FROZEN EDITION — ISSUE 260 / 2026-09-20. THE CURRENT SHEET IS AT GPUDECK.COM →

SINCE ISSUE 259: Lium direct B300 SXM6 AC 269GB ON_DEMAND_SINGLE P50 moved -5.9% ($8.50 → $8.00, n=5) · Lium direct H100 80GB HBM3 80GB ON_DEMAND_SINGLE P50 moved -18.8% ($1.60 → $1.30, n=14) · Lium direct H100 PCIe 80GB ON_DEMAND_SINGLE P50 moved +49.3% ($1.50 → $2.24, n=3) · Lium direct H200 140GB ON_DEMAND_MULTI P50 moved +27.7% ($3.25 → $4.15, n=3) · Lium direct H200 140GB ON_DEMAND_SINGLE P50 moved -19.6% ($4.23 → $3.40, n=9) · Lium direct RTX 3090 24GB ON_DEMAND_SINGLE P50 moved +22.2% ($0.18 → $0.22, n=2) · machine diff · ← issue 259 · all issues
GPUDECK OPS SOFTWARE FOR GPU OPERATORS
JOIN THE FOUNDING COHORT →
Free while in pilot · no card, no agent phoning home

Your rigs earn
while you sleep.
Or they don't.

If your GPUs earn by completing work — Sogni and other job-settled inference and render networks — the money leaks in ways your dashboard calls healthy. GPUDeck is the control plane that catches it: built by an operator, running nightly on a real production fleet.

Join the founding cohort →

NIGHTLY ON A LIVE FLEET SINCE DEC 2025 · 278 DAYS PUBLISHED · HOW IT WORKS

WHAT IT DOES

  • Restarts a worker between jobs, not mid-render. Memory climbs, and a worker the kernel kills mid-render takes the job with it — the container exits clean, so nothing tells you. The deck watches for an idle window and recycles there, so a restart costs idle seconds instead of landing on paid work.
  • Treats earning as the only truth. Checks verify at the payout layer, never the process layer. A green container that stopped earning is an incident.
  • Acts on incidents before you wake up. Lane down, model set wrong after a reboot, image update stalled — recovery runs first, the report arrives after.
  • Changes config on evidence. Memory limits, headroom, routing: rolled out as controlled comparisons, reverted automatically when the control wins.
See a real incident, log lines and all →
WHY TRUST IT

We publish our own fleet's numbers, in public, every day.

Most ops tools ask you to take their word for it. GPUDeck runs on our production fleet and the results are published as a dated, checksummed, citable record — including the days it looks bad. That record is the standard GPUDeck holds its own fleet to — and the standard we propose for customer fleets. No outside fleet has run the deck yet, so it is a proposal until the first one does. It is not a price comparison site, and it does not pretend to cover the market: it covers one fleet, measured honestly, plus the 5 rental sources we have a documented right to republish. Across our last 15 archived snapshots our cards finished 83–97% of the work they started — observed non-completion, not a claim about how much of it we can fix. See the record →

Who this is not for, yet. GPUDeck is for fleets paid for completed work. If you rent your cards out by the hour, or you mine, your earning loop is a different one — you are paid for delivered availability, and whether your renter’s job finished is neither yours to lose nor yours to inspect; a miner is paid per accepted share. We do not measure those loops, so we do not sell to them. That boundary moves when the measurement does, not before.

How the measurement works →
TODAY’S VALUE BOARD

RENTAL RATE × GPU-HOURS ÷ COMPLETED RENDERS

↓

COST PER FINISHED RENDER — WITHDRAWN

price side has no recorded legal basis

GOODPUT MEASUREMENTS UNAFFECTED AND STILL PUBLISHED

WHY IT IS WITHDRAWN →

BUILT BY AN OPERATOR · YOUR MACHINES, YOUR KEYS · NO AGENT PHONING HOME · NO SPONSORS

Not ready for the cohort? Get the field notes — what broke on a real fleet this week and what fixed it. No spam; unsubscribe is one reply.
REBUILD FROM CSV direct-tariffs.csv lium-price-context.csv goodput-ratios.csv census.csv value-board.csv methodology.md citation.cff