GPUDECK · THE 60-SECOND VERSION

Why hourly price lies.

Every rental marketplace shows you one number: dollars per hour. But you don't buy hours because you want hours — you buy them to finish work. And two GPUs holding the same hourly price can finish very different amounts of it.

Two GPUs, one paid hour each: the same spend finishes different amounts of work GPU A · ONE PAID HOUR 7 done GPU B · ONE PAID HOUR · SAME PRICE unfinished 3 done Same hourly price. Same spend. GPU A finished each piece of work at less than half GPU B’s cost.
Illustration of the mechanism, not a measurement: the blocks are completed jobs inside one paid hour. The measured version of this picture is the table below.

This is not hypothetical.

GPUDeck measures it nightly on a real production render fleet and publishes the result as cost per 1,000 completed renders — the licensed hourly price divided by what an occupied hour actually finished. One real cell from the value board:

MEASURED · SAME WORKLOAD, SAME WINDOW · SOURCE: /VALUE
WORKLOADGPUPER 1,000 COMPLETEDVS BEST
dark_beast_z_image_turbo_v9_bf16 RTX 5090$1.15—
RTX PRO 4500 Blackwell$2.442.1× more

The same hardware pair can also flip on a different workload — which is why GPUDeck never averages workloads together and never publishes a single “best GPU”. A ranking would be a lie with better typography. What you get instead: the per-workload record, dated, with sample sizes, and every number downloadable so you can check the arithmetic.

So read the record in this order.

What does a card cost per hour? Per-card pages: median, floor, trend — marketplace and cloud kept apart. Which card finishes work cheapest? The value board: cost per 1,000 completed renders, per workload. What do operators actually run? The census: cards entering and leaving a live network, daily since Dec 2025.