Net cost if you stopped at each month
What you have paid, plus what you still owe (contracts, loans, leases), minus what the hardware would sell for. 64 × H100 SXM at 70% utilization, in today's dollars at 14.9% cost of capital.
Rent Own
Show the numbers
| Month | Rent | Own |
|---|---|---|
| 0 | $0 | $0 |
| 6 | $809K | $924K |
| 12 | $1.56M | $1.61M |
| 18 | $2.27M | $2.22M |
| 24 | $2.92M | $2.76M |
| 30 | $3.54M | $3.25M |
| 36 | $4.11M | $3.68M |
Owning saves $428K (10%) over 3 yr and pulls ahead for good in month 16. Buying wins above 62% utilization or when renting costs more than $3.45/hr per GPU.
Rent
Crusoe · H100 80GB HGX
$4.11M
Price$3.90/hr per GPU
BillingPer hour used
Per useful GPU-hour$4.29/hr
Cash out (undiscounted)$5.05M
Own
8 × HGX H100 server · 90 kW
$3.68M
Up-front hardware$2.71M
Running cost$60K/mo
Per useful GPU-hour$3.84/hr
Cash out (undiscounted)$3.66M
Cost per useful GPU-hour vs utilization
Pay-as-you-go is flat: idle hours cost nothing. Owning and committing get cheaper the busier you keep the GPUs. The lines cross at 62%.
Your rent option Own Other rent options
Show the numbers
| Utilization | Cheapest 1-yr+ commit | Median spot / market | Your rent option | Own |
|---|---|---|---|---|
| 10% | $22.4/hr | $2.50/hr | $4.29/hr | $23.7/hr |
| 15% | $14.9/hr | $2.50/hr | $4.29/hr | $16.0/hr |
| 20% | $11.2/hr | $2.50/hr | $4.29/hr | $12.1/hr |
| 25% | $8.96/hr | $2.50/hr | $4.29/hr | $9.80/hr |
| 30% | $7.47/hr | $2.50/hr | $4.29/hr | $8.26/hr |
| 35% | $6.40/hr | $2.50/hr | $4.29/hr | $7.15/hr |
| 40% | $5.60/hr | $2.50/hr | $4.29/hr | $6.33/hr |
| 45% | $4.98/hr | $2.50/hr | $4.29/hr | $5.68/hr |
| 50% | $4.48/hr | $2.50/hr | $4.29/hr | $5.17/hr |
| 55% | $4.07/hr | $2.50/hr | $4.29/hr | $4.75/hr |
| 60% | $3.73/hr | $2.50/hr | $4.29/hr | $4.39/hr |
| 65% | $3.45/hr | $2.50/hr | $4.29/hr | $4.10/hr |
| 70% | $3.20/hr | $2.50/hr | $4.29/hr | $3.84/hr |
| 75% | $2.99/hr | $2.50/hr | $4.29/hr | $3.62/hr |
| 80% | $2.80/hr | $2.50/hr | $4.29/hr | $3.43/hr |
| 85% | $2.64/hr | $2.50/hr | $4.29/hr | $3.26/hr |
| 90% | $2.49/hr | $2.50/hr | $4.29/hr | $3.11/hr |
| 95% | $2.36/hr | $2.39/hr | $4.11/hr | $2.95/hr |
| 100% | $2.24/hr | $2.27/hr | $3.90/hr | $2.91/hr |
Your numbers
Rent
$3.90/hr
Big buyers pay below list. Enter your own price, or compare against the committed options on the price board (CoreWeave advertises up to 60% off and Nebius up to 35% off for commitments).
0%
Today's prices for the whole horizon. It moves every rental price here: renting, renting while servers arrive or are down, and resold idle hours. A committed price holds for its term and renews at the price of the day.
+10%
Hours billed beyond useful work on on-demand, spot and serverless: spin-up, idle tails, spot interruptions. Ignored for committed options.
14.9%
Treasuries plus 10 points, about what venture lenders earned on their loans in Q2 2026 (12–15%). Use this if you would otherwise borrow. 3-year Treasuries yielded 4.85% on Oct 15, 2026, 4.91% a year compounded. Every payment is discounted at this rate: at 14.9%, $1 paid in 3 years is worth $0.66 today. That is how renting or borrowing can beat paying cash.
Own
$275K
$34K per GPU. Dell XE9680 8× H100 listed at $275K new (Viperatech) and $249K refurbished; range $225K–$350K.
$3.4K
NICs, switches, optics and cables for the GPU-to-GPU fabric (400G per GPU, non-blocking).
$205/kW/mo
Per kW of IT load reserved. You need 90 kW.
$0.089/kWh
0.75 FTE
At $250K/yr loaded. Someone has to swap the dead GPUs.
$73K
Designing the cluster, choosing a colo, contracts, network fabric, scheduler and acceptance tests, paid while you wait for delivery. Our estimate: 3.5 engineer-months at $250K/yr (3 plus 6 per 1,000 GPUs).
$75/GPU/mo
A managed service that watches the hardware, triages failures and handles part swaps, on top of the support contract and remote hands. Our estimate; set it to $0 if your own staff covers it. $58K a year for 64 GPUs.
3 mo
You keep renting at $3.90/hr until it arrives.
18%
This GPU's default: 18% a year for HGX servers (20% for NVL72 racks and PCIe-card servers, 25% for AMD and used A100s), so an HGX server keeps about 55% after 3 years. Mercatus puts 3-year-old H100s at 40–75% of list in 2026. Worth $1.53M at the end of the horizon.
97.0%
Hours lost to failures and maintenance. If you need more than this, the shortfall is rented on-demand.
Price board · ways to rent H100 SXM
Last refresh Oct 16, 19:17 UTC. Prices from provider APIs update hourly (AWS spot, Azure, Akash) or daily (AWS, Oracle Cloud, Verda, Modal, OpenRelay, GetDeploying market medians). The rest were read by hand, most recently Oct 2, 2026.
| Pricing | vs owning | Source | ||||
|---|---|---|---|---|---|---|
Mithril H100 SXM5 1x | Spot no commitment | $1.00 | $1.10 | 71% cheaper | mithril.ai checked Oct 2, 2026 | |
Mithril H100 SXM5 8x · 8 GPUs | Spot no commitment | $1.50 | $1.65 | 57% cheaper | mithril.ai checked Oct 2, 2026 Spot, 7-day average; spot multi-node runs on same InfiniBand fabric as reservations. | |
Together AI GPU Cluster HGX H100 (preemptible) · 8 GPUs | Spot no commitment | $1.99 | $2.19 | 43% cheaper | together.ai checked Oct 2, 2026 "Preemptible Compute" column on Together pricing page. | |
Voltage Park HGX H100 on-demand (starting at) | On-demand no commitment | $1.99 | $2.19 | 43% cheaper | voltagepark.com checked Oct 2, 2026 "On-demand access starts at $1.99/hr"; pricing table at https://www.voltagepark.com/pricing says contact for pricing. Ethernet tier 1-1016 GPUs, InfiniBand (3200 Gbps) tier 8-1016 GPUs, long-term reserve 6+ months… | |
GMI Cloud H100 SXM "from" | On-demand no commitment | $2.00 | $2.20 | 43% cheaper | gmicloud.ai checked Oct 2, 2026 Lowest "from" price on the pricing page; the page does not say whether it is on-demand or reserved. | |
Market median H100, all variants | Spot no commitment | $2.04 | $2.25 | 41% cheaper | getdeploying.com weekly median Median listed price across 10 providers (42 listings), week of 2026-09-28; range $0.54–$6.64. Mixes H100 SXM, PCIe and NVL listings. Data: GetDeploying (CC BY 4.0), from the copy saved with the site. | |
Verda 8H100.80S.176V · 8 GPUs | Spot no commitment | $2.08 | $2.29 | 40% cheaper | verda.com live | |
SF Compute 8x H100 SXM5 node (VM), market average · 8 GPUs | Marketplace no commitment | $2.10 | $2.31 | 40% cheaper | old.sfcompute.com checked Oct 2, 2026 Daily average H100 price embedded in old.sfcompute.com page payload (pricesByHardwareType.H100, date 2026-10-02: avg 2.10, top 2.88, bottom 1.30); /prices page payload read minutes later showed avg 2.11. Unclear… | |
Nebius Preemptible H100 SXM | Spot no commitment | $2.15 | $2.36 | 38% cheaper | docs.nebius.com checked Oct 2, 2026 | |
Azure ND96isr_H100_v5 · 8 GPUs · eastus | Spot no commitment | $2.27 | $2.50 | 35% cheaper | azure.microsoft.com live Can be evicted at any time | |
Vultr Bare metal vbm-112c-2048gb-8-h100-gpu (preemptible) · 8 GPUs | Spot no commitment | $2.30 | $2.53 | 34% cheaper | vultr.com checked Oct 2, 2026 | |
Denvr 8x H100 SXM, IB 3200G · 8 GPUs | On-demand no commitment | $2.45 | $2.70 | 30% cheaper | denvr.com checked Oct 2, 2026 Per GPU-hour, per-minute billing. Reserved: "save up to 20%" (multi-month/annual, cluster size, prepayment); H200, GB200, B200, Gaudi 3, MI300X are reserved-only (no published rates). | |
CoreWeave HGX H100 · 8 GPUs · North America | Spot no commitment | $2.46 | $2.71 | 29% cheaper | coreweave.com checked Oct 2, 2026 | |
Akash Network H100 80Gi | Marketplace no commitment | $2.56 | $2.82 | 27% cheaper | akash.network live Median of recent provider bids on a decentralized network (range $2.04–$3.16); 69 GPUs listed, 3 free. | |
OpenRelay H100 SXM | On-demand no commitment | $2.60 | $2.86 | 26% cheaper | openrelay.inc live Verified hosts. Billed per minute; storage and egress included. | |
io.net H100 x1 US (vmaas) | On-demand no commitment | $2.60 | $2.86 | 25% cheaper | api.io.solutions checked Oct 2, 2026 io.cloud VMaaS listing, external supplier; form factor unclear | |
RunPod Community Cloud pod - H100 SXM | Community hosts no commitment | $2.69 | $2.96 | 23% cheaper | runpod.io checked Oct 2, 2026 | |
Paperspace HGX H100 3-year commitment · 8 GPUs | Reserved 3 yr term | $2.24 | $3.20 | 17% cheaper | paperspace.com checked Oct 2, 2026 "$2.24/hour pricing is for a 3-year commitment." | |
Thunder Compute Sandbox H100 | Serverless no commitment | $2.95 | $3.25 | 16% cheaper | thundercompute.com checked Oct 2, 2026 Sandboxes billed per second, scale to zero; vCPU/RAM/storage billed separately; access by request. | |
Vultr Bare metal vbm-112c-2048gb-8-h100-gpu · 8 GPUs | On-demand no commitment | $2.99 | $3.29 | 14% cheaper | vultr.com checked Oct 2, 2026 | |
Massed Compute 8x H100 SXM5 · 8 GPUs | On-demand no commitment | $3.14 | $3.45 | 10% cheaper | massedcompute.com checked Oct 2, 2026 | |
Hyperbolic H100 SXM on-demand (starting from) | On-demand no commitment | $3.19 | $3.51 | 9% cheaper | hyperbolic.ai checked Oct 2, 2026 "Starting from" price; refreshed weekly from best available supplier rates. Reserved is "discounted prepaid", no published rates. | |
Hyperstack H100 SXM | On-demand no commitment | $3.20 | $3.52 | 8% cheaper | hyperstack.cloud checked Oct 2, 2026 | |
Thunder Compute h100_x1 | On-demand no commitment | $3.20 | $3.52 | 8% cheaper | thundercompute.com checked Oct 2, 2026 Same $3.20/GPU-hr for x1/x2/x4/x8 (x8 = $25.60/hr). | |
Market median H100, all variants | On-demand no commitment | $3.36 | $3.69 | 4% cheaper | getdeploying.com weekly median Median listed price across 41 providers (134 listings), week of 2026-09-28; range $1.30–$12.29. Mixes H100 SXM, PCIe and NVL listings. Data: GetDeploying (CC BY 4.0), from the copy saved with the site. | |
AWS p5.48xlarge · 8 GPUs · us-east-1 | Savings plan 3 yr term, all upfront | $2.59 | $3.70 | 4% cheaper | aws.amazon.com live EC2 Instance Savings Plan (locked to this instance family), all upfront, shown as the effective hourly rate. | |
AWS p5.48xlarge · 8 GPUs · us-east-1 | Reserved 3 yr term, all upfront | $2.59 | $3.70 | 4% cheaper | aws.amazon.com live Standard Reserved Instance, all upfront, shown as the effective hourly rate. | |
Cerebrium H100 ($0.000944/sec) | Serverless no commitment | $3.40 | $3.74 | 3% cheaper | cerebrium.ai checked Oct 2, 2026 CPU/RAM extra; H100 form factor not stated | |
AWS p5.48xlarge · 8 GPUs · us-east-1 | Spot no commitment | $3.46 | $3.81 | 1% cheaper | aws.amazon.com live Lowest current zone price in the region. Can be reclaimed with two minutes' notice. | |
Jarvislabs H100 SXM (1-8 GPUs per instance) | On-demand no commitment | $3.49 | $3.84 | 0% cheaper | jarvislabs.ai checked Oct 2, 2026 Rate from 2026-10-05. | |
RunPod Secure Cloud pod - H100 SXM | On-demand no commitment | $3.49 | $3.84 | 0% cheaper | runpod.io checked Oct 2, 2026 | |
Hyperstack H100 SXM reservation "starting from" · 8 GPUs | Reserved 1 mo term | $2.72 | $3.89 | 1% more | hyperstack.cloud checked Oct 2, 2026 | |
Scaleway H100-SXM-8-80G · 8 GPUs · fr-par-2 (Paris) | On-demand no commitment | $3.55 | $3.91 | 2% more | api.scaleway.com checked Oct 2, 2026 Public API price EUR 25.3308/hr per instance, converted at 1 EUR = 1.1225 USD (ECB rate 2026-10-02 via api.frankfurter.dev). Excludes VAT. | |
Verda 1H100.80S.30V reserved 24 mo | Reserved 2 yr term | $2.77 | $4.06 | 6% more | verda.com checked Oct 2, 2026 Derived: on-demand $3.7 minus published 25% reserved discount for 24-month commitment. | |
Together AI GPU Cluster HGX H100 reserved 7-30 days · 8 GPUs | Capacity block 8 days term, all upfront | $3.69 | $4.06 | 6% more | together.ai checked Oct 2, 2026 Reserved 7-30 days; reserved is pay upfront, up to 6 months on the clusters page; 181+ days contact sales. termMonths = shortest duration in the band. | |
Market median H100, all variants | Contract 2 yr term | $2.83 | $4.11 | 7% more | getdeploying.com weekly median Median listed price across 3 providers (9 listings), week of 2026-09-28; range $2.59–$3.30. Mixes H100 SXM, PCIe and NVL listings. Data: GetDeploying (CC BY 4.0), from the copy saved with the site. | |
AWS p5.48xlarge · 8 GPUs · us-east-1 | Reserved 3 yr term | $2.97 | $4.25 | 10% more | aws.amazon.com live Standard Reserved Instance, no upfront. | |
AWS p5.48xlarge · 8 GPUs · us-east-1 | Savings plan 3 yr term | $2.97 | $4.25 | 10% more | aws.amazon.com live EC2 Instance Savings Plan (locked to this instance family), no upfront. | |
Market median H100, all variants | Contract 3 yr term | $2.97 | $4.25 | 10% more | getdeploying.com weekly median Median listed price across 5 providers (18 listings), week of 2026-09-28; range $2.49–$5.40. Mixes H100 SXM, PCIe and NVL listings. Data: GetDeploying (CC BY 4.0), from the copy saved with the site. | |
Crusoe H100 80GB HGX | On-demand no commitment | $3.90 | $4.29 | 12% more | crusoe.ai checked Oct 2, 2026 | |
Modal Nvidia H100 SXM5 | Serverless no commitment | $3.95 | $4.34 | 13% more | modal.com live Billed per second, scales to zero. GPU functions can be preempted. Picking a region costs 1.15–1.75× more; CPU and memory are extra. | |
Lambda 8x NVIDIA H100 SXM · 8 GPUs | On-demand no commitment | $3.99 | $4.39 | 14% more | lambda.ai checked Oct 2, 2026 | |
Together AI GPU Cluster HGX H100 (on-demand) · 8 GPUs | On-demand no commitment | $3.99 | $4.39 | 14% more | together.ai checked Oct 2, 2026 Instant GPU Clusters, 8 to 256 GPUs, InfiniBand included, hourly, terminate anytime. | |
Market median H100, all variants | Contract 6 mo term | $3.09 | $4.41 | 15% more | getdeploying.com weekly median Median listed price across 9 providers (23 listings), week of 2026-09-28; range $2.58–$3.91. Mixes H100 SXM, PCIe and NVL listings. Data: GetDeploying (CC BY 4.0), from the copy saved with the site. | |
Market median H100, all variants | Contract 1 mo term | $3.17 | $4.53 | 18% more | getdeploying.com weekly median Median listed price across 10 providers (34 listings), week of 2026-09-28; range $1.53–$4.34. Mixes H100 SXM, PCIe and NVL listings. Data: GetDeploying (CC BY 4.0), from the copy saved with the site. | |
Market median H100, all variants | Contract 3 mo term | $3.18 | $4.54 | 18% more | getdeploying.com weekly median Median listed price across 5 providers (15 listings), week of 2026-09-28; range $2.30–$4.12. Mixes H100 SXM, PCIe and NVL listings. Data: GetDeploying (CC BY 4.0), from the copy saved with the site. | |
Together AI GPU Cluster HGX H100 reserved 91-180 days · 8 GPUs | Reserved 3 mo term, all upfront | $3.19 | $4.56 | 19% more | together.ai checked Oct 2, 2026 Reserved 91-180 days; reserved is pay upfront, up to 6 months on the clusters page; 181+ days contact sales. termMonths = shortest duration in the band. | |
Verda 8H100.80S.176V · 8 GPUs | On-demand no commitment | $4.17 | $4.59 | 19% more | verda.com live Formerly DataCrunch. 2-year reservations are 25% off. | |
DigitalOcean GPU Droplet H100 SXM 12-month reserved · 8 GPUs | Reserved 1 yr term | $3.26 | $4.66 | 21% more | digitalocean.com checked Oct 2, 2026 | |
Market median H100, all variants | Contract 1 yr term | $3.26 | $4.66 | 21% more | getdeploying.com weekly median Median listed price across 13 providers (35 listings), week of 2026-09-28; range $2.14–$8.03. Mixes H100 SXM, PCIe and NVL listings. Data: GetDeploying (CC BY 4.0), from the copy saved with the site. | |
Lambda 1x NVIDIA H100 SXM | On-demand no commitment | $4.29 | $4.72 | 23% more | lambda.ai checked Oct 2, 2026 | |
DigitalOcean GPU Droplet H100 SXM (1 or 8 GPUs) · 8 GPUs | On-demand no commitment | $4.41 | $4.85 | 26% more | digitalocean.com checked Oct 2, 2026 | |
Verda 1H100.80S.30V reserved 12 mo | Reserved 1 yr term | $3.40 | $4.86 | 27% more | verda.com checked Oct 2, 2026 Derived: on-demand $3.7 minus published 8% reserved discount for 12-month commitment. | |
Cirrascale 8x NVIDIA H100 (standalone), 12-month term ($19,999/server/month) · 8 GPUs | Contract 1 yr term | $3.42 | $4.89 | 27% more | cirrascale.com checked Oct 2, 2026 | |
Together AI GPU Cluster HGX H100 reserved 31-90 days · 8 GPUs | Reserved 1 mo term, all upfront | $3.45 | $4.93 | 28% more | together.ai checked Oct 2, 2026 Reserved 31-90 days; reserved is pay upfront, up to 6 months on the clusters page; 181+ days contact sales. termMonths = shortest duration in the band. | |
Nebius NVIDIA HGX H100 VM (1 or 8 GPUs) | On-demand no commitment | $4.50 | $4.95 | 29% more | nebius.com checked Oct 2, 2026 Price effective October 1, 2026 (was $3.85 before). Billed per second. Commitments: "up to 35% less" for multi-month large-cluster reservations, no published rates. InfiniBand on 8-GPU HGX clusters. | |
Google Cloud a3-highgpu-8g (DWS Flex-start) · 8 GPUs · us-central1 | Capacity block no commitment | $4.79 | $5.27 | 37% more | cloud.google.com checked Oct 2, 2026 Flex-start: waits in a queue until capacity is free, then runs at most 7 days, billed per use. A3 High; DWS prices are variable and Google says they will be updated Nov 2026. Same value shown on accelerator-optimized… | |
Google Cloud a3-highgpu-1g (DWS Flex-start) · us-central1 | Capacity block no commitment | $4.79 | $5.27 | 37% more | cloud.google.com checked Oct 2, 2026 Flex-start: waits in a queue until capacity is free, then runs at most 7 days, billed per use. A3 High 1-GPU shape: Flex-start or Spot only; DWS prices are variable and Google says they will be updated Nov 2026. Same… | |
RunPod Serverless flex worker - H100 PRO 80GB | Serverless no commitment | $4.79 | $5.27 | 37% more | runpod.io checked Oct 2, 2026 | |
Cirrascale 8x NVIDIA H100 (standalone), 6-month term ($22,499/server/month) · 8 GPUs | Contract 6 mo term | $3.85 | $5.50 | 43% more | cirrascale.com checked Oct 2, 2026 | |
Google Cloud a3-megagpu-8g (DWS Flex-start) · 8 GPUs · us-central1 | Capacity block no commitment | $5.04 | $5.54 | 44% more | cloud.google.com checked Oct 2, 2026 Flex-start: waits in a queue until capacity is free, then runs at most 7 days, billed per use. A3 Mega; DWS prices are variable and Google says they will be updated Nov 2026. Same value shown on accelerator-optimized… | |
Google Cloud a3-highgpu-8g · 8 GPUs · us-central1 | Capacity block no commitment | $5.20 | $5.72 | 49% more | cloud.google.com checked Oct 2, 2026 A3 High; DWS Calendar mode: future reservation of 1-90 days, pay for full reserved duration; 8-GPU shapes only | |
AWS p5.48xlarge · 8 GPUs · us-east-1 | Savings plan 1 yr term, all upfront | $4.05 | $5.78 | 50% more | aws.amazon.com live EC2 Instance Savings Plan (locked to this instance family), all upfront, shown as the effective hourly rate. | |
Cirrascale 8x NVIDIA H100 (standalone), 3-month term ($23,749/server/month) · 8 GPUs | Contract 3 mo term | $4.07 | $5.81 | 51% more | cirrascale.com checked Oct 2, 2026 | |
Replicate gpu-h100 ($0.001525/sec) | Serverless no commitment | $5.49 | $6.04 | 57% more | replicate.com checked Oct 2, 2026 private deploys bill setup+idle; H100 form factor not stated | |
Replicate gpu-h100-8x ($0.0122/sec) · 8 GPUs | Serverless no commitment | $5.49 | $6.04 | 57% more | replicate.com checked Oct 2, 2026 private deploys bill setup+idle | |
Google Cloud a3-megagpu-8g · 8 GPUs · us-central1 | Capacity block no commitment | $5.50 | $6.05 | 57% more | cloud.google.com checked Oct 2, 2026 A3 Mega; DWS Calendar mode: future reservation of 1-90 days, pay for full reserved duration; 8-GPU shapes only | |
Lambda 1-Click Cluster H100, 256 GPUs · 256 GPUs | Capacity block 15 days term | $5.54 | $6.09 | 59% more | lambda.ai checked Oct 2, 2026 | |
Cirrascale 8x NVIDIA H100 (standalone), 1-month term ($24,999/server/month) · 8 GPUs | Contract 1 mo term | $4.28 | $6.12 | 59% more | cirrascale.com checked Oct 2, 2026 | |
AWS p5.48xlarge · 8 GPUs · us-east-1 | Savings plan 1 yr term | $4.33 | $6.19 | 61% more | aws.amazon.com live EC2 Instance Savings Plan (locked to this instance family), no upfront. | |
Lambda 1-Click Cluster H100, 64 GPUs · 64 GPUs | Capacity block 15 days term | $5.85 | $6.43 | 67% more | lambda.ai checked Oct 2, 2026 | |
Paperspace H100 80GB | On-demand no commitment | $5.95 | $6.55 | 70% more | paperspace.com checked Oct 2, 2026 Footnote: "On-demand pricing for H100 is $5.95/hour under our special promo price." Page looks stale; DigitalOcean GPU Droplets is the current product. | |
AWS p5.48xlarge · 8 GPUs · us-east-1 | Savings plan 3 yr term, all upfront | $4.58 | $6.55 | 70% more | aws.amazon.com live Compute Savings Plan, all upfront, shown as the effective hourly rate. | |
AWS p5.48xlarge · 8 GPUs · all regions (non-GovCloud) | Capacity block no commitment | $5.97 | $6.57 | 71% more | aws.amazon.com checked Oct 2, 2026 Announced on the pricing page: $5.97/accelerator-hour effective 2026-10-07 in all available Regions. Instance price = per-GPU x 8 (computed). P6e-GB200 UltraServer and Trainium prices unchanged. The lower non-US P5… | |
AWS p5.4xlarge · all regions (non-GovCloud) | Capacity block no commitment | $5.97 | $6.57 | 71% more | aws.amazon.com checked Oct 2, 2026 Announced on the pricing page: $5.97/accelerator-hour effective 2026-10-07 in all available Regions. Instance price = per-GPU x 1 (computed). P6e-GB200 UltraServer and Trainium prices unchanged. Per-accelerator rate… | |
CoreWeave HGX H100 · 8 GPUs · North America | On-demand no commitment | $6.16 | $6.77 | 76% more | coreweave.com checked Oct 2, 2026 | |
Lambda 1-Click Cluster H100, 16 GPUs · 16 GPUs | Capacity block 15 days term | $6.16 | $6.78 | 76% more | lambda.ai checked Oct 2, 2026 1-Click Clusters: term 2 weeks to 1 year (termMonths = 2-week minimum). InfiniBand included. 1 year+ is contact sales. | |
Google Cloud a3-highgpu-8g · 8 GPUs · us-central1 | Reserved 3 yr term | $4.86 | $6.94 | 81% more | cloud.google.com checked Oct 2, 2026 A3 High; 3-yr resource-based committed use discount (Compute Resource CUD), billed monthly | |
Baseten Dedicated deployment H100 80GiB ($0.10833/min) | Serverless no commitment | $6.50 | $7.15 | 86% more | baseten.co checked Oct 2, 2026 per-minute billing; H100 MIG 40GiB $0.0625/min also listed | |
AWS p5.48xlarge · 8 GPUs · us-east-1 | Savings plan 3 yr term | $5.05 | $7.22 | 88% more | aws.amazon.com live Compute Savings Plan, no upfront. | |
AWS p5.48xlarge · 8 GPUs · us-east-1 | Savings plan 1 yr term, all upfront | $5.06 | $7.22 | 88% more | aws.amazon.com live Compute Savings Plan, all upfront, shown as the effective hourly rate. | |
Google Cloud a3-megagpu-8g · 8 GPUs · us-central1 | Reserved 3 yr term | $5.08 | $7.26 | 89% more | cloud.google.com checked Oct 2, 2026 A3 Mega; 3-yr resource-based committed use discount (Compute Resource CUD), billed monthly | |
Google Cloud a3-highgpu-8g · 8 GPUs · us-central1 | Spot no commitment | $6.62 | $7.28 | 89% more | cloud.google.com checked Oct 2, 2026 A3 High; current Spot price (variable, preemptible); same value on accelerator-optimized page | |
Google Cloud a3-highgpu-1g · us-central1 | Spot no commitment | $6.62 | $7.28 | 89% more | cloud.google.com checked Oct 2, 2026 A3 High 1-GPU shape: Flex-start or Spot only; current Spot price (variable, preemptible); same value on accelerator-optimized page | |
AWS p5.48xlarge · 8 GPUs · us-east-1 | On-demand no commitment | $6.88 | $7.57 | 97% more | aws.amazon.com live | |
Google Cloud a3-megagpu-8g · 8 GPUs · us-central1 | Spot no commitment | $6.99 | $7.69 | 100% more | cloud.google.com checked Oct 2, 2026 A3 Mega; current Spot price (variable, preemptible); same value on accelerator-optimized page | |
Azure ND96isr_H100_v5 · 8 GPUs · eastus | Reserved 3 yr term | $5.40 | $7.71 | 101% more | azure.microsoft.com live Reserved VM instance, paid up front or monthly | |
AWS p5.48xlarge · 8 GPUs · us-east-1 | Savings plan 1 yr term | $5.42 | $7.74 | 101% more | aws.amazon.com live Compute Savings Plan, no upfront. | |
Azure ND96isr_H100_v5 · 8 GPUs · eastus | Reserved 5 yr term | $4.92 | $10.3 | 168% more | azure.microsoft.com live Reserved VM instance, paid up front or monthly | |
Google Cloud a3-highgpu-8g · 8 GPUs · us-central1 | Reserved 1 yr term | $7.67 | $11.0 | 185% more | cloud.google.com checked Oct 2, 2026 A3 High; 1-yr resource-based committed use discount (Compute Resource CUD), billed monthly | |
Oracle Cloud BM.GPU.H100.8 · 8 GPUs · all regions | On-demand no commitment | $10.00 | $11.0 | 186% more | oracle.com live Bare metal. Committed discounts are negotiated, not published. | |
Azure ND96isr_H100_v5 · 8 GPUs · eastus | Reserved 1 yr term | $7.87 | $11.2 | 192% more | azure.microsoft.com live Reserved VM instance, paid up front or monthly | |
Google Cloud a3-megagpu-8g · 8 GPUs · us-central1 | Reserved 1 yr term | $8.03 | $11.5 | 198% more | cloud.google.com checked Oct 2, 2026 A3 Mega; 1-yr resource-based committed use discount (Compute Resource CUD), billed monthly | |
Google Cloud a3-highgpu-8g · 8 GPUs · us-central1 | On-demand no commitment | $11.06 | $12.2 | 217% more | cloud.google.com checked Oct 2, 2026 A3 High; on-demand list, us-central1 (page default region) | |
Google Cloud a3-megagpu-8g · 8 GPUs · us-central1 | On-demand no commitment | $11.68 | $12.8 | 234% more | cloud.google.com checked Oct 2, 2026 A3 Mega; on-demand list, us-central1 (page default region) | |
Azure ND96isr_H100_v5 · 8 GPUs · eastus | On-demand no commitment | $12.29 | $13.5 | 252% more | azure.microsoft.com live ND H100 v5 with InfiniBand |
“Your cost” is what each GPU-hour of real work costs at your utilization and horizon: pay-as-you-go prices plus overhead, committed prices divided by utilization. “vs owning” compares it with owning at $3.84/hr. List prices, mostly US regions.
Median on-demand price, week by week
H100 SXM: $2.95/hr in the week of Sep 2025, $3.36/hr now (+14%), median across 41 providers. The H100 series includes PCIe and NVL cards. Data: GetDeploying (CC BY 4.0), weekly medians.
H100 SXM Other GPUs