FROZEN EDITION — ISSUE 253 / 2026-08-26. THE CURRENT SHEET IS AT GPUDECK.COM →

SINCE ISSUE 252: methodology v3 → v4 (dated note in METHODOLOGY.md) · coverage: source lium-public-nodes added to the register · tracker A100 80GB marketplace P50 moved +5.9% ($1.69 → $1.79, n=39) · tracker H100NVL 94GB marketplace P50 moved +7.2% ($2.97 → $3.19, n=30) · tracker RTX4090 24GB marketplace P50 moved +31.0% ($0.56 → $0.74, n=7) · tracker RTX5090 32GB marketplace P50 moved -6.6% ($0.61 → $0.57, n=9) · machine diff · ← issue 252 · all issues
GPUDECK THE OPEN RECORD OF THE GPU RENTAL MARKET
ISSUE 253 / 26 AUG 2026 / OPEN CSV / SHA-256 MANIFEST
THE DECK / OPERATIONS WAITLIST / JOIN FIRST COHORT

Hourly price
is half the story.

Price says what the clock costs. Goodput says what the paid hour finishes.

39 GPU MODELS IN CENSUS / 27 PRICED / 252 DAYS / LATEST CELLS 2026-08-26
THE 60-SECOND METHOD →

PRICE CONTEXT / 2026-08-26

CONTEXT, NOT RANKING
GPUP50FLOORWINDOW TRACE
RTX 5090 32GB ●$0.57/h$0.37RTX 4090 24GB ●$0.74/h$0.34H100 80GB$3.25/h$3.25H200 141GB$4.00/h$4.00A100 80GB$1.79/h$1.39L40S 48GB$1.37/h$1.37

● = this model also has measured goodput cells on the observatory

Every card, its own page →

THE OTHER HALF / PRODUCTION GOODPUT

MEASURED ON ONE DISCLOSED FLEET · NEVER A RANKING

RENTAL RATE × GPU-HOURS ÷ COMPLETED RENDERS

↓

LATEST MEASURED CELL — USD PER 1,000 COMPLETED RENDERS

$9.82

NO CROSS-CELL COMPARISON IN THIS WINDOW / ONE DISCLOSED FLEET

workload: chroma-v48-detail-svd_fp8 · RTX PRO 4500 Blackwell · CELLS OBSERVED 2026-08-26 · 75 ATTEMPTS · 7-DAY PRICE WINDOW

DAILY HARDWARE CENSUS

9 PRESENT NOW / 39 OBSERVED / 96 WORKERS

08 1608 25RTX 409048RTX 509029RTX PRO 6000 Black6RTX PRO 5000 Black6RTX 3090LAST 07 19×H100 80GB HBM3LAST 07 29×

CONTEXT ONLY — NEVER WEIGHTED INTO GOODPUT · FULL LIFELINES →

SUPPRESSED WHEN UNDER-REPLICATED · NO SPONSORS · NO RANKINGS · REBUILD FROM CSV

FOR OPERATORS · FIRST COHORT

The ops deck: your rigs earn while you sleep, or they don't.

The operations layer this record is measured with — pre-emptive memory recycling that saves paid jobs, earning-layer truth checks, incidents that heal themselves. Productizing next.

Join the waitlist →
THE METHOD · 60 SECONDS

Why hourly price lies.

Two GPUs at the same hourly price can finish very different amounts of work — the measured spread on this fleet is printed above. One page, one picture, dated evidence.

Read the method →
A short weekly market brief is coming — price moves, cards entering and leaving the census, one measured finding. No spam; unsubscribe is one reply.
REBUILD FROM CSV price-context.csv goodput-ratios.csv census.csv value-board.csv methodology.md citation.cff