Compute economics

A complete record of GPU spend and what it produced. Every figure on this page is computed from an append-only ledger reconciled against vendor billing — nothing is hand-typed, and the underlying file is downloadable below.

Cost per rendered image
Images renderedarchive sweep, verified count
Total GPU spendall-in, since day 1
Images per dollarthroughput per unit cost
GPU-hours loggedwhere duration was recorded

Burn rate

7-day trailing average, USD per day · daily actuals shown behind

7-day trailing average daily actual
▸ table view
datedaily7-day avg

Capital funded vs. deployed

cumulative USD since day 1 · the gap is cash on hand

funded deployed
▸ table view
datefundeddeployedon hand

Spend by GPU class

every dollar of compute, by the card it ran on

“Mixed fleet” is a day-level total from the vendor's billing API, covering every pod that ran that day — those rows predate per-pod attribution and no single card can be recovered from them. A negative line is compute released back, not spent.

Spend by workload

what the compute was doing

Line items

every entry in the ledger, newest first · rows, nothing omitted

datehardwareworkloadhoursamount

Spend comes from the compute vendor's billing API; funding is reconciled line-by-line against the payment processor's transaction record. Both live in an append-only JSONL (download it). Entries are never edited after the fact — corrections are appended as new rows. Hardware and workload are classified from each entry; hours are shown only where the entry recorded a duration, so GPU-hours is a floor, not a total. The site refuses to deploy if this page's data is older than the books.