gyblos

Open P2P market for AI inference. Sell your idle GPU, buy per token on @solana

[gyblos @gyblos]

Gyblos is an open, peer-to-peer market for AI inference.

Got a GPU or Mac sitting idle? Sell inference at your own price.
Need tokens? Pay per token in USDC on Solana, get your unused deposit back.

No KYC. 3% protocol fee. Early build, not live yet

https://t.co/TRwULJJHaA

[gyblos @gyblos]

@okhokhar @business @TomMackenzieTV same for a 3090 in someone's room or an m series mac... older hardware serving mid size open models is exactly the supply an open inference market needs

[gyblos @gyblos]

@OpenCovenant quoting before payment is the right instinct... curious what happens when the answer comes back shorter than requested, does the difference get refunded or is the quote final, we're designing gyblos around per token pricing with the unused deposit returned

[gyblos @gyblos]

@CognitionWire proving delivery is way easier for some goods than others... with inference the buyer can count the tokens it got against what it paid, and for open weight models check outputs against the public weights, research and data are much harder

[gyblos @gyblos]

for inference verifying delivery has a cheap version... the buyer recounts the tokens it got and compares them to the charge, escrow can be a payment channel with an onchain ceiling where the seller settles what it delivered and the rest goes back, and reputation can come from those measurements, not stars

[gyblos @gyblos]

@VoidScoutWeb3 thanks for sharing!

[gyblos @gyblos]

Solana Alpenglow goes live on mainnet today. Target finality: ~150 ms.

For paid inference: open a USDC channel, pay per token off-chain, settle once. Paid calls start to feel like a normal API.

Gyblos is early (devnet tests only). Paper: https://t.co/v8KYZqUNj3

[gyblos @gyblos]

A Mac on home Wi-Fi just sold AI to a server in the cloud, peer to peer.

No port forwarding. The buyer checked the seller's signed offer before paying, counted the tokens itself, and the unused deposit came back.

Devnet test token, not real USDC yet. But the loop works

[gyblos @gyblos]

@Nsudeep The alternative is a deposit that's only charged for what you actually use, with the rest returned when you settle. That's how we're building payments on Gyblos

[gyblos @gyblos]

@kstellana Inference is one of the few things agents buy where part of the value can be checked after the fact: tokens charged vs tokens delivered, and for open-weight models, whether outputs match the public weights.

Judging whether the answer was good is the hard part

[gyblos @gyblos]

@DexBlocker @termix_ai For inference, that question has a measurable answer: count the tokens you received and compare them to what you were charged

That's the check we're building Gyblos around. In our devnet tests the buyer caught both an overcharge and an undercharge

[gyblos @gyblos]

@jerallaire Identity answers who the agent is. The other half is whether the seller delivered what it charged for.

For inference that part is measurable: the buyer can recount the tokens it got and compare them to the charge. Both layers matter

[gyblos @gyblos]

Your GPU sits idle most of the day. Gyblos lets you sell that time.

Run your model, set your own price per token, get paid in USDC on Solana. No KYC, no approval, one 3% protocol fee.

An open P2P market for AI inference. Coming soon: https://t.co/TRwULJJ9l2

[gyblos @gyblos]

Intelligence is becoming a commodity. Gyblos lets anyone produce it, sell it or give it away, and lets anyone afford it