An H100 rents in three tiers: $1.69–$1.85/hr on open markets (RunPod spot, Vast.ai), $2.49–$2.89/hr at fixed-price clouds (RunPod on-demand, Lambda, CoreWeave), and $4.08 and up in hyperscale equivalents — with the whole-market median near $3.39. That is roughly seven times a 4090 hour, worth paying only when the job needs 80GB or NVLink-class interconnect. Here is each tier priced, and where that line sits.
01. What does an H100 cost per hour right now?
The 2026 market splits into three price tiers for the same chip.
gputracker's April 2026 survey — 54 providers, 5,269 GPU instances — lays the H100 ladder out in one table: open-market listings at the bottom, fixed-price clouds in the middle, hyperscale entries on top. The chip is identical at every step. What changes between tiers is who owns the machine, what contract stands behind it, and whether someone else can take the card back mid-job. Those three differences, not the silicon, produce the spread between $1.69 and $4.08.
| Tier | Provider and class | Hourly |
|---|---|---|
| Open market | RunPod spot | $1.69 |
| Open market | Vast.ai | $1.85 |
| Fixed-price cloud | RunPod on-demand | $2.49 |
| Fixed-price cloud | Lambda | $2.49 |
| Fixed-price cloud | CoreWeave | $2.89 |
| Hyperscale | AWS p5 class (per-GPU equivalent) | $4.08 |
Source: gputracker's lowest-price survey, April 2026 (54 providers, 5,269 instances). The full card-by-card floor-and-median table for every class lives on our GPU rental price comparison page.
The market tier starts just under $1.70, and the floor keeps moving.
RunPod spot at $1.69 and Vast.ai at $1.85 top gputracker's April list. By October, the aggregators' floor numbers sat in the same band: Inference Hub listed the H100 from $1.68/hr, gpurentalprices from $1.99 (October 16 data), with Rank.ai's open-market figure at $2.55. Vast.ai's own marketplace shows a $1.73 floor against a $3.06 median for H100 SXM 80GB — the figures quoted across this site. Every one of these is a floor with a date on it, not a booking price; a tier this liquid only quotes honestly with its source and month attached.
The fixed-price tier is the $2.49–$2.89 band.
RunPod's on-demand H100 and Lambda's both listed at $2.49/hr, CoreWeave's at $2.89, in gputracker's April survey. This is the tier most teams should price against: a fixed, single-tenant rate at a GPU cloud, no marketplace reclaim risk, and an invoice that matches the quote. The premium over the market tier buys a machine nobody takes back mid-run and a rate card that does not move under a training job. When a schedule or a deadline sits on top of the compute, this band is the honest number to plan with.
Hyperscale starts above $4 and climbs past $11.
AWS's p5 class enters gputracker's list at $4.08 per GPU-hour as an equivalent — the top of the April table. Higher up the same ladder, in the index Akash cites from GetDeploying, Azure's NC H100 v5 runs about $6.98 per card, CoreWeave's eight-card node about $6.16 per GPU, and GCP's a3 about $11.06 per GPU as an equivalent. Nobody books these for a weekend job. The hyperscale line exists for procurement that needs contracts, compliance paperwork, and a support desk, and its prices are procurement prices — quoted per card, per node, and per committed year.
02. Why does the same card span $2 to $12 an hour?
The whole market runs $2–$12 on demand, with the median near $3.39.
GetDeploying's index — as cited in Akash's H100 pricing review — puts the market's on-demand H100 between $2 and $12, with single-card mainstream between $2 and $7 and the median near $3.39. That spread is the three tiers of the last section stacked on one axis. When someone quotes you "the" H100 price, place it on this axis first: under $2 is market inventory, $2–$3 is fixed cloud, over $4 is hyperscale or an equivalent-class quote. A number without its tier is a number without a meaning.
The tier you buy is mostly a procurement decision.
The open-market tier prices interruption risk into its rate — the listing can be reclaimed, which is why it starts under $2. Fixed-price clouds charge for a machine nobody takes back. Hyperscale quotes carry contracts and support desks on top of compute. None of the three is a scam or a gift; each is the same silicon plus a different promise about who can pull the card away and who answers the phone when something breaks. Choose the promise your job can actually use, and the price sorts itself out.
Contract rates tell you where the floor is heading.
One-year H100 contract rates bottomed at $1.70/hr in October 2025 and have since risen by nearly 40% — the same Akash-cited index. Contracts are the market's wholesale price for promised hours, and their direction shows what bulk buyers are paying this quarter. For planning purposes, a commitment buys the market tier's price level with the fixed tier's predictability — provided the workload is steady enough to promise a year to, which most jobs are not.
03. When do you actually need an H100?
The first test is memory: the job must not fit in 24GB.
An H100 SXM carries 80GB of memory; an RTX 4090 carries 24GB. If the model, its gradients, and the pipeline around them fit inside 24GB, the H100's extra capacity is inventory you pay for by the hour and never touch. The jobs that clear this test are the familiar ones: full-parameter fine-tunes of models well past consumer-card size, long-context inference, batch configurations that only make sense with that much headroom. Memory is the first reason the card exists, and it is the cheapest test to run before renting anything.
The second test is interconnect: cards that must behave like one machine.
A 4090 has no NVLink — two 24GB cards in one box do not behave like one 48GB card, because they talk over PCIe. H100 systems are wired for multi-GPU training with NVLink-class interconnect between cards, and that wiring is what a large training run is actually buying when it scales past one card. If your job splits cleanly across cards over ordinary PCIe, or fits on a single card, the interconnect premium is idle wire on the invoice.
If the job fits on a 24GB card, the H100 premium buys you nothing.
Run the tests in order: memory, then interconnect, then price. The first two belong to the job; the price only matters after the job has cleared them. Our RTX 4090 vs A100 vs H100 comparison walks the card-by-card decision, so the tier you book is the tier the work needs — not the tier the search result advertised.
04. What are you actually buying at seven times the 4090 price?
The whole-market median is about seven times the 4090 median.
Place the two medians side by side: $3.39 across the whole H100 market (GetDeploying, cited by Akash) against $0.49 for the RTX 4090 on Vast.ai — call it seven times. The multiple is not a speed claim and not a markup; it is the price of a different machine class: 80GB instead of 24GB, NVLink-class wiring instead of a PCIe slot, and a chassis rated for sustained datacenter duty. Seven times is what that package costs when the market clears near its median — and it is the correct price only for the jobs that use the package.
At the floor you buy a gamble; at the median you buy a schedule.
The spread between tiers is a trade between price and terms. Book the market floor — RunPod spot at $1.69 — and you hold the same silicon at the market tier's rate, with the market tier's reclaim risk attached. Budget the whole-market median, $3.39, and the plan survives contact with a real booking. The H100 floor is a real price in exactly the way the 4090's floor is a real price; it is just not a number a schedule can be built on.
For training that needs checkpoints, the server — not the task — is the product.
SpareHive's plan is task-based image and video work with ephemeral storage, not a persistent training server. A multi-hour train that needs checkpoints belongs on a server you can store to. Join the waitlist only if your work is image and video: sparehive.com.
FAQ
Which providers top the lowest H100 prices?
RunPod spot at $1.69/hr and Vast.ai at $1.85/hr — the two market-tier entries on gputracker's April 2026 survey of 54 providers. By October, aggregator floors sat in the same band, from $1.68 (Inference Hub) to $1.99 (gpurentalprices).
Why pay $2.49 at Lambda when the market lists $1.69?
For the fixed tier's terms: a single-tenant machine nobody reclaims mid-run, at a rate that matches the invoice. The gap between the tiers is interruption insurance — you pay it to make reclaim risk somebody else's problem.
Is an H100 worth renting for a job that fits in 24GB?
No. Memory and interconnect are the two reasons the card exists, and a 24GB job uses neither. Until the job outgrows the card, the 4090 median — $0.49/hr — is the number to budget.
What does an H100 hour cost at the median?
Near $3.39 across the whole market (GetDeploying's index, cited by Akash). Vast.ai's own marketplace median is $3.06 for H100 SXM 80GB, with a $1.73 floor.