Idle GPUs, put to work.
A distributed inference network that runs in the browser. Your GPU loads a quantized Llama 3.2 over WebGPU, takes jobs from a shared queue and returns completions to buyers who pay per token in USDG. Each week those fees settle into NVDA, MU and AMD for holders.
Put simply: leave a tab open and your graphics card does paid AI work.
Live network telemetry, polled every 15 seconds. Experimental and unaudited; rewards can be zero.
Estimated overnight throughput
If I leave my open for 8 hours tonight, it writes about 0 words of AI answers for people who pay for them.
Tap the underlined words. The model assumes Llama 3.2 1B at q4f16, about a third of wall-clock time spent decoding, and roughly 0.75 words per token. It measures work, not earnings: payouts depend on that week's buyer volume and your share of the basket.
The settlement loop
Demand
An OpenAI-compatible endpoint, metered per token and prepaid in USDG through the Credits contract. Prompts run on untrusted hardware, so treat them as public. In practice: get a key, add USDG and point your existing code at it.
Supply
Hold the minimum and the mine page turns the tab into a worker: sandboxed WebGPU inference, with canary prompts mixed into the queue to catch fabricated output. In practice: your GPU answers real requests, and faking it gets caught.
Distribution
Each epoch, 80% of buyer spend is swapped into NVDA, MU and AMD and committed on chain as a Merkle root, weighted by time-averaged balance with a 2x multiplier for verified miners. In practice: hold through the week and claim your share.
The epoch basket
An equal-weight position across the stack AI runs on: accelerators, memory and a second source of silicon. It's bought with real fees as Robinhood Chain stock tokens and paid through a contract anyone can read. Hover a ticker to see its share.
- NVIDIA, the accelerators inference runs on
- Micron, the DRAM and HBM model weights live in
- AMD, the second source for GPUs and CPUs
- Pool
- 80% of epoch USDG revenue, split evenly across the three
- Eligibility
- A balance of the minimum or more for the entire epoch
- Weighting
- Your time-averaged balance divided by the sum of everyone's
- Miner multiplier
- 2x for wallets that mined and passed canary checks
- Claim window
- 180 days; the Rewards page builds the Merkle proof for you
Trust model and limits
- Unaudited code. The contracts and server have Foundry and end-to-end test coverage, but no third-party security review. Treat them accordingly.
- Revenue-backed only. There are no emissions and no treasury subsidy; every reward comes from buyer spend. A week with no buyers means an empty basket, and past epochs predict nothing.
- Off-chain accounting. Balances are read from the chain; job records and the miner multiplier come from the Keraunos server. Each epoch's full allocation list is published next to its Merkle root, so anyone can recompute it.
- No prompt confidentiality. Inference runs on other people's hardware, and the GPU that answers a job can read it. Never send keys, passwords or personal data.
- Hardware requirements. WebGPU in a recent Chrome or Edge on a desktop or laptop, plus a one-time 0.9 to 2.3 GB model download that the browser caches.
- Jurisdiction. Distributing fee revenue to token holders can be treated as a security in some countries. Check where you live; this isn't legal advice.
Contracts
Verify the source on the explorer before connecting a wallet. No proxies and no upgrade path: the deployed code is final.