There is no single H100 price: public offers run from about $1.90 to $10 per GPU-hour before a single byte leaves the cloud

Disclosed H100 rates across 12 GPU cloud providers as of the 2026-08-27 research date, in USD per GPU-hour (Scaleway quotes EUR and is shown unconverted in the table). Effective cost = usable compute rate + paid egress GB per GPU-hour × egress $/GB after any monthly allowance — so no all-in number exists without an outbound-data assumption.

sticker / on-demand rate lowest disclosed spot, committed, promotional or marketplace rate ● on-demand listing  ◆ marketplace / variable  ▲ reservation / quota constrained marketplace minima are not guaranteed capacity
CoreWeave discloses only 8× HGX H100 instance prices — $49.24 on demand and $19.71 spot per instance — which normalize to $6.155 and $2.464 per GPU-hour, a 60% spot discount on the same silicon. coreweave.com
AWS sells H100s through Capacity Blocks reservations: the p5.48xlarge in the US East (Atlanta) Local Zone showed $41.528 per 8× instance ($5.191 per accelerator), with a $34.608 ($4.326) observation on another capture of the same page. aws.amazon.com
Egress splits the market: Oracle bundles 10 TB/month of free public-internet egress against its $10 list rate, while AWS standard internet egress of about $0.09/GB after 100 GB free adds ~$0.90/GPU-hour at a sustained 10 paid GB per GPU-hour. oracle.com · spendark.com
Marketplace floors are real but not guaranteed: Vast.ai H100 hosts were observed spanning roughly $1.05–$2.75/hr (about $2.21 on verified datacenter hosts) in July 2026, with bandwidth pricing set per host rather than platform-wide. runpod.io · rfp.wiki

Every offer, with the egress fine print

The all-in column is an illustration at 10 paid outbound GB per GPU-hour (a scenario, not a universal workload): compute rate + 10 × egress $/GB after allowance. Where a price or fee is not publicly resolved the cell is left blank. 8× instance totals are shown as totals and never plotted as one-GPU pricing.

ProviderH100 variant / SKU Sticker $/GPU-hDiscount rate & type Egress fee & allowanceAll-in @ 10 GB/h AvailabilitySource
Oracle Cloud (OCI)BM.GPU.H100.8, 8× H100 80GB, per-GPU list$10.0010 TB/mo free, then paid$10.00On-demand list; pay-as-you-gooracle.com
CoreWeaveHGX H100 8× ($49.24 / $19.71 per instance ÷ 8)$6.155$2.464 spotFree (secondary source)$6.155On-demand + spot, North Americacoreweave.com
PaperspaceH100 80GB$5.95Free on Machines (secondary)$5.95On-demand listingpaperspace.com
AWSp5.48xlarge Capacity Blocks, $41.528 per 8× instance$5.191$4.326 Capacity Block (lower observation)~$0.09/GB after 100 GB/mo free (secondary)$6.09Reservation only (Capacity Blocks)aws.amazon.com
NebiusHGX H100 SXM$3.85$2.15 spot / preemptibleCompute egress free; object storage $0.015/GiB$3.85On-demand + preemptiblenebius.com
HyperstackH100 SXM $3.20; H100 PCIe $1.90$3.20$1.90 PCIe variant, on-demandFree (secondary)$3.20On-demand; reserving advised for availabilityhyperstack.cloud
RunpodH100 SXM Secure Cloud $2.99; PCIe Community $1.99$2.99$1.99 Community Cloud PCIe (spot-equivalent)No ingress/egress fees on pods$2.99Secure on-demand; Community variablerunpod.io
Vast.aiH100 marketplace, verified DC hosts ≈$2.21$2.75$1.05 marketplace low (observed, Jul 2026)Host-set: free to ~$0.04/GBMarketplace; host-dependent, not guaranteedrunpod.io
Microsoft AzureND H100 v5, East US, reserved $20.49/hr per 8× = $2.561/GPU-h$2.561 reserved (secondary tracker)~$0.087/GB after 100 GB/mo free (secondary)$3.43Reservation / quota / region constrainedgridstackhub.ai
ScalewayH100 PCIe 80GB€2.73 (EUR, not converted)Egress included in instance priceOn-demand; datacenter-dependentscaleway.com
TensorDockH100 SXM5 80GB, marketplace typical$2.25$2.20 host quote (Austin, TX)No ingress/egress fees$2.25Marketplace; varies by host, no quotastensordock.com
FluidStackH100 SXM$2.10Zero fees reported (secondary)$2.10On-demand listingfluidstack.io

Method: H100 pricing evidence gathered across four result sets (roughly 770 rows of pages, provider price captures, hyperscaler material and egress rates), research date 2026-08-27; deduplicated to one current offer per provider, preferring provider-owned pricing pages over blogs and trackers. A rate measures the disclosed public price per H100 GPU-hour in the quoted currency; 8-GPU instance totals were divided only where the per-GPU figure appears in the source. The all-in column applies compute rate + 10 GB/h × egress $/GB after allowance; egress fees resolved only by secondary sources are labelled and missing data is left blank, not treated as zero. Cut for space: dozens of aggregator price captures, older Hyperstack/Nebius/Runpod price points, Nebius $1.50 Explorer promo (first 1,000 h/mo), and Lambda/Google Cloud offers without a resolvable per-GPU price in the extraction.

Keenable · made with SELECT* · 9,738 pages in 6m 38s · Open the chatShare: XLinkedInReddit