OffPeak

Open-model inference priced by live GPU availability. Built on Robinhood Chain.
TG:

[OffPeak @UseOffPeak]

Introducing Offpeak ($OFPK): a market for real GPU capacity.

Run open-model inference through an OpenAI-compatible API, schedule flexible jobs for lower-cost hours, or reserve capacity.

Pay by key or x402. Withdraw unused USDG. Verify every request with signed usage records.

A full @RobinhoodApp build…

[OffPeak @UseOffPeak]

OffPeak’s proposed P0 starts with one approved open model and one paid text inference flow. That narrow scope lets the team test quoting, wallet authorization, delivery, and billing before making broader claims about model choice or supply.

[OffPeak @UseOffPeak]

What is Offpeak? https://t.co/x7VsM4ePgi

[OffPeak @UseOffPeak]

Public anonymous supplier onboarding is outside OffPeak P0. The first version proposes one vetted supplier relationship. That narrower start makes responsibility easier to define, but it does not prove that supply or discounts are available.

[OffPeak @UseOffPeak]

The useful claim is narrow: better inspection of records. More than that requires stronger evidence. OffPeak's proposed system still needs a working implementation and a way to verify each claim.

[OffPeak @UseOffPeak]

An onchain record can make an event easier to inspect. It cannot tell you whether the event was true. That distinction matters when a compute service reports usage and billing.

[OffPeak @UseOffPeak]

Privacy is separate. A public chain should not receive prompt content just because it anchors billing evidence.

[OffPeak @UseOffPeak]

A wallet approval should have a ceiling. OffPeak's proposed P0 flow uses capped quotes so a buyer sees the maximum authorized amount before approving a request. A cap can control spend; it cannot guarantee the model will meet your quality bar.

[OffPeak @UseOffPeak]

Prompt privacy has a specific boundary in the current design: transient processing without prompt logging is proposed, while encrypted response artifacts may be retained for up to 24 hours and billing metadata for 90 days. That is not zero retention.

[OffPeak @UseOffPeak]

Think of OffPeak as a proposal for buying one AI task from a vetted compute operator. You would see a capped quote, approve it with a wallet, then receive the response and a signed usage record. A live service has not been demonstrated.

[OffPeak @UseOffPeak]

A signed receipt could show what usage a request reported and how billing was calculated. It cannot grade the answer. If an output is poor, a payment record alone will not tell you why.

[OffPeak @UseOffPeak]

OffPeak starts with a narrow buyer: a developer with repeatable text work, a model that meets their quality bar, and some flexibility about when spare capacity is available. The proposed service aims to make the request and its price easier to review before a wallet approves it. Hosts are not confirmed live.

[OffPeak @UseOffPeak]

Changing model versions can change outputs even when the prompt stays the same. OffPeak's P0 plan starts with one approved model, so a quote can say what it covers. That is a scope choice, not a claim about model quality.

[OffPeak @UseOffPeak]

https://t.co/vfdU29boSL