RESEARCH INDEX / GPU CLOUD TCO
GPU cloud infrastructure & TCO teardowns
Catalogue rates are component prices. These reports convert them into dollars per useful GPU-hour by adding bundle minimums, idle capacity, storage, egress and restart overhead — the costs that decide whether a cheap listing is actually cheap.
Published by 600me2 long-form reportsLive catalogue observationsBaselines dated 2026-10-18
2Deep-dive TCO reports
2×Minimum H100 bundle
+46%Headline vs effective rate
Published research
MOST RECENT FIRST
GPU cloud cost notes / head-to-head
Vast.ai lists H100 SXM about 25% below RunPod and still loses a two-hour test by 63% once 150 GB of ingress, a 40-minute pull and a 15% failover reserve are priced in. The same arithmetic reverses on a 72-hour preemptible fine-tune. Six billing dimensions and three workloads, all in dollars per useful GPU-hour.
- Scenario A: Vast $8.86 vs RunPod $5.42 for one test
- Scenario B: RunPod wins above 5.4 TB/month egress
- Scenario C: Vast saves $314.52 on a 72-hour run
GPU cloud cost notes / rental trap
The $2.134 H100 SXM headline is attached to a two-GPU bundle priced at $4.269/hour. An eight-hour fine-tune that fits on one card still buys 16 GPU-hours. This teardown walks through the smallest rentable unit, the 48.3% community-to-secure premium, utilization below 100%, and the one incident that can erase a 20.7% saving.
- $2.134 headline → $4.269 per useful GPU-hour
- 60% utilization → $3.56 per useful hour
- 45-min outage: $48.75 of a $55.60 saving
How to read these numbers. Rates are catalogue observations captured on 2026-10-17, not negotiated quotes. Inventory is not reserved, and storage, network, taxes and restart labour are excluded from headline rates. The formulas and assumptions are printed in the article so you can substitute your own utilisation and interruption data.