Value Delivered
What it would have cost to rent this compute instead of owning it.
Data freshness unknown
Read this before quoting the number
This is the figure most likely to be challenged, so here is exactly what it is and is not.
What it is. Allocated GPU-hours for each model, multiplied by that model’s public on-demand hourly rate, summed. Nothing else.
What it is not. It is not a budget, a saving that appears in any ledger, or a claim that the same work would have been purchased at list price. Anyone running this volume in the cloud would negotiate committed-use discounts, use spot capacity, or restructure the work — all of which would lower the bill substantially. Treat it as an upper bound on replacement cost at list price.
What it leaves out in the other direction. It excludes the staff, facilities, power, cooling, and networking that make the cluster run, and it excludes the data-transfer charges that dominate real cloud bills for research workloads. It is a comparison of compute rental rates, not a total cost of ownership analysis.
The rate table, its source, and the date it was retrieved are printed below and versioned in the repository. The site refuses to display a dollar figure at all unless a source and an as-of date are recorded alongside the rates.