H100s really cost $1.4–$2.4 an hour at neoclouds; hyperscaler “$0.47” deals hide waitlists and egress

The question: what is the real effective price per hour for an NVIDIA H100 across GPU cloud providers — sticker price, spot or committed discounts, egress fees, and current availability?

13 providers, prices normalized to one H100 GPU-hour in USD where GPU counts were known, from pages scanned through 30 Aug 2026, plus a primary-source check of Lambda and Nebius pricing pages. Collected rates span $0.47 (AWS 1-yr committed, secondary claim) to $5.95 (Paperspace sticker).

Price range per provider: sticker, spot, committed — cheapest option marked

$ per H100 GPU-hour · sorted by cheapest compute option · hover a row for detail, egress and sources

sticker → cheapest range sticker (on-demand) spot / interruptible committed / reserved ✓ primary-source verified ⚠ secondary-only extreme — verify before purchase

Self-service neoclouds and marketplaces cluster at $1.4–$2.4/H100-hour — Vast.ai $1.38, Azure spot $1.40, TensorDock spot $1.60, DataCrunch spot $1.65, CoreWeave committed $1.79, RunPod $1.89, Lambda $1.99, Nebius $2.00 — with easier access than hyperscaler quota. gpusmith.com

The two extreme lows are secondary-source claims: AWS $0.473 (1-yr committed, vs $3.83 spot and a months-long waitlist) and Oracle $1.40 (annual commitment against a $10.00 PCIe sticker, 86% off). Verify exact instance, GPU count, region and term before relying on them. gpuaas.com, deploybase.ai

Egress is resolved for only 3 of 13 providers: Lambda $0/GB (explicitly free, primary source), RunPod $0/GB, AWS $0.08/GB — at 100 GB out per GPU-hour AWS's modeled total jumps to $8.47. Blank egress means unresolved, not free; those rows are compute-only floors. lambda.ai

Availability is the hidden price: AWS reports a months-long waitlist, Azure "very limited", Google Cloud "limited quota", while Lambda advertises self-serve launch in minutes — a point-in-time signal from the scanned pages, not a capacity guarantee. runpod.io

Primary source ✓

Lambda — $3.99 per H100-SXM GPU-hour, egress free

Lambda's own instance page lists NVIDIA H100 SXM at $3.99/GPU/hr (80 GB VRAM, 208 vCPU, 1,800 GiB RAM, 22 TiB SSD) with "transparent pricing with no egress fees" and self-serve, pay-by-the-minute billing. The $1.99 figure in the comparison set is a monthly-commitment (730 hrs) rate from a secondary comparison. lambda.ai

Primary source ✓

Nebius — $3.85 on-demand, $2.15 preemptible (eu-north1)

Nebius's pricing docs list NVIDIA H100 NVLink at $3.85 per GPU-hour on demand and $2.15 preemptible in eu-north1 (updated 19 Aug 2026). The comparison set's $2.00 sticker for Nebius conflicts with this primary source — treat $3.85/$2.15 as the documented rates. docs.nebius.com

All 13 providers — click a header to sort

Effective hourly cost = selected compute rate + (outbound GB per GPU-hour × egress $/GB). Excludes storage, public IPs, taxes and idle time. Where egress is unresolved the 10/100-GB columns add $0 and the total is a compute-only floor, not all-in.

Provider	Sticker	Spot	Committed	Cheapest	Disc %	Egress $/GB	Eff @10GB	Eff @100GB	Availability	Source

Buyer checklist

Confirm the H100 variant (SXM vs PCIe vs NVL/NVLink) and node CPU/RAM/storage — the rows here mix them and they are not interchangeable for training throughput.
Get the egress rate in writing; a blank cell here means unresolved, not free. Model your own GB-out per GPU-hour before comparing totals.
Separate interruptible from firm capacity: spot/preemptible rates (orange) can be reclaimed mid-run; committed rates (green) bind you to a term.
Ask for current quota and delivery date — advertised availability was a point-in-time signal from pages scanned through 30 Aug 2026.
Re-verify the two extreme lows (AWS $0.473, Oracle $1.40) directly with the provider for exact instance, GPU count, region and commitment.

Data: 13 provider rows aggregated from 107 scanned pricing pages and comparison articles (latest source dates 22 Jun – 19 Aug 2026, scan window through 30 Aug 2026), plus 3 primary-source rows from Lambda and Nebius pricing pages. Numbers are USD per one NVIDIA H100 GPU-hour, normalized where a GPU count was available. Effective totals model internet egress only; storage, IPs, taxes and idle time excluded. Availability text and long marketplace notes were shortened for space; regions, minimum terms and interruption risk differ per provider and were not normalized.