Specialist clouds beat hyperscalers on H100 cost; egress widens the gap
A live scan of public and provider pricing found the lowest usable headline rates in marketplace or specialist clouds. “Effective” cost is not one universal number: it is compute price plus outbound data multiplied by the workload’s GB per GPU-hour, and spot prices also carry interruption and capacity risk.
$0.50
Lowest observed H100 GPU-hour
$1.25
Lowest preemptible specialist rate
$0
Egress at several GPU specialists
12
Providers checked in final set
Best public rate found, USD per H100 GPU-hour
Blue: provider/contract · light blue: marketplace
What changes the ranking
1
Zero-egress specialist clouds protect the headline rate. CoreWeave explicitly states zero ingress/egress fees, while Lambda and Runpod advertise no egress fees. coreweave.com
2
3
Commitment can reverse “sticker” comparisons. CoreWeave’s $2.23 rate is tied to a two-year contract; Nebius’s $1.25 is preemptible; these are not on-demand equivalents. nebius.com
4
Availability is part of effective cost. Quotas, capacity blocks, first-come allocation and unavailable SKUs can turn a low published rate into no usable capacity.
Provider detail
| Provider / H100 offer | Sticker | Best discount found | Internet egress | Current availability signal | Source |
|---|---|---|---|---|---|
| Vast.ai · SXM marketplace | $0.50 | $0.50 spot | Host-specific / not extracted | Thousands listed across 30+ data centers | vast.ai |
| Nebius · HGX H100 | $2.95 | $1.25 preemptible | Not extracted | Advertised as ready without waitlists | nebius.com |
| Runpod · H100 PCIe | $1.99 community | $1.52 observed community offer | No egress advertised on comparison page | Dynamic by host and region | runpod.io |
| Lambda · H100 SXM | $1.49 | Committed clusters quoted separately | $0 advertised | First-come access | lambda.ai |
| Hyperstack · H100 SXM | $2.40 | $1.90 reserved | Free advertised | On-demand access advertised | hyperstack.cloud |
| CoreWeave · HGX H100 | Sales quote | $2.23, two-year contract | $0 advertised | “Available at supercomputer scale” | coreweave.com |
| Oracle · BM.GPU.H100.8 | $2.50/GPU-h found | None verified | Not extracted | Capacity/limit approval not verified | oracle.com |
| Crusoe · H100 80GB SXM | $3.90 | None public found | Not extracted | 16 GPUs shown in one current capacity signal | crusoecloud.com |
| AWS · P5 capacity block | On-demand varies by P5 size/region | $3.933 capacity block | Paid internet egress; region/tier dependent | At-scale on-demand advertised; quotas apply | aws.amazon.com |
| Google Cloud · A3 | Instance/region dependent | Spot and CUD available | Paid internet egress; destination/tier dependent | Quota and regional capacity constraints | cloud.google.com |
| Microsoft Azure · ND H100 v5 | Region dependent | Spot/reserved available | Paid bandwidth; zone/tier dependent | General availability, quota/capacity dependent | microsoft.com |
| DataCrunch · 8× H100 SXM | $2.29/GPU-h | Contract quote; low extracted outlier excluded | Not extracted | SKU page showed unavailable in scan | datacrunch.io |
How to calculate your real rate: effective USD/GPU-hour = compute rate + (internet-egress USD/GB × outbound GB per GPU-hour) + storage/CPU/network extras. Add an interruption/restart penalty for spot. A provider with an unverified egress cell should not be assumed free.
Method: live web scan on 21 Aug 2026; 12 final providers, using provider pages where available and cross-provider search results for discovery. Prices are normalized per H100 GPU-hour where the GPU count was stated. H100 PCIe and SXM/HGX, on-demand, marketplace, spot and committed capacity are labeled because they are not performance- or risk-equivalent. Dynamic offers can change minute to minute.