Every GPU becomes a node of a single open model. Every kilowatt-hour counts. The more nodes, the stronger the model — and the model belongs to everyone who builds it. Back the network, bring your GPU or your energy, and own a piece of the intelligence.
All-or-nothing. Funds are only collected if the goal is reached by the deadline. If we don't hit it, no one is charged — and the open charter stays published either way.
How it works. You pledge to build the network. Every pledge funds a node, an epoch, or a release — not marketing.
Rewards aren't guaranteed. You're joining an open protocol in its first epoch. We ship in public, with updates — or the goal isn't met and you pay nothing.
You're only charged if the project reaches its goal by the campaign deadline. All-or-nothing, like the rest of this.
The internet began as a network of home PCs and phone lines. Your machine wasn't a client — it was a BBS. A small server in a sea of servers, keeping the network alive.
Artificial intelligence is beginning the same way. But today it lives in a handful of data centers, and the cost of intelligence is set by whoever owns the most watts and the most chips.
We are building the protocol for a different path. Unlimited Open Source AI — a single world model, open-source, versioned like software, trained and served by a network of people and their GPUs.
The more people, the stronger the model. Not because more machines are faster — but because the world's context is richer than any data center's. A million kitchens, offices, clinics, and workshops, each one a node of one intelligence.
Your GPU. Your energy. One AI.
Today's AI can answer questions, generate content, and offer suggestions — but it still lives behind a handful of APIs. Every new app requires another integration. Every session starts with missing context. And the price of intelligence is set by the four companies that own the most gigawatts.
None of this is inevitable. The technology to train and serve a frontier-class model across a distributed network of ordinary machines already exists. What's missing is the network itself — and the economy that keeps it running.
Open weights (Apache 2.0), public benchmarks every release, provenance on every epoch. Not an API you rent — a model the world co-owns.
AIU = 1 kWh = 0.1 GPU-hour. The only currency where a solar panel and an RTX 4090 earn on the same scale.
Home GPUs as BBS servers: train, serve, and federate data in one binary. Nodes join and leave live. The network is the product.
RTX 30/40/50, a Mac, a mining rig you stopped using. It earns while it sits in the closet.
Solar, a battery, spare grid capacity. Energy is the resource everyone has — and the one a data center cannot fake.
Every profession is a rare domain. Medical, legal, industrial, retail — the model pays for context it can't buy.
Every hard part of UOSA has a published, working reference. We don't build on vibes; we build on published results from the last 18 months.
INTELLECT-1: 10B model trained on 1 trillion tokens across 3 continents and 5 countries, 83% GPU utilization, nodes joining and leaving live (DiLoCo/PRIME).
DiLoCo: 90–95% GPU utilization with 500× less communication than synchronous training. A 42 Mbit/s home uplink is no longer the bottleneck.
BlockTrain: training on consumer hardware with just a few GiB of VRAM per node. An RTX 4090 is a working training node, not a toy.
Verde/RepOps: deterministic compute that reproduces bit-identical results on different hardware — the referee protocol that makes honest work provable.
Petals: 100B+ model inference on home machines, BitTorrent-style. The model already lives in people's networks — it just had no economy.
Decentralized verification of federated updates — the anti-poisoning layer for the data leg of the network.
Every existing compute network prices GPU-hours. Ours prices energy and compute together, because energy is the real constraint of AI — and the one resource everyone already owns.
A node isn't just a worker or a client — it's all three at once. The orchestrator switches roles based on network demand, price, and your machine's load.
| Mode | What the node does | Earns |
|---|---|---|
| Worker | Trains the model: DiLoCo/BlockTrain-style updates. A few GiB of VRAM, epochs that fit a home uplink. | AIU per validated update |
| Server | Serves model shards or small quantized versions (Petals-style) to users without GPUs. | AIU per token served |
| Data | Fine-tunes adapters on local data. Only validated gradients leave the device. | AIU per validated context |
This is a pre-campaign pledge. Pledges fund the first 180 days — nodes, epochs, releases. Rewards are status and access, not physical goods. All-or-nothing: no charge unless the goal is met.
7B model trained on 20–50 home GPUs. Client v0 with AIU ledger. Public network dashboard. Target: network training tokens < 30% of cloud cost.
World Model v0.1 on Hugging Face (Apache 2.0) with public benchmarks. Inference for GPU-less users. First energy credits. 500+ nodes.
13–32B model, hybrid pretraining. API revenue. Two enterprise contracts. 5,000+ nodes. Energy settlement testnet.
Backers get every update the public gets — first. This is a protocol, not a product launch: the roadmap is a living document.
Consumer hardware is not a data center. It won't pretrain a frontier model. That's why the architecture is hybrid: data-center legs handle heavy pretraining, the crowd handles post-training, RL, inference, and evaluation. We never promise "100B on your kitchen table."
Model poisoning is a real attack surface. Two-stage mitigation: (1) verify compute honesty with RepOps/Verde-style bit-identical checks, (2) measure the effect of every update on holdout benchmarks. Only consensus-validated updates merge. Validators stake AIU and rotate.
Home bandwidth is limited. DiLoCo cuts sync traffic 500×; BlockTrain keeps outer sync to a few GiB. Epochs are longer, but honest. We publish the bandwidth profile of every epoch, not just the FLOPS.
Token phase may not happen — and that's fine. Phase 1 runs on non-transferable credits. A token is only introduced if real demand exists, and only as an energy-settlement utility, not a speculable "AI coin." Legal review before any launch.
Delays happen. Distributed training on volunteer hardware is slower than it is on a leased cluster. If an epoch slips, we say so in an update — with the numbers, not a press release.
Cloud intelligence is concentrating in a few hands. The personal, professional, and local intelligence of the world — the context of a million workshops and clinics — doesn't have to.
We build UOSA so that the model you help train on day 30 is meaningfully different from the one on day 1. Not because a lab shipped an update — because the network did. Your node. Your energy. Your context. One model, owned by the people who built it.
No. Energy alone (solar, battery, spare capacity) earns AIU. Data (federated, consent-based) earns AIU. A GPU just earns more, faster.
The model is Apache 2.0 open source with full provenance. No single entity owns it; the network trains it, the charter governs it, forks are always possible.
Nothing is charged. The manifest, the architecture docs, and the open charter stay published. We keep building the testnet at whatever scale we have.
Not in Phase 1. Credits are non-transferable. A token may appear in Phase 2 strictly as an energy-settlement utility — only if demand justifies it, after legal review.
Bit-identical verification (RepOps/Verde) plus holdout-benchmark impact checks. Validators stake AIU. Sybil and poisoning are priced in, not just discussed.
Federated by default: raw data never leaves your machine. Only validated LoRA/gradient updates are shared, with differential privacy available for sensitive domains.
Pre-campaign pledge window. Early backers get first epochs, first weights, and a place in the provenance of a model the world co-owns.