Nebius delivers the cheapest verifiable H100 at $1.75 all-in — a fifth of a hyperscaler's delivered price
Asked:
“What is the real effective price per hour for an NVIDIA H100 across GPU cloud providers? Include the sticker price, spot or committed discounts, egress fees, and current availability for each provider.”
Eleven GPU cloud providers, snapshot 2026-09-05. Prices are normalized to one H100 GPU-hour in USD; the effective figure adds 10 GB of public-internet egress to the lowest verified public compute price. Four marketplace providers publish no comparable egress rate, so they carry no defensible all-in number.
Sticker (on-demand)
Lowest public spot / commit
Effective all-in at 10 GB egress/GPU-h
Egress not publicly comparable — no all-in figure
Nebius is the cheapest computable all-in at
$1.75/GPU-h — a $1.25 committed compute figure plus $0.50 of egress; verify term and region.
nebius.comAWS Spot cuts a $6.88 sticker to about
$2.96 all-in ($2.06 compute + $0.90 egress) — but spot is interruptible and region-gated.
deploybase.aiLambda's $2.86 PCIe on-demand stays
$2.86 all-in within its reported 1 TB/month free-egress allowance.
deploybase.aiVast.ai ($1.47), RunPod ($1.59 midpoint), Crusoe (~$2.17) and TensorDock ($3.66) post cheap compute headlines, but none published a comparable egress rate — their delivered cost is unverifiable here.
deploybase.aiHyperscaler on-demand stays expensive delivered: Oracle
$10.00 (within its 10 TB free outbound), Azure
$11.93, Google Cloud
$12.26 — Google's own docs call A3 H100 capacity limited and point to account teams.
docs.cloud.google.com
Buyer takeaway. Egress-light or in-cloud training favors low-compute marketplaces and discounted specialists; sustained export-heavy inference can erase a headline discount at $0.09–0.12/GB. Where secure capacity and SLAs matter, committed hyperscaler or specialist capacity can be worth more than the lowest nominal price. Confirm region, term, quota, and live inventory before procurement — every figure here is indicative.
All eleven providers, row by row
| Provider | Offer basis | Sticker $/GPU-h | Lowest discounted $/GPU-h | Egress $/GB & caveat | Effective $/GPU-h | Discount detail | Availability |
|---|
Snapshot checked 2026-09-05. 11 providers, one row each; prices normalized to one H100 GPU-hour in USD. “Effective” = lowest verified public compute option (spot/commit when available, otherwise sticker) plus 10 GB public-internet egress in that hour — a comparison benchmark, not a workload assumption. Free allowances honored when stated; null means no comparable public egress price, so no all-in figure. Excluded: taxes, persistent disks, CPU/RAM add-ons beyond bundled instance resources, orchestration, idle time. Spot is interruptible; committed prices can impose duration, prepayment, or capacity constraints. H100 form factors differ (PCIe, SXM, NVL) — see offer basis. Some rows use official provider pages; others independent trackers where no official public rate existed. All figures indicative.