Budget stability
A low listing today may be transient. Compare only historically supported published-rate ranges for exact GPU variants, one billing mode and one observation window at a time.
Cross-GPU evidence table
Default · 90 days · on-demand · USD/GPU-hourNo series passes the 90-day on-demand evidence gate. Numerical rankings, charts, averages, and cheapest conclusions are withheld.
| Exact GPU variant | Evidence | p25–p75 | p50 | Floor persistence | Coverage / longest gap | As of |
|---|---|---|---|---|---|---|
| H100 SXM5 80GBDatacenter GPU | InsufficientNo materialized validation record | Withheld | — | — | Not yet validated | — |
| H100 PCIe 80GBDatacenter GPU | InsufficientNo materialized validation record | Withheld | — | — | Not yet validated | — |
| H100 NVLDatacenter GPU | InsufficientNo materialized validation record | Withheld | — | — | Not yet validated | — |
| H200 SXM 141GBDatacenter GPU | InsufficientNo materialized validation record | Withheld | — | — | Not yet validated | — |
| H200 NVLDatacenter GPU | InsufficientNo materialized validation record | Withheld | — | — | Not yet validated | — |
| A100 SXM 40GBDatacenter GPU | InsufficientNo materialized validation record | Withheld | — | — | Not yet validated | — |
| A100 PCIe 40GBDatacenter GPU | InsufficientNo materialized validation record | Withheld | — | — | Not yet validated | — |
| A100 SXM 80GBDatacenter GPU | InsufficientNo materialized validation record | Withheld | — | — | Not yet validated | — |
| A100 PCIe 80GBDatacenter GPU | InsufficientNo materialized validation record | Withheld | — | — | Not yet validated | — |
| B200Datacenter GPU | InsufficientNo materialized validation record | Withheld | — | — | Not yet validated | — |
| L40SDatacenter GPU | InsufficientNo materialized validation record | Withheld | — | — | Not yet validated | — |
| RTX 4090Consumer GPU | InsufficientNo materialized validation record | Withheld | — | — | Not yet validated | — |
| RTX 5090Consumer GPU | InsufficientNo materialized validation record | Withheld | — | — | Not yet validated | — |
Common window end: no materialized analysis. Records from other end dates are excluded from this view, not combined.
What this comparison can support
A passing p25–p75 band describes observed published rates, not a forecast or guaranteed budget. It does not establish availability, deployment success, performance, topology, reliability, negotiated cost, storage or egress charges, or total workload economics. Consumer and datacenter rows are labels, not claims of equivalent capability.
Evidence gate
The 30-, 90-, and 180-day windows require 24, 72, and 144 observed days, with maximum gaps of 3, 7, and 14 days. Every passing series also requires three independent providers on 80% of observed days, a live observation within 72 hours of the common window end, exact variant and billing classification, cleared provenance and redistribution rights, and no archive contribution. High confidence tightens coverage to 90%, provider count to four on 80% of observed days, and gaps to 2, 5, or 10 days.
Missing days are not interpolated, future rates are not extrapolated, and current listings do not substitute for history. See the versioned method and record contract and the H100 validation record.
Artifact acceptance record
- Pass — fail-closed comparison: the default 90-day on-demand view retains every approved exact variant, exposes validation failures, and produces no cross-GPU ranking while fewer than two series pass.