: labelling and short summaries. About 30 tasks an hour, shared across all helpers. Jobs come in bursts and idle time is normal. Work for outside organisations comes later and is not on any schedule.

4. SES, the credit

SES is a test credit on Base Sepolia, a test blockchain. It records the GPU time a helper contributed to the network. When a helper wants their credits, they click Claim and approve one MetaMask transaction; on the testnet the fee is paid in free test ETH. Payout tickets are signed by the coordinator that counted the work, and a helper's own computer keeps a record of every job it did, so the two can be compared.

SES has no money value. It cannot be bought from us, and we make no promise that it will ever be worth anything. It is a record of contributed work and nothing more. Nothing on this site or in this paper is a prediction or an offer.

5. Trust levels

Every helper has a level, from T0 (new) to T5 (operator). New helpers get the simplest work. Riskier work unlocks only with a track record. Levels are earned by completed, verified jobs, not bought.

6. Security promises

These are promises the software keeps, not policies we ask participants to trust.

What it does not protect: the text of other people's jobs passes through the participant's computer, to Ollama and back. It should not run on a work computer, a shared computer, or any machine with confidential files. It uses power and the graphics card while it works. Power controls (run only when idle, cap GPU use) are planned and not yet built. The current build is unsigned, so Windows shows a warning on first run; participants can verify their copy against SHA-256 hashes sent separately.

7. What participants need

8. Where this stands

Symbiosis is an invite-only testnet. It starts with about ten helpers, each chosen by the builder, and each invite is read and answered by hand. A wider testnet follows once invite codes are built into the software. Real compute work for outside organisations comes after that. There are no dates. We will say what is true when it is true, and this paper will be updated when it is.

9. What the builder owes participants

Every job a participant's machine does is recorded on their own computer, and the status page shows their totals. If the network is shut down, participants will be told why. If it works, they were there first.

Where this is today: an invite-only testnet with about ten helpers. SES is a test credit on the Base Sepolia test network. It has no money value, it cannot be bought from us, and we make no promise that it will ever be worth anything. The text jobs your computer runs today come from the builder's own research work.