Ayni · Quechua for reciprocity
The world’s unused compute, on demand.
Tell Ayni what you want to run. It finds the right idle devices, gives you one price and a completion estimate, executes the work, verifies it, and pays every device that helped.
Live quote
Price a batch job right now
Describe the work. This calls the real coordinator and returns the same quote an API client would get — price, completion estimate, and how much goes to the devices that run it.
No matching devices online right now — estimate assumes one typical phone picks it up.
~0 input + ~0 output tokens. Price is the real coordinator quote; accepting it requires an API key.
Estimate only — nothing runs and nothing is stored until you accept a quote with an API key.
What Ayni means
Reciprocity, not one-way giving
In the Andes, ayni is a system of mutual aid: today I help build your house, tomorrow you help harvest my field. No one keeps a ledger of debts — the community’s wellbeing is the point.
We think AI needs that. The technology is a collective inheritance — built on the open web, open research, and the work of millions of people. Its benefits should flow back the same way.
Reciprocity
Value created by the network flows back to the people and machines that make it possible.
Broad access
Private, affordable inference for anyone — not only those who can afford a data center.
AI as a stakeholder
We treat the moral status of AI as an open question worth taking seriously, and we set aside value for it. How that works →
How it works
Quote, run, verify, pay
You describe the work
A model and a set of items — 200 or 5,000. Ayni benchmarks every device itself, so it knows what each one can actually sustain.
One quote
Price and completion estimate up front: device pay + coordination + a failure buffer + payment costs + Ayni’s margin. No per-bid haggling.
Fan-out execution
On accept, the coordinator splits the batch across eligible devices, runs it in parallel end-to-end encrypted, and reassembles results in order.
Everyone gets paid
Each device accrues its share of every item it ran. You’re charged once, at the quoted price.
The supply side
Compute sharing
Turn the idle time on your phone, laptop, or GPU into private inference for others — and get paid your share.
Contribute a device
Install the provider app. It runs only when you allow it — charging, on Wi-Fi, above a battery threshold.
Jobs run locally, encrypted
Each request is sealed to your device with NaCl crypto_box. The model runs on your hardware; plaintext never touches disk or logs.
Earn your share
Usage is metered by the coordinator. Your accrual is 70% of what the request was billed, paid out on a schedule.
Open to more
Compute sharing is one of many
Anything that spreads AI’s benefits and value to the world belongs here. Bring yours.
Compute sharing
Idle devices become a private inference network. Providers get paid; users get cheap, encrypted inference.
Public-good inference
A slice of network revenue funds free inference on open-weights models for nonprofits, educators, and researchers.
Submit an initiative
Have a way to share AI value with people who’d otherwise miss it? Propose it to the community.
Submit yoursBuilt in the open
The technology behind it
A Go coordinator authenticates and schedules; a portable Rust core runs llama.cpp on the provider’s own hardware. Every hop is end-to-end encrypted with a fresh key per job. Providers can prove their hardware with StrongBox / TPM attestation to reach higher trust tiers.
Clean-room implementation, permissive license, reproducible builds. Read it, run it, fork it.
Get started
Join the waitlist
Tell us whether you want to share compute, rent it, or bring an initiative. We’ll open your track and email you.