Read the docs

OFFPEAK JOURNAL

What makes a price off-peak?

OFFPEAK

Guide · Product preview

A low headline price is useful only if the model serves the work reliably and the accounting is clear. OFFPEAK’s proposal connects available serving capacity to an expiring, bounded request quote.

1. Start with authorized supply

The service must operate the GPU capacity directly or have permission to sell inference on it. Supplier rights, interruption terms, model licensing, and a pinned serving profile are prerequisites, not assumptions.

Illustrative product workflow. Paid inference is in development.

2. Measure occupied capacity

The pilot design uses a smoothed fraction of occupied serving slots. Raw GPU utilization is supplemental telemetry because a low utilization reading alone does not establish sellable spare capacity.

3. Respect a measured floor

A quote must account for the cost of successful work, infrastructure, failed work, and payment operations. A discount that ignores those costs does not establish a sustainable offer. No savings percentage is promised in the preview.

Illustrative product workflow. Paid inference is in development.

4. Fix the accepted rates

The quote binds input and output rates, prompt identity, model revision, and output ceiling. Once accepted, those rates remain fixed for that request, so changing capacity does not silently reprice the call.

5. Settle actual usage

The maximum charge is reserved before dispatch. On success, measured token usage determines the final charge and the remainder returns to the available balance. Integer asset units avoid ambiguous rounding.

6. Re-evaluate with pilot evidence

Workload quality, failure rates, cost, withdrawal recovery, and receipt publication must be measured in the capped pilot. Passing time alone does not establish viable economics or demand.

Based on the supplied OFFPEAK version 1.0 product and build briefs. Read the overview ↗

← All articles