How much does it cost to rent an H100 per hour?

Between $1.99 and $11.06 per GPU-hour on published on-demand prices read on 13 Oct 2026: Voltage Park was the cheapest and Google Cloud (A3, A4) the dearest, 5.6× apart for the same chip. Spot and marketplace listings go lower but can be interrupted or depend on someone else's machine, and committed or negotiated rates are lower than every on-demand price — nobody publishes those.

Why is the same GPU so much cheaper at one company than another?

Because you are not only buying the chip. The cheapest end is often brokered or resold capacity, priced per GPU with everything else extra. The dearest end is a hyperscaler list price for a whole 8-GPU machine with its CPUs, memory and storage, which exists mainly to make the committed rate look like a discount. Between them sit differences in support, networking, region, billing and contract terms — and in who the counterparty is.

Does Macrostack measure these prices itself?

No. We do not rent clusters and we do not benchmark them. Every figure is a price the company publishes, read on its own price page on the date in its row, with a link to that page. Each row says how its number was made.

How often are the prices checked?

Every 14 days. A row older than 30 days says how old it is, and the site will not build with an on-demand price older than 45 days.

Can I reuse this data?

Yes, under CC BY 4.0, commercial use included. The one condition is credit with a link: Macrostack (2026). GPU rental prices, read 13 Oct 2026. CC BY 4.0. https://www.macrostack.net/compute/gpu-prices

GPU cost calculator · prices read 13 Oct 2026

What will renting a GPU cost you this month?

Name: Macrostack GPU rental prices
Creator: Macrostack
Published: 2026-09-09
License: https://creativecommons.org/licenses/by/4.0/

Pick a chip, how many, and for how long. You get the monthly bill at every company that publishes an on-demand price for it — 23 companies across 8 chips, each price read on the company’s own price page and linked to it. Cheapest first, ordered by price and nothing else. The widest gap today: the same H100 SXM rents for $1.99 to $11.06 per GPU-hour — 5.6×.

Your month · H100 SXM × 1 GPU × 100 h

Same chip, 5.6× apart.

An NVIDIA H100 SXM 80GB, 1 GPU, 100 hours a month: $199 at Voltage Park and $1,106 at Google Cloud (A3, A4). 23 companies publish an on-demand price for it; every one is below, cheapest first, ordered by price and nothing else.

We read these prices on each company’s own price page. We do not measure them.

Committed, reserved and negotiated rates are lower than anything shown here.

On-demand price for an NVIDIA H100 SXM 80GB at every company that publishes one, cheapest first, with the monthly bill for 1 GPU × 100 hours.
#	Company · your month · per GPU-hourCompany	Per GPU-hour	Your month	Price page
1	1. Voltage Park$199 $1.99 per GPU-hour The price cards say 'Contact for pricing'. The figure is the page's own FAQ: one H100 for an hour, no contract. On-demand · US (company is US-only; region not stated on page) · checked 12 Oct 2026 · voltagepark.com ↗ How this number was made Already per GPU 'starting at $1.99' for one H100 GPU for 1 hour without a contract (FAQ); cards say 'Contact for pricing'. GPUs are HGX H100	$1.99	$199	voltagepark.com ↗
2	2. Vultr$230 $2.30 per GPU-hour On-demand · not stated (FAQ: some locations priced differently; per-location prices are in the Control Panel) · checked 12 Oct 2026 · vultr.com ↗ How this number was made Already per GPU (8-GPU HGX H100, 640 GB GPU RAM) 'Price starting at' $2.300/GPU/hr, Cloud GPU (VM), billed hourly	$2.30	$230	vultr.com ↗
3	3. GMI Cloud$260 $2.60 per GPU-hour The price card does not name its tier. Recorded as the on-demand starting rate — flagged, not certain. On-demand · not stated on page · checked 12 Oct 2026 · gmicloud.ai ↗ How this number was made 'from $2.60 /GPU-hour'; tier not labelled on the card (see company note); listed as 'NVIDIA H100 GPU', form factor not stated	$2.60	$260	gmicloud.ai ↗
4	4. Civo$299 $2.99 per GPU-hour On-demand · not stated on page · checked 12 Oct 2026 · civo.com ↗ How this number was made 1-GPU instance; 8-GPU instance $23.92/hr ÷ 8 = $2.99 (identical) Per hour, 1 x H100 80GB (H100 SXM table)	$2.99	$299	civo.com ↗
5	5. OVHcloud$299 $2.99 per GPU-hour H100 80GB (h100-380: 1 GPU; form factor not stated) On-demand · GRA9 / GRA11 (Gravelines, France); not offered in OVHcloud US regions · checked 12 Oct 2026 · ovhcloud.com ↗ How this number was made h100-380 $2.99/hr, 1 GPU, no division. USD, excl. VAT, hourly, from the international page. h100-760 (2x) $5.98 and h100-1520 (4x) $11.97 are the same per GPU (within $0.01).	$2.99	$299	ovhcloud.com ↗
6	6. Hyperstack (NexGen Cloud)$320 $3.20 per GPU-hour On-demand · not stated on page · checked 12 Oct 2026 · hyperstack.cloud ↗ How this number was made Already per GPU Per hour per GPU (H100 SXM 80 GB); billed per minute	$3.20	$320	hyperstack.cloud ↗
7	7. JarvisLabs$349 $3.49 per GPU-hour H100 SXM 80GB On-demand · not stated (India regions mentioned) · checked 12 Oct 2026 · jarvislabs.ai ↗ How this number was made none — page prices per GPU On-demand list price, per GPU-hour, per-minute billing. Read 20:44 UTC (16:44 EDT).	$3.49	$349	jarvislabs.ai ↗
8	8. Crusoe$390 $3.90 per GPU-hour On-demand · not stated on page · checked 12 Oct 2026 · crusoe.ai ↗ How this number was made Per GPU-hour (H100 80GB HGX)	$3.90	$390	crusoe.ai ↗
9	9. Modal$395 $3.95 per GPU-hour Nvidia H100 SXM5 Base price: GPU time only, preemptible by default. Non-preemptible runs cost 3x; CPU and memory are extra. On-demand · Not pinned: base price lets Modal place the job; region selection costs 1.15-1.75x · checked 12 Oct 2026 · modal.com ↗ How this number was made Per GPU per second x 3600, no division. Per-second billing: $0.001097/sec x 3600. GPU only; CPU/memory extra. Base (preemptible) rate; non-preemptible = 3x.	$3.95	$395	modal.com ↗
10	10. RunPod$399 $3.99 per GPU-hour H100 SXM 80GB On-demand · not stated (30+ regions) · checked 13 Oct 2026 · runpod.io ↗ How this number was made none — page prices per GPU Secure Cloud Pod — RunPod's own on-demand list price per GPU-hour, billed per second; ranked from 13 Oct on Abdul's call (one rule for every company: a published hourly price is ranked). Re-read 13 Oct 2026, ~22:35 EDT (14 Oct ~02:35 UTC), on runpod.io/pricing with the Secure Cloud toggle (the default); unchanged since 9 Oct.	$3.99	$399	runpod.io ↗
11	11. Together AI$399 $3.99 per GPU-hour NVIDIA HGX H100 On-demand · Not stated on the page · checked 12 Oct 2026 · together.ai ↗ How this number was made Printed per GPU per hour, no division. GPU Clusters, on-demand, pay as you go; hourly.	$3.99	$399	together.ai ↗
12	12. Lambda$399–$429 $3.99–$4.29 per GPU-hour On-demand · not stated on page · checked 12 Oct 2026 · lambda.ai ↗ How this number was made Already per GPU. By instance size: 8x $3.99, 4x $4.09, 2x $4.19, 1x $4.29 Per GPU per hour, plus applicable sales tax/VAT/GST (H100 SXM 80GB)	$3.99–$4.29	$399–$429	lambda.ai ↗
13	13. DataCrunch (now Verda)$405 $4.05 per GPU-hour On-demand · not stated on page · checked 12 Oct 2026 · verda.com ↗ How this number was made Already per GPU (1x instance) Per hour, 1x H100 SXM5 80GB (two configs listed, same price)	$4.05	$405	verda.com ↗
14	14. Paperspace (DigitalOcean)$441 $4.41 per GPU-hour On-demand · not stated on page · checked 12 Oct 2026 · digitalocean.com ↗ How this number was made Already per GPU Per GPU per hour, GPU Droplet NVIDIA HGX H100 (1 or 8 GPUs); billed per second, 5-minute minimum	$4.41	$441	digitalocean.com ↗
15	15. Nebius$450 $4.50 per GPU-hour On-demand · not stated on page · checked 12 Oct 2026 · nebius.com ↗ How this number was made Per GPU-hour, column 'Effective October 1, 2026'; the page's other column 'On-demand, GPU-hour' still shows $3.85	$4.50	$450	nebius.com ↗
16	16. Replicate$549 $5.49 per GPU-hour Nvidia H100 80GB (gpu-h100; form factor not stated) On-demand · Not stated on the page · checked 12 Oct 2026 · replicate.com ↗ How this number was made 1-GPU hardware class, no division. Per-second billing: $0.001525/sec; Replicate prints $5.49/hr.	$5.49	$549	replicate.com ↗
17	17. CoreWeave$616 $6.16 per GPU-hour On-demand · North America (Europe on-demand identical) · checked 12 Oct 2026 · coreweave.com ↗ How this number was made $49.24/hr for an 8-GPU HGX H100 (80GB) instance ÷ 8 Instance price per hour	$6.16	$616	coreweave.com ↗
18	18. Baseten$650 $6.50 per GPU-hour H100 80 GiB VRAM (form factor not stated) On-demand · Not stated (Baseten: 'Talk to Sales about compute in other countries and regions') · checked 12 Oct 2026 · baseten.co ↗ How this number was made Single-GPU instance; per-minute price x 60. Per-minute billing: $0.10833/min x 60. Instance price (GPU instance with its vCPU/RAM).	$6.50	$650	baseten.co ↗
19	19. AWS (P5, P6, G6)$688 $6.88 per GPU-hour H100 80GB HBM3, 8 per instance on 900 GB/s NVSwitch (AWS does not use the word SXM) On-demand · us-east-1 (US East, N. Virginia) · checked 12 Oct 2026 · aws.amazon.com ↗ How this number was made p5.48xlarge on-demand Linux $55.04/hr ÷ 8 GPUs = $6.88 (cross-check: p5.4xlarge, 1x H100, $6.88/hr). Read from the price file AWS's on-demand page loads (b0.p.awsstatic.com/pricing/2.0/meteredUnitMaps/ec2/USD/current/ec2-ondemand-without-sec-sel/US East (N. Virginia)/Linux/index.json, published 2026-10-08T18:48:50Z); the page's table is rendered from it. Billed per second. The 1-GPU p5.4xlarge is also $6.88/hr, the same per-GPU rate.	$6.88	$688	aws.amazon.com ↗
20	20. Fireworks AI$800 $8.00 per GPU-hour H100 80 GB (form factor not stated) On-demand · Not pinned (region-restricted deployments cost 1.5x, via sales) · checked 12 Oct 2026 · fireworks.ai ↗ How this number was made Printed per GPU, no division. Billed per GPU second; Fireworks prints $0.134/min and $8.00/hr; the hourly figure is used.	$8.00	$800	fireworks.ai ↗
21	21. Oracle Cloud (OCI)$1,000 $10.00 per GPU-hour H100 80GB (BM.GPU.H100.8, bare metal, 8 GPUs; form factor not named by Oracle) On-demand · Not region-specific (Oracle's USD price list carries one rate per part number) · checked 12 Oct 2026 · oracle.com ↗ How this number was made Oracle prices per GPU (metric 'GPU Per Hour'), PAY_AS_YOU_GO $10, no division. BM.GPU.H100.8, part B98415. The page's price cells are empty in its HTML and filled from Oracle's public price-list API (https://apexapps.oracle.com/pls/apex/cetools/api/v1/products/?partNumber=B98415&currencyCode=USD, lastUpdated 2026-10-08), which is where the figure was read. Pay-as-you-go.	$10.00	$1,000	oracle.com ↗
22	22. Microsoft Azure (ND, NC)$1,106–$1,229 $11.06–$12.29 per GPU-hour H100 80GB (ND96is H100 v5 family: 8 GPUs; high = ND96isr H100 v5, the same machine with InfiniBand; Microsoft does not use the word SXM) On-demand · East US 2 (eastus2) · checked 13 Oct 2026 · azure.microsoft.com ↗ How this number was made Low: ND96is_H100_v5 $88.488/hr ÷ 8 — the same $88.488 for ND96is_noIB, ND96isf and ND96is_flex. High: ND96isr_H100_v5 $98.32/hr ÷ 8 (the InfiniBand machine this row showed alone until 13 Oct). Both Linux, East US 2. Pay-as-you-go, listed per hour. Shown as a range like Google's (a3-highgpu to A3 Mega), on Abdul's call of 13 Oct: Azure sells several 8-GPU H100 machines and the InfiniBand one is the dearest. Azure also sells a 1-GPU H100 NVL 94GB PCIe VM, NC40ads H100 v5, at $6.98/hr — a different variant, kept out of this row. Re-read 13 Oct 2026, ~22:30 EDT (14 Oct ~02:30 UTC), on Microsoft's public Retail Prices API (prices.azure.com), East US 2, Linux pay-as-you-go.	$11.06–$12.29	$1,106–$1,229	azure.microsoft.com ↗
23	23. Google Cloud (A3, A4)$1,106–$1,168 $11.06–$11.68 per GPU-hour H100 80GB (a3-highgpu-8g; high = a3-megagpu-8g, A3 Mega) On-demand · Not stated by Google for this table (the A4/A3/A2 table has no region selector; changing the page's region selector moves only the G-series table, checked in the built-in browser) · checked 12 Oct 2026 · cloud.google.com ↗ How this number was made Low: a3-highgpu-8g on-demand $88.490000119/hr ÷ 8. High: a3-megagpu-8g $93.400712807/hr ÷ 8. Per-second billing; listed per hour. Google prices the whole machine (GPUs + vCPU + memory + bundled Local SSD); not eligible for sustained-use discounts.	$11.06–$11.68	$1,106–$1,168	cloud.google.com ↗

H100 SXM specifications on Layer 2

Priced only in another currency — not converted

Scaleway: EUR 3.17–3.31 per GPU-hour, fr-par-2 (Paris, France); no US region, checked 9 Oct 2026. We do not convert a price at an exchange rate the company did not publish. scaleway.com ↗

No public price for the H100 SXM: not published

CUDO Compute — 'Quote on request' on CUDO's pricing page. cudocompute.com ↗
Genesis Cloud — Company page unreachable on 2026-10-09 (TLS error / Cloudflare 522); no price read. genesiscloud.com ↗
Nscale — Listed on the GPU nodes page; price only via 'Contact sales' / 'Get a quote' or inside the console after sign-up. nscale.com ↗
Prime Intellect — Listed as 'Unavailable' in the picker at read time; no price shown. app.primeintellect.ai ↗
SF Compute — SF Compute prices sit behind a sign-in (pricing page redirects to login). sfcompute.com ↗

Other tiers for the H100 SXM — 47 prices, never in the ranking

Kept apart on purpose. A spot price can vanish mid-job; a marketplace price is set by independent hosts or is a live pick across other clouds, and the machine behind it is not one the company owns; a reserved or committed price is a contract, not a rental. Each is a real published price, and none of them is comparable with the on-demand list above.

Spot, community and marketplace — 27 prices

Spot, community and marketplace prices for the NVIDIA H100 SXM 80GB
Company	Tier	Per GPU-hour	Source
Vast.ai H100 SXM 80GB Marketplace: independent hosts, a mix of verified, unverified and deverified. checked 9 Oct 2026 How this number was made min_bid ÷ GPU count; one value per machine Interruptible: the MINIMUM BID per GPU-hour — a floor, you can be outbid and preempted. usdPerGpuHour = lowest, usdPerGpuHourHigh = MEDIAN across 2 machines. Read 20:40 UTC (16:40 EDT). Read from https://console.vast.ai/api/v0/bundles/ (public, unauthenticated marketplace search — POST, no key).	Marketplace · Spot — can be interrupted	$0.65–$1.66	vast.ai ↗
Vast.ai H100 PCIe 80GB Marketplace: independent hosts, a mix of verified, unverified and deverified. checked 9 Oct 2026 How this number was made min_bid ÷ GPU count; one value per machine Interruptible: the MINIMUM BID per GPU-hour — a floor, you can be outbid and preempted. usdPerGpuHour = lowest, usdPerGpuHourHigh = MEDIAN across 3 machines. Read 20:40 UTC (16:40 EDT). Read from https://console.vast.ai/api/v0/bundles/ (public, unauthenticated marketplace search — POST, no key).	Marketplace · Spot — can be interrupted	$0.67–$0.75	vast.ai ↗
Nebius checked 9 Oct 2026 How this number was made Preemptible 'from $0.79' per GPU-hour; dynamic, page says the highest rate is one cent below on-demand	Spot — can be interrupted	$0.79	nebius.com ↗
JarvisLabs H100 SXM 80GB checked 9 Oct 2026 How this number was made none Spot tab: 'may be preempted when capacity is needed'. Read 20:44 UTC (16:44 EDT).	Spot — can be interrupted	$1.54	jarvislabs.ai ↗
Vast.ai H100 SXM 80GB Marketplace: independent hosts, a mix of verified, unverified and deverified. checked 9 Oct 2026 How this number was made none — page prices per GPU vast.ai/pricing headline: usdPerGpuHour = 'from', usdPerGpuHourHigh = 'median'. The page does not say which instance types (on-demand / interruptible) or VRAM sizes it pools — do not read as on-demand. Updated hourly by Vast. Read 20:41 UTC (16:41 EDT).	Marketplace listing	$1.73–$2.40	vast.ai ↗
Vast.ai H100 SXM 80GB Marketplace: independent hosts, a mix of verified, unverified and deverified. checked 9 Oct 2026 How this number was made offer price ÷ GPU count in the offer; one value per machine Marketplace on-demand (non-interruptible). usdPerGpuHour = lowest, usdPerGpuHourHigh = MEDIAN across 3 machines listed. Billed per second. Read 20:39–20:40 UTC (16:39–16:40 EDT). Read from https://console.vast.ai/api/v0/bundles/ (public, unauthenticated marketplace search — POST, no key).	Marketplace on-demand (hosts vary)	$1.80–$2.83	vast.ai ↗
TensorDock H100 SXM5 80GB Marketplace: hosts list at staggered prices, and the floor is not guaranteed. checked 9 Oct 2026 How this number was made table header 'Price per GPU, 8x minimum' gpu-h100.html table 'Spot Bid $1.91/hr' — 'the minimum bid is the lowest price you can bid'. The same page's text says '$1.30/hour if deployed as a spot instance'. Undated; treat as stale-risk. Read 20:44 UTC (16:44 EDT).	Marketplace · Spot — can be interrupted	$1.91	tensordock.com ↗
RunPod H100 PCIe 80GB checked 9 Oct 2026 How this number was made none Community Cloud Pod. Variant row, not SXM. Read 20:41 UTC (16:41 EDT).	Community cloud — other people's hardware	$1.99	runpod.io ↗
Together AI NVIDIA HGX H100 checked 9 Oct 2026 How this number was made Printed per GPU per hour. GPU Clusters, Preemptible Compute, pay as you go.	Preemptible — can be interrupted	$1.99	together.ai ↗
Vast.ai H100 NVL 94GB Marketplace: independent hosts, a mix of verified, unverified and deverified. checked 9 Oct 2026 How this number was made min_bid ÷ GPU count; one value per machine Interruptible: the MINIMUM BID per GPU-hour — a floor, you can be outbid and preempted. usdPerGpuHour = lowest, usdPerGpuHourHigh = MEDIAN across 2 machines. Read 20:40 UTC (16:40 EDT). Read from https://console.vast.ai/api/v0/bundles/ (public, unauthenticated marketplace search — POST, no key).	Marketplace · Spot — can be interrupted	$2.00–$2.13	vast.ai ↗
DataCrunch (now Verda) checked 12 Oct 2026 How this number was made Already per GPU (1x instance) Spot price per hour, 1x H100 SXM5 80GB	Spot — can be interrupted	$2.02	verda.com ↗
Akash Network H100 SXM5 80GB Marketplace: providers bid in a reverse auction. checked 9 Oct 2026 How this number was made API is per GPU-hour Recent provider bids: usdPerGpuHour = MIN, usdPerGpuHourHigh = MEDIAN (avg $2.58, weighted avg $2.73 — the page's 'Starting at'). 21 of 69 GPUs available. Read 20:46 UTC (16:46 EDT).	Marketplace listing	$2.04–$2.56	console-api.akash.network ↗
TensorDock H100 SXM5 80GB Marketplace: hosts list at staggered prices, and the floor is not guaranteed. checked 9 Oct 2026 How this number was made none — per GPU as published Homepage 'FROM $2.25/HR'; gpu-h100.html: 'hosts are listed at staggered pricing from $1.90 to $2.50/hr. Availability at $1.90/hr is not guaranteed' — usdPerGpuHourHigh = that $2.50 top. CPU/RAM/storage billed on top ('a la carte'). Undated page. Read 20:44 UTC (16:44 EDT).	Marketplace on-demand (hosts vary)	$2.25–$2.50	tensordock.com ↗
Microsoft Azure (ND, NC) H100 80GB (ND96isr H100 v5, 8x) checked 9 Oct 2026 How this number was made ND96isr_H100_v5 spot $18.1695/hr ÷ 8 GPUs (read on the page). Spot: evictable, variable.	Spot — can be interrupted	$2.27	azure.microsoft.com ↗
Vast.ai H100 NVL 94GB Marketplace: independent hosts, a mix of verified, unverified and deverified. checked 9 Oct 2026 How this number was made offer price ÷ GPU count; one value per machine Variant row. Marketplace on-demand. usdPerGpuHour = lowest, usdPerGpuHourHigh = MEDIAN across 3 machines. Read 20:39–20:40 UTC (16:39–16:40 EDT). Read from https://console.vast.ai/api/v0/bundles/ (public, unauthenticated marketplace search — POST, no key).	Marketplace on-demand (hosts vary)	$2.39–$2.47	vast.ai ↗
Vast.ai H100 NVL (page label 80GB) Marketplace: independent hosts, a mix of verified, unverified and deverified. checked 9 Oct 2026 How this number was made none — page prices per GPU vast.ai/pricing headline: usdPerGpuHour = 'from', usdPerGpuHourHigh = 'median'. The page does not say which instance types (on-demand / interruptible) or VRAM sizes it pools — do not read as on-demand. Updated hourly by Vast. Read 20:41 UTC (16:41 EDT).	Marketplace listing	$2.39–$2.93	vast.ai ↗
CoreWeave checked 9 Oct 2026 How this number was made $19.71/hr for an 8-GPU HGX H100 instance ÷ 8 Spot instance price per hour; Europe spot $19.51/instance	Spot — can be interrupted	$2.46	coreweave.com ↗
Vast.ai H100 PCIe 80GB Marketplace: independent hosts, a mix of verified, unverified and deverified. checked 9 Oct 2026 How this number was made offer price ÷ GPU count; one value per machine Variant row. Marketplace on-demand. usdPerGpuHour = lowest, usdPerGpuHourHigh = MEDIAN across 3 machines. Read 20:39–20:40 UTC (16:39–16:40 EDT). Read from https://console.vast.ai/api/v0/bundles/ (public, unauthenticated marketplace search — POST, no key).	Marketplace on-demand (hosts vary)	$2.59–$2.77	vast.ai ↗
RunPod H100 NVL 94GB checked 9 Oct 2026 How this number was made none Community Cloud Pod. Variant row, not SXM. Read 20:41 UTC (16:41 EDT).	Community cloud — other people's hardware	$2.59	runpod.io ↗
Vast.ai H100 PCIe 80GB Marketplace: independent hosts, a mix of verified, unverified and deverified. checked 9 Oct 2026 How this number was made none — page prices per GPU vast.ai/pricing headline: usdPerGpuHour = 'from', usdPerGpuHourHigh = 'median'. The page does not say which instance types (on-demand / interruptible) or VRAM sizes it pools — do not read as on-demand. Updated hourly by Vast. Read 20:41 UTC (16:41 EDT).	Marketplace listing	$2.59–$2.77	vast.ai ↗
io.net H100 PCIe 80GB Marketplace: partner-supplied capacity, some of it resold, quoted from a live catalogue. io.net's own listings were sold out on 9 Oct. checked 9 Oct 2026 How this number was made US 4-GPU VM $10.416 ÷ 4; FR/PL 1-GPU VM $3.67 Variant row (catalogue interconnect 'pcie'). usdPerGpuHour = lowest per-GPU, usdPerGpuHourHigh = highest per-GPU of 4 available listings. Read 20:45 UTC (16:45 EDT).	Marketplace on-demand (hosts vary)	$2.60–$3.67	cloud.io.net ↗
RunPod H100 SXM 80GB checked 9 Oct 2026 How this number was made none — page prices per GPU Community Cloud Pod (other people's hardware), list price per GPU-hour shown with the Community Cloud toggle; billed per second. Read 20:41 UTC (16:41 EDT).	Community cloud — other people's hardware	$2.69	runpod.io ↗
Spheron H100 80GB (menu and detail page: SXM5) Marketplace: the cheapest live rate among its partner data centres. checked 9 Oct 2026 How this number was made none Spot rate shown under the on-demand rate; 'can be reclaimed at any time'. Read 20:43 UTC (16:43 EDT).	Marketplace · Spot — can be interrupted	$2.77	spheron.network ↗
Spheron H100 80GB (menu and detail page: SXM5) Marketplace: the cheapest live rate among its partner data centres. checked 9 Oct 2026 How this number was made none — page prices per GPU Cheapest live on-demand rate on Spheron's marketplace (only the cheapest is shown). Per-minute billing, 20-minute minimum. Read 20:43 UTC (16:43 EDT).	Marketplace on-demand (hosts vary)	$2.98	spheron.network ↗
AWS (P5, P6, G6) H100 80GB (p5.48xlarge, 8x) checked 9 Oct 2026 How this number was made p5.48xlarge spot $28.8881/hr ÷ 8. Spot: variable, interruptible; the Linux value in the feed AWS's spot pricing page loads (website.spot.ec2.aws.a2z.com/spot.json) at read time. Spot moves with demand and differs by Availability Zone. p5.4xlarge (1x) spot $3.9147/hr.	Spot — can be interrupted	$3.61	aws.amazon.com ↗
io.net H100 SXM5 80GB Marketplace: partner-supplied capacity, some of it resold, quoted from a live catalogue. io.net's own listings were sold out on 9 Oct. checked 9 Oct 2026 How this number was made $8.439 machine-hour ÷ 2 GPUs Only one SXM5 listing (2-GPU VM, $8.44/hr). Partner-supplied. Read 20:45 UTC (16:45 EDT).	Marketplace on-demand (hosts vary)	$4.22	cloud.io.net ↗
Google Cloud (A3, A4) H100 80GB (a3-highgpu-8g, 8x) checked 9 Oct 2026 How this number was made a3-highgpu-8g spot $50.315338521/hr ÷ 8 GPUs. Spot ('Current Spot pricing' column): preemptible, variable. Odd: Google's spot figure here is ABOVE its DWS Flex-start figure. Google prices the whole machine (GPUs + vCPU + memory + bundled Local SSD); not eligible for sustained-use discounts.	Spot — can be interrupted	$6.29	cloud.google.com ↗

Reserved, committed and other plans — 20 prices

Reserved, committed and other plan prices for the NVIDIA H100 SXM 80GB
Company	Tier	Per GPU-hour	Source
Paperspace (DigitalOcean) checked 9 Oct 2026 How this number was made Per GPU H100 card price $2.24*/hour; footnote: '$2.24/hour pricing is for a 3-year commitment'	Committed 3-year (Paperspace CORE)	$2.24	paperspace.com ↗
Vultr checked 9 Oct 2026 How this number was made Already per GPU Bare Metal 'Starting from' per GPU per hour for pre-paid contract terms (page: 'contact sales to discuss pricing'). 8x H100 80 GB SXM.	Reserved — a contract term	$2.30	vultr.com ↗
Civo checked 9 Oct 2026 How this number was made 1-GPU instance Commitment, per hour: 6 months $2.79, 12 months $2.69, 24 months $2.59, 36 months $2.49 (1 GPU)	Committed — a contract term	$2.49–$2.79	civo.com ↗
Hyperstack (NexGen Cloud) checked 9 Oct 2026 How this number was made Already per GPU Reservation 'Starting from' per GPU per hour; term not stated. H100 SXM.	Reserved — a contract term	$2.72	hyperstack.cloud ↗
RunPod H100 PCIe 80GB checked 9 Oct 2026 How this number was made none Secure Cloud Pod. Variant row, not SXM. Read 20:41 UTC (16:41 EDT).	Secure Cloud pod	$2.89	runpod.io ↗
RunPod H100 NVL 94GB checked 9 Oct 2026 How this number was made none Secure Cloud Pod. Variant row, not SXM. Read 20:41 UTC (16:41 EDT).	Secure Cloud pod	$3.19	runpod.io ↗
Together AI NVIDIA HGX H100 checked 9 Oct 2026 How this number was made Printed per GPU per hour. GPU Clusters, Reserved, 91-180 days (181+ days: contact us).	Reserved, 91–180 days	$3.19	together.ai ↗
Paperspace (DigitalOcean) checked 9 Oct 2026 How this number was made Already per GPU 12 Month Reserved Price, per GPU per hour (contractual 12-month commitment, via sales)	Reserved — a contract term	$3.26	digitalocean.com ↗
Together AI NVIDIA HGX H100 checked 9 Oct 2026 How this number was made Printed per GPU per hour. GPU Clusters, Reserved, 31-90 days.	Reserved, 31–90 days	$3.45	together.ai ↗
Together AI NVIDIA HGX H100 checked 9 Oct 2026 How this number was made Printed per GPU per hour. GPU Clusters, Reserved, 3-30 days.	Reserved, 3–30 days	$3.69	together.ai ↗
Google Cloud (A3, A4) H100 80GB (a3-highgpu-8g, 8x) checked 9 Oct 2026 How this number was made a3-highgpu-8g dws-flex-start $38.32/hr ÷ 8 GPUs. Dynamic Workload Scheduler Flex-start: queued, capacity not guaranteed at request time; Google says DWS prices are variable but updated infrequently. Google prices the whole machine (GPUs + vCPU + memory + bundled Local SSD); not eligible for sustained-use discounts.	Flex-start — queued, capacity not guaranteed	$4.79	cloud.google.com ↗
Google Cloud (A3, A4) H100 80GB (a3-highgpu-8g, 8x) checked 9 Oct 2026 How this number was made a3-highgpu-8g committed-use-3yr $38.864383195/hr ÷ 8 GPUs. Compute Resource CUD, 3-year commitment. Google prices the whole machine (GPUs + vCPU + memory + bundled Local SSD); not eligible for sustained-use discounts.	Committed use, 3 years	$4.86	cloud.google.com ↗
Google Cloud (A3, A4) H100 80GB (a3-highgpu-8g, 8x) checked 9 Oct 2026 How this number was made a3-highgpu-8g dws-calendar $41.6/hr ÷ 8 GPUs. Dynamic Workload Scheduler Calendar mode: reserved future window; variable price, updated infrequently. Google prices the whole machine (GPUs + vCPU + memory + bundled Local SSD); not eligible for sustained-use discounts.	Calendar mode — a reserved future window	$5.20	cloud.google.com ↗
Microsoft Azure (ND, NC) H100 80GB (ND96isr H100 v5, 8x) checked 9 Oct 2026 How this number was made ND96isr_H100_v5 reserved-3yr $43.1625/hr ÷ 8 GPUs (read on the page). Reserved instance, 3-year term, effective hourly rate. Azure shows no savings-plan rate for this VM.	Reserved, 3 years	$5.40	azure.microsoft.com ↗
Together AI NVIDIA HGX H100 checked 9 Oct 2026 How this number was made Printed per GPU per hour. Dedicated Inference (single-tenant managed endpoint), pay as you go. Reserved: contact sales. Tier named without 'on-demand' so the merge keeps it out of the on-demand spread (the GPU Clusters row is Together's on-demand price).	Dedicated inference endpoint	$5.49	together.ai ↗
Lambda checked 9 Oct 2026 How this number was made Already per GPU 1-Click Cluster ('NVIDIA H100', HGX systems), 2 weeks–1 year term, per GPU per hour: 16 GPUs $6.16, 64 GPUs $5.85, 256 GPUs $5.54	Reserved — a contract term	$5.54–$6.16	lambda.ai ↗
Paperspace (DigitalOcean) checked 9 Oct 2026 How this number was made Per GPU (page: multiply by number of GPUs for multi-GPU) Footnote on the H100 card: 'On-demand pricing for H100 is $5.95/hour under our special promo price.' NVIDIA HGX H100, 8x multi-GPU type	Promo price (Paperspace CORE legacy page)	$5.95	paperspace.com ↗
AWS (P5, P6, G6) H100 80GB (p5.48xlarge, 8x) checked 9 Oct 2026 How this number was made AWS prints $47.757/hr per instance ($5.970 per accelerator). Capacity Blocks for ML: reservation fee for a fixed future window, paid up front; Linux OS fee $0.000/hr. AWS says prices are updated regularly, next update scheduled January 2027. p5.4xlarge (1x H100) also $5.970/hr.	Capacity Block — a reserved window, paid up front	$5.97	aws.amazon.com ↗
Google Cloud (A3, A4) H100 80GB (a3-highgpu-8g, 8x) checked 9 Oct 2026 How this number was made a3-highgpu-8g committed-use-1yr $61.383674231/hr ÷ 8 GPUs. Compute Resource CUD, 1-year commitment. Google prices the whole machine (GPUs + vCPU + memory + bundled Local SSD); not eligible for sustained-use discounts.	Committed use, 1 year	$7.67	cloud.google.com ↗
Microsoft Azure (ND, NC) H100 80GB (ND96isr H100 v5, 8x) checked 9 Oct 2026 How this number was made ND96isr_H100_v5 reserved-1yr $62.9248/hr ÷ 8 GPUs (read on the page). Reserved instance, 1-year term, shown by Azure as an effective hourly rate. Azure shows no savings-plan rate for this VM.	Reserved, 1 year	$7.87	azure.microsoft.com ↗

Every chip at a glance

The cheapest and dearest published on-demand price for each chip, read 13 Oct 2026. The gap between them is the number that lasts: any single price here will move within weeks.

Chip	On-demand prices	Cheapest per GPU-hour	Dearest per GPU-hour	Same chip
NVIDIA H100 SXM 80GB	23	$1.99 Voltage Park	$11.06 Google Cloud (A3, A4)	5.6× apart
NVIDIA H200 141GB	19	$3.20 GMI Cloud	$10.60 Google Cloud (A3, A4)	3.3× apart
NVIDIA B200	14	$3.50 Vultr	$14.24 AWS (P5, P6, G6)	4.1× apart
NVIDIA A100 80GB	17	$1.35 Hyperstack (NexGen Cloud)	$5.07 Google Cloud (A3, A4)	3.8× apart
NVIDIA L40S 48GB	13	$1.09 RunPod	$3.51 Replicate	3.2× apart
NVIDIA GeForce RTX 4090 24GB	1	—	—	one price
NVIDIA L4 24GB	7	$0.44 JarvisLabs	$1.00 OVHcloud	2.3× apart
AMD Instinct MI300X 192GB	5	$1.85 Vultr	$6.00 Oracle Cloud (OCI)	3.2× apart

Every price for one chip on its own page: H100 SXM rental prices · H200 rental prices · B200 rental prices · A100 80GB rental prices · L40S rental prices · L4 rental prices

What changed since the last check

16 Oct 2026
- The headline spread now compares each company at its own cheapest published on-demand price for the chip. For the H100 that is $1.99 (Voltage Park) to $11.06 (Google Cloud) — 5.6× for the same chip. Until today the dearest end was the top of a company's range (Azure's InfiniBand machine, $12.29, which read 6.2×); that machine is still listed as the top of Azure's range. No price changed — only the rule that picks the dearest end, so it is the same for every company. The other chips' spreads are unchanged.
13 Oct 2026
- RunPod’s Secure Cloud prices are now ranked. They are RunPod’s own on-demand prices — GPUs in its own data centres, billed by the second — and until today they sat under “other tiers”. One rule for every company: a published hourly price is ranked. Re-read on RunPod’s page today, unchanged since 9 Oct: B200 $7.99, H100 SXM $3.99, L40S $1.09, L4 $0.59 and RTX 4090 $0.89 per GPU-hour; RunPod’s H200 now reads $4.31 (Instant Clusters) to $5.29 (a Secure Cloud machine). Its Community Cloud prices — other people’s hardware — stay unranked.
- The L40S has a new cheapest price: $1.09 at RunPod (it was $1.29 at Civo), so the same L40S is now 3.2× apart (was 2.7×). No other cheapest or dearest price moved.
- Microsoft Azure’s H100 is now a range, like Google’s: $11.06 per GPU-hour for its standard 8-GPU H100 machines (ND96is H100 v5, $88.488 an hour) up to $12.29 for the same machine with InfiniBand networking (ND96isr H100 v5, $98.32 an hour). Until today the row showed only the $12.29 machine. Re-read on Microsoft’s public price list today, East US 2.
12 Oct 2026
- Re-check, 12 Oct 2026: every ranked on-demand price (94 prices from 23 companies) re-read on the company’s own price page. 89 were unchanged.
- DataCrunch (now Verda) raised its on-demand prices by about 3%: B200 $7.42 → $7.64, H200 $5.12 → $5.28, H100 $3.93 → $4.05, A100 80GB $1.89 → $1.95, L40S $1.62 → $1.66 per GPU-hour. Its spot prices rose the same way.
- No cheapest or dearest on-demand price moved, so no spread changed: the same H100 is still $1.99 at Voltage Park and $12.29 at Microsoft Azure.
- Voltage Park’s price page now says it has merged with Lightning AI; its $1.99 H100 price is unchanged.
- The five companies without a readable price on 9 Oct are still without one: Genesis Cloud’s site still does not load, CUDO Compute, SF Compute and Nscale still quote privately or behind a sign-in, and Fly.io still rents no GPUs.
9 Oct 2026
- First full read, 9 Oct 2026: every price read on the company's own price page. It replaces the 9 Sep survey of three articles, which stays below as background.
- Fly.io no longer rents GPUs. Its pricing page lists none, and a Fly email quoted on its own community forum dates the retirement to 31 Jul 2026.

Download the table — free to reuse with credit

Every row behind this page — 283 published prices, plus the companies that publish none — with the source URL, the date read and how each number was made. Published under CC BY 4.0: use it commercially, change it, republish it. The one condition is credit with a link.

Download CSV Download JSON

Cite as

Macrostack (2026). GPU rental prices, read 13 Oct 2026. CC BY 4.0. https://www.macrostack.net/compute/gpu-prices

Or as a link

<a href="https://www.macrostack.net/compute/gpu-prices">GPU prices from Macrostack</a>, CC BY 4.0

Think a price here is wrong? It may well be — tell us with the company’s page and we correct it the same day. We never remove a correct price because it is unflattering.

How the numbers were made

Read, not measured. Each price was read on the company’s own public price page on the date in its row, and the row links to that page. We do not rent clusters and we do not benchmark them.
On-demand, per GPU, per hour, in US dollars. A machine sold as 8 GPUs is divided by 8; per-second and per-minute billing is multiplied up to an hour. Each row says how its number was made, and names the region where the company does.
Ordered by price and nothing else. Lowest price first, then the top of its range; an exact tie is listed alphabetically. No company pays to be on this page or to move on it, and there are no affiliate or sponsored links here.
Kept apart, never ranked: spot, preemptible, community and marketplace prices; reserved, committed and promo prices; listings that were sold out on the day; and prices published only in another currency, which we do not convert.
Marketplaces are never ranked. Where a company’s price is set by independent hosts, or is a live pick of the cheapest offer across other clouds — TensorDock, io.net, Spheron, Prime Intellect, Vast.ai and Akash on 13 Oct 2026 — every row of it is listed with the marketplace prices, whatever its tier says. Where a ranked figure carries a caveat of its own, the row says so.
Kept fresh by the build. Prices are re-read every 14 days. A row older than 30 days says how old it is, and the site refuses to build with an on-demand price older than 45 days.
List prices are a ceiling. Committed, reserved and negotiated rates are lower than any on-demand price here; nobody publishes what large buyers actually pay.

Companies without a readable public price

36 companies checked on 13 Oct 2026; 31 publish prices we could read. The rest, and why — listed, never guessed.

Company	Status	What we found
Fly.io	no GPU rental	No GPU rental. Fly's pricing page lists no GPUs on 9 Oct 2026, fly.io/docs/gpus/ now redirects to its February 2025 post 'We Were Wrong About GPUs', and a Fly email quoted on Fly's own community forum (11 Feb 2026) says GPUs are fully deprecated as of 31 Jul 2026. their page ↗
Genesis Cloud	page unreadable	Its site could not be reached on 9 Oct 2026: www.genesiscloud.com failed the TLS handshake and the bare domain returned a Cloudflare 522 (origin timed out), at 20:43 and 20:48 UTC. Third-party figures exist but are not used. It is re-read on the next pass. their page ↗
CUDO Compute	login/quote only	Pricing page lists B300, GB300, VR200, H100 PCIe, H100 SXM, H200, HGX B200 and GB200 NVL72 — every one 'Quote on request'. Product pages (e.g. /products/nvidia-h100, /products/nvidia-b200) say 'Contact us for pricing'. No self-serve on-demand price is published. A100, L40S, RTX 4090 and L4 do not appear. their page ↗
SF Compute	login/quote only	sfcompute.com/pricing answers 307 to a WorkOS sign-in. The homepage, /specs and docs.sfcompute.com publish no current per-GPU price; market prices appear only in authenticated CLI / dashboard output. The homepage resale illustration ('Reserved price $3.00/gpu/hr', 'Realized price $4.29/gpu/hr') names no GPU and no date — an example, not used. their page ↗
Nscale	login/quote only	No public GPU price anywhere on nscale.com: /pricing returns 404, the GPU nodes page and the H100/H200 node pages offer only 'Contact sales' / 'Get a quote', and the docs say to add $5 credit in the console (sign-up required). GPU nodes page lists H100, H200, GB200, GB300 and Vera Rubin only. MI300X appears in Nscale blog benchmarks but not as a priced rental. their page ↗

Questions

How much does it cost to rent an H100 per hour?: Between $1.99 and $11.06 per GPU-hour on published on-demand prices read on 13 Oct 2026: Voltage Park was the cheapest and Google Cloud (A3, A4) the dearest, 5.6× apart for the same chip. Spot and marketplace listings go lower but can be interrupted or depend on someone else's machine, and committed or negotiated rates are lower than every on-demand price — nobody publishes those.
Why is the same GPU so much cheaper at one company than another?: Because you are not only buying the chip. The cheapest end is often brokered or resold capacity, priced per GPU with everything else extra. The dearest end is a hyperscaler list price for a whole 8-GPU machine with its CPUs, memory and storage, which exists mainly to make the committed rate look like a discount. Between them sit differences in support, networking, region, billing and contract terms — and in who the counterparty is.
Does Macrostack measure these prices itself?: No. We do not rent clusters and we do not benchmark them. Every figure is a price the company publishes, read on its own price page on the date in its row, with a link to that page. Each row says how its number was made.
How often are the prices checked?: Every 14 days. A row older than 30 days says how old it is, and the site will not build with an on-demand price older than 45 days.
Can I reuse this data?: Yes, under CC BY 4.0, commercial use included. The one condition is credit with a link: Macrostack (2026). GPU rental prices, read 13 Oct 2026. CC BY 4.0. https://www.macrostack.net/compute/gpu-prices

Background: what three surveys said on 9 Sep 2026

The first edition of this page showed price ranges read from three third-party articles, not from each company. They are kept here, with their sources, so anyone who cited them can see what replaced them. They are not mixed into the table above.

NVIDIA H100 SXM 80GB: $1.38 to $12.29

Marketplace and boutique floor: $1.38 - 1.49 (on-demand, marketplace) intuitionlabs.ai
Specialist neocloud and marketplace on-demand: $2.00 - 3.00 (on-demand) intuitionlabs.ai
Broad market median: $3.38 (on-demand) shattered.io
Hyperscaler on-demand ceiling: $11.68 - 12.29 (on-demand, hyperscaler list) intuitionlabs.ai

NVIDIA H200 SXM 141GB: $3.59 to $4.59

RunPod community cloud: $3.59 (community) gpusmith.com
RunPod secure cloud: $4.39 - 4.59 (secure) gpusmith.com

NVIDIA B200: $3.44 to $16.11

Vast.ai: $3.44 (marketplace) intuitionlabs.ai
Lambda: $4.99 - 5.29 (on-demand) intuitionlabs.ai
RunPod: $5.89 (on-demand) intuitionlabs.ai
Median across surveyed providers: $6.11 (on-demand) intuitionlabs.ai
Google Cloud: $16.11 (on-demand, hyperscaler list) intuitionlabs.ai

Neoclouds and marketplaces price 40% to 400% below the three major hyperscalers for the same accelerator. spheron.network
H100 on-demand rates roughly halved during 2026, while Blackwell-class parts held a wide spread — the signature of a market where the previous generation has become a commodity and the current one is still allocated. shattered.io

Who these companies are, who owns them and how hard each is to leave: Layer 2.5 — Compute. The chips themselves: Layer 2 — Silicon. The whole map: the five layers.

Prices read 13 Oct 2026. Each price read on the company's own public price page on the date shown. On-demand per GPU per hour in USD; multi-GPU machines divided by GPU count (see normalization); per-second billing converted to per hour. Not measured by Macrostack. Committed, reserved and negotiated rates are lower and are not published by anyone.