BlogPricing

The cheapest GPU cloud providers in 2026, ranked

By Compute CheapPrices checked August 15, 2026

Every major GPU cloud’s H100, H200, and Blackwell price per GPU-hour, ranked from cheapest to most expensive.

01 / The short answer

We’re the cheapest, at $1.19 per H100-hour and $1.99 per H200-hour.

As of August 15, 2026, we — Compute Cheap — have the lowest published GPU-hour prices of any cloud in this table: NVIDIA H100 at $1.19 reserved ($1.38 interruptible) and H200 at $1.99 reserved ($1.67 interruptible), and we list B200 at $3.10 and GB300 at $3.39, with GB300 currently sold out. The next cheapest guaranteed H100 is $1.68, and the hyperscalers list $6.88 to $12.29.

Below, 19 providers ranked cheapest first on the non-preemptible price you can plan around — pick the GPU — then a few plain words on each one, in H100 order.

02 / The ranking

19 GPU clouds, cheapest first.

#	Provider	Interruptible	On-demand	Reserved
1	Compute Cheap	$1.38	—	$1.192,000 GPU-hr block
2	Latitude.sh	—	$1.68	$1,230/momonthly
3	SF Compute	≈0.5× market	$1.95 avg	orderbook
4	Voltage Park	—	$1.99	contact6+ months
5	Vast.ai	50%+ cheaper (per host)	$1.33–$2.00	up to 50% off1–6 months
6	Prime Intellect	$0.94 listing	$2.43	—
7	Verda (formerly DataCrunch)	$1.63	$3.25	$2.4424 months
8	RunPod	~50% off (spot pods)	$2.69 / $3.29	3–6 mo plansdiscount unpublished
9	Lyceum	$1.10	$2.79	—
10	Together AI	$1.99 preemptible	$3.99	$3.1991–180 days
11	Hyperbolic	—	from $3.19	discounted, upfront
12	Thunder Compute	—	$3.20 (PCIe)	—
13	Nebius	$2.15	$3.85	up to 35% offmulti-month clusters
14	Modal	—	≈$3.95	—
15	Lambda	—	$3.99–$4.29	$5.54–$6.161-Click Cluster, 2 wk–1 yr
16	AWS	$2.59	$6.88	$5.19Capacity Block
17	CoreWeave	$2.46	$6.16	up to 60% offmulti-year
18	Google Cloud	—	$10.98	CUD, unpublished
19	Microsoft Azure	$2.27	$12.29	unpublished

03 / Provider by provider

The providers, in order.

1. Compute Cheap — That’s us. The lowest guaranteed rate on the list for every GPU: $1.19 per H100-hour reserved and $1.15 interruptible, H200 at $1.99 and $1.39, and B200 at $3.10 and GB300 at $3.39 (GB300 currently sold out), in the United States and European Union, with a 99.9% uptime SLA on reserved capacity. We sell capacity in prepaid blocks of 2,000 GPU-hours or more, confirmed before you pay.

2. Latitude.sh — Bare-metal servers billed hourly. The $1.68 rate is a single-GPU H100 server; there is no multi-GPU H100 node on the rate card and no H200; an 8× B300 node is $64 an hour.

3. SF Compute — A live orderbook where you bid for whole 8-GPU nodes by the hour. The $1.95 figure is the site’s average clearing price, not a rate you can lock, and the SLA sits behind a login.

4. Voltage Park — A US-only owned cloud with $1.99 on-demand H100s on Ethernet networking. InfiniBand nodes, H200, and every Blackwell SKU are quote-only.

5. Vast.ai — A peer-to-peer marketplace with per-second billing. H100 listings ran from $1.33 to a $2.00 median on the day; each host sets its own price, uptime, and terms.

6. Prime Intellect — An aggregator of 50+ clouds paid from one wallet. H100 listings start at $2.43 on-demand, with spot offers as low as $0.94, H200 listings anywhere from $0.47 to $3.14, B200 at $3.49 and B300 at $4.99; hardware and reliability depend on whichever provider posted the listing.

7. Verda (formerly DataCrunch) — An EU-based owned cloud, formerly DataCrunch. $3.25 on-demand and $1.63 spot for H100, reserved discounts of 2–25% for terms up to two years, and H200, B200, B300, and GB300 all on the rate card.

8. RunPod — Per-second billing with no minimum. Community Cloud (peer hosts) is $2.69 for an H100 SXM; Secure Cloud in T3/T4 datacenters is $3.29. H200 is $3.59 and $4.59 on the same two tiers, B200 from $5.98.

9. Lyceum — An EU-only cloud (Denmark, France, Germany) with per-second billing: $2.79 on-demand and $1.10 spot for H100, $4.29 and $1.60 for H200. Its pricing page blocks crawlers, so these figures come from GetDeploying.

10. Together AI — Sells whole 8-GPU nodes: $3.99 on-demand, $3.19 for a 91–180-day reservation paid upfront, and $1.99 preemptible. H200 from $5.99, B200 from $8.19.

11. Hyperbolic — A marketplace with prices from rotating suppliers, refreshed weekly and billed hourly. H100 from $3.19, H200 from $3.99, B200 from $5.99.

12. Thunder Compute — Virtualized H100 PCIe instances at $3.20, billed per minute with no egress fees. No SXM, no NVLink, no H200.

13. Nebius — $3.85 on-demand and $2.15 preemptible for H100, $4.50 and $2.45 for H200, billed per second. H100 capacity is in Finland only; US regions carry H200 and B200.

15. Lambda — $3.99 on-demand for 8× H100 SXM ($4.29 for a single GPU), billed per minute. InfiniBand clusters run $5.54–$6.16 with a 16-GPU, two-week minimum; H200 is not on the price page.

16. AWS — p5 instances at $6.88 on-demand and $2.59 spot per GPU, or $5.19 through a Capacity Block reserved in advance; H200 (p5en) blocks at $6.87. Egress is billed on top.

17. CoreWeave — Sold by the 8-GPU node: $6.16 on-demand and $2.46 spot per H100, $6.31 and $2.62 per H200, B200 at $8.60. Committed contracts get up to 60% off; since April 2026 the minimum term is three years.

18. Google Cloud — a3-highgpu-8g at $10.98 per GPU on-demand, the highest list price in this table. Newer A3 tiers are sold on commitment only.

19. Microsoft Azure — ND96isr H100 v5 at $12.29 per GPU on-demand, with spot at $2.27 — roughly ten times the cheapest guaranteed rate in this table.

04 / Company

About us.

Frontier labs used capital and scale to make intelligence cheap. Nobody did the same for the compute underneath it, so we did. We’re Compute Cheap, based in San Francisco, and we sell NVIDIA H100, H200, and B300 interruptible capacity by the GPU-hour, in the United States and European Union. Our prices come from competition between capacity owners, distributed supply, and a leaner stack with no platform fee, which is why the first row of this table is ours. Reserved blocks carry our 99.9% uptime SLA, and you pay nothing until we’ve confirmed availability. Here’s our story.

05 / FAQ

Common questions.

What is the cheapest GPU cloud in 2026?

We are — Compute Cheap. Our H100 capacity is $1.19 per GPU-hour reserved and $1.15 interruptible, H200 is $1.99 and $1.39, and we list B200 and GB300 at $3.10 and $3.39 (GB300 currently sold out) — the lowest published non-preemptible rates of any provider in this table for every GPU. The next cheapest guaranteed H100 is Latitude.sh at $1.68 for a single-GPU server; hyperscalers list $6.88 to $12.29.

Why do H100 prices range from about $1 to over $12 per hour?

Whether the GPU can be reclaimed (spot and community tiers are cheapest), whether you are buying one GPU or a whole 8-GPU node, how long you commit for, and who the counterparty is — a peer host on a marketplace, a cloud that owns its hardware, or a hyperscaler bundling networking, support, and a brand premium into the rate.

Is spot or interruptible capacity safe for training?

Yes, if the job checkpoints. Interruptible capacity can be reclaimed, so it suits fault-tolerant training and batch inference. For a serving fleet or a deadline, reserve non-preemptible capacity — with us that is $1.19 per H100-hour with a 99.9% uptime commitment.

Are hyperscalers ever the cheapest option?

Only on spot: AWS ($2.59) and Azure ($2.27) land near neocloud on-demand pricing, and startup credits can make them free for a while. On list price they are four to ten times the cheapest guaranteed rate, and egress is billed on top.

What does per-node pricing mean?

Some providers only sell servers with eight GPUs attached. CoreWeave’s $49.24 per hour is $6.16 per GPU-hour, and you pay for all eight whether your job uses one or eight. Per-GPU pricing lets you rent one, two, or four.

How often do these prices change?

Often. Marketplace listings move hourly and the H100 index moved 11% in one quarter. We re-check and re-date this table; our own prices are published live through the pricing API.