At the providers compared in this post, a GPU server costs between about $0.71 and $10 per GPU-hour on demand in November 2026, depending on the card and the provider. Run around the clock for a 730-hour month, a 24 GB card costs $515.99 to $2,336.00, a 96 GB NVIDIA RTX PRO 6000 Blackwell costs $1,963.70 to $5,741.45, and a single H100 costs $2,766.70 to $7,300.00. A dedicated server billed monthly costs less than those on-demand prices when the GPU is busy most of the month. Below the break-even hours, paying by the hour costs less.

This post puts the published prices side by side, all checked on 16 November 2026, and works out the number of hours a month at which a fixed monthly bill becomes the lower one. MassiveGRID sells four GPU dedicated servers, so our own product is in the tables and labelled as such. Where paying by the hour is the better choice, the post says so.

How much does a GPU server cost per hour and per month?

The table lists on-demand prices for one GPU, grouped by memory class, at the providers compared in this post. It is not a survey of the whole market. The monthly column is the hourly rate multiplied by 730 hours, which is what the instance costs if it is never stopped. Where a provider bills by the month, the monthly figure is the one it prints itself. Prices are in US dollars before tax.

Provider and instanceGPUPer hour, on demand730-hour month
24 GB class
MassiveGRID RTX PRO 4000 server (ours, billed monthly)RTX PRO 4000 Blackwell, 24 GB$0.62 equivalent$449.99
Google Cloud g2-standard-4, IowaL4, 24 GB$0.7068$515.99
AWS g6.xlarge, us-east-1L4, 24 GB$0.8048$587.50
AWS g5.xlarge, us-east-1A10G, 24 GB$1.006$734.38
Liquid Web bare metal, listed regular priceL4, 24 GB$1.07$781.10
Paperspace A5000RTX A5000$1.38$1,007.40
OVHcloud Scale-GPU-1 bare metal, billed monthlyNVIDIA L4-from $1,145 (OVHcloud's figure, plus an equal installation fee)
Oracle Cloud VM.GPU.A10.1, list priceA10, 24 GB$2.00$1,460.00
Azure Standard_NV36ads_A10_v5, East USA10, 24 GB$3.20$2,336.00
96 GB class (RTX PRO 6000 Blackwell)
MassiveGRID RTX PRO 6000 server (ours, billed monthly)RTX PRO 6000 Blackwell, 96 GB$2.27 equivalent$1,656.15
RunPod Secure Cloud96 GB$2.69$1,963.70
Akamai Cloud, RTX PRO 6000 Blackwell x196 GB$3.00$2,190.00
Modal, serverless, GPU only96 GB$3.0312$2,212.78
AWS g7e.2xlarge, us-east-196 GB$3.36312$2,455.08
Google Cloud g4-standard-48, Iowa96 GB$4.49993$3,284.95
Oracle Cloud, RTX PRO 6000, list price96 GB$4.50$3,285.00
Azure NC144lds_xl_RTXPRO6000BSE_v6, East US96 GB$5.50$4,015.00
AWS g7e.2xlarge, Frankfurt96 GB$5.71915$4,174.98
Azure NC144lds_xl_RTXPRO6000BSE_v6, West Europe96 GB$7.865$5,741.45
Four or more 96 GB cards
MassiveGRID 4x RTX PRO 6000 server (ours, billed monthly)4x RTX PRO 6000 Blackwell, 384 GB$8.17 equivalent ($2.04 per GPU)$5,966.44 ($1,491.61 per GPU)
Vultr bare metal, billed monthly4x A100$9.589 equivalent$7,000 (Vultr's figure)
Akamai Cloud, RTX PRO 6000 Blackwell x44x 96 GB$12.00$8,760.00
AWS g7e.24xlarge, us-east-14x 96 GB$16.57216$12,097.68
Google Cloud g4-standard-192, Iowa4x 96 GB$17.99972$13,139.80
CoreWeave, 8-GPU instance only8x 96 GB$20.00 per instance ($2.50 per GPU)$14,600.00 ($1,825.00 per GPU)
Latitude.sh g4.rtx6kpro.large bare metal, 8 GPUs only8x 96 GB$24.00 per server ($3.00 per GPU)$17,520 (Latitude.sh's figure; $2,190.00 per GPU)
H100, 80 GB (market reference)
Lambda H100 PCIe (SXM: $4.89)H100, 80 GB$3.79$2,766.70
AWS p5.4xlarge, us-east-1H100, 80 GB$6.88$5,022.40
Azure Standard_NC40ads_H100_v5, East USH100, 80 GB$6.98$5,095.40
Oracle Cloud, H100, list priceH100, 80 GB$10.00$7,300.00

Sources, all checked on 16 November 2026: AWS, Azure, Google Cloud, Oracle and Vultr figures come from their price files and pricing APIs (Vultr listed no location for its GPU plans that day), OVHcloud from its bare-metal price list, and the Akamai, Modal, RunPod, Paperspace, Liquid Web, Latitude.sh, CoreWeave and Lambda figures are the prices listed on their pricing pages. Modal bills per second ($0.000842) with CPU and memory extra, so its row is the GPU alone as if always on.

The instances are not identical around the GPU. AWS g7e.2xlarge has 8 vCPUs and 64 GiB of RAM, Google's g4-standard-48 has 48 vCPUs and 180 GiB, and the Azure size has 144 vCPUs and 264 GiB; the MassiveGRID 96 GB server has a 20-core CPU and 128 GB of RAM to itself. The cloud instances carry the Server Edition of the card and our servers the Max-Q edition. The 24 GB rows are older cards (the A5000's 24 GB is NVIDIA's figure, not Paperspace's), OVHcloud does not state how many L4 cards its server carries, and the Vultr server uses A100 cards.

What do the four MassiveGRID GPU servers cost?

The servers are bare-metal machines with the whole GPU, CPU, memory and NVMe storage for one customer and root access. They are billed monthly (quarterly up to triennial terms are available), not by the hour, and each has a one-time setup fee. The hourly equivalent is the monthly price divided by 730.

ServerGPU memoryMonthlySetup, onceFirst month in totalHourly equivalentPer GB of GPU memory
RTX PRO 4000 (Core i5-13500, 64 GB RAM, 2x 512 GB NVMe)24 GB$449.99$346.48$796.47$0.62$18.75
RTX PRO 6000 (Core Ultra 7 265, 128 GB RAM, 2x 1.92 TB NVMe)96 GB$1,656.15$1,656.15$3,312.30$2.27$17.25
RTX PRO 6000 Xeon (Xeon Gold 5412U, 256 GB ECC RAM, 2x 960 GB NVMe)96 GB$2,485.05$2,485.05$4,970.10$3.40$25.89
4x RTX PRO 6000 (Xeon Gold 5412U, 128 GB ECC RAM, 2x 1.92 TB NVMe)384 GB$5,966.44$5,966.44$11,932.88$8.17 ($2.04 per GPU)$15.54

Every server includes a 1 Gbps port with unlimited traffic and one dedicated IPv4 address, and is hosted in an ISO 27001 certified data center. If the location matters for latency or data residency, ask sales which data center a server will be delivered in before ordering. NVIDIA H100, H200 and B200 systems are not in the store and are quoted on request.

The setup fee changes the first-year arithmetic. Spread over twelve months, the 96 GB server costs $1,794.16 a month (thirteen payments of $1,656.15 divided by twelve), or $2.46 per hour. The 24 GB server works out at $478.86 a month, or $0.66 per hour. From the thirteenth month the plain monthly price applies.

At how many hours a month does a dedicated GPU server win?

The break-even point is the monthly price of the dedicated server divided by the on-demand hourly price of the alternative. Below that many hours a month, paying by the hour costs less, provided you actually stop the instance when idle. Above it, the fixed monthly price is the lower one. A month has 730 hours.

On-demand alternativeHourly priceBreak-even hours per monthShare of the month
96 GB server, $1,656.15 a month
Azure NC144lds, West Europe$7.86521129%
AWS g7e.2xlarge, Frankfurt$5.7191529040%
Azure NC144lds, East US$5.5030141%
Oracle Cloud, RTX PRO 6000$4.5036850%
Google Cloud g4-standard-48, Iowa$4.4999336850%
AWS g7e.2xlarge, us-east-1$3.3631249267%
Modal, GPU only$3.031254675%
Akamai Cloud$3.0055276%
RunPod Secure Cloud$2.6961684%
24 GB server, $449.99 a month
Azure NV36ads_A10_v5, East US$3.2014119%
Oracle Cloud A10$2.0022531%
Paperspace A5000$1.3832645%
AWS g5.xlarge (A10G), us-east-1$1.00644761%
AWS g6.xlarge (L4), us-east-1$0.804855977%
Google Cloud g2-standard-4 (L4), Iowa$0.706863787%
4x 96 GB server, $5,966.44 a month
Google Cloud g4-standard-192 (4 GPUs), Iowa$17.9997233145%
AWS g7e.24xlarge (4 GPUs), us-east-1$16.5721636049%
Akamai Cloud, 4 GPUs$12.0049768%

In the first year, with the setup fee spread over twelve months, the 96 GB break-even moves from 492 to 533 hours against AWS us-east-1 and from 616 to 667 hours against RunPod Secure Cloud.

The pattern: against Azure, Google Cloud and Oracle the 96 GB server pays for itself once the GPU is busy for 30% to 50% of the month, against AWS in N. Virginia two thirds, and against Modal, Akamai and RunPod Secure Cloud three quarters or more. Beyond the hourly figure, the dedicated server brings a fixed bill, the whole machine for one tenant and unlimited traffic on its 1 Gbps port.

Which costs are not in the hourly rate?

The hourly rate is the largest line, but three others decide close calls.

Two costs sit on the dedicated side only. A dedicated server is a single machine, so redundancy and backups are your job: mirror the two NVMe drives with RAID and keep copies of anything you cannot re-download somewhere else. And the server arrives with a clean operating system and root access, so the first hour goes on drivers and containers. The first-hour setup guide for Ubuntu covers that.

Scenario 1: 40 hours a month of experiments

A developer tests prompts and fine-tuning runs on a 96 GB card for about two hours each working day, 40 hours a month.

OptionCalculationMonthly cost
RunPod Secure Cloud40 x $2.69$107.60
Akamai Cloud40 x $3.00$120.00
AWS g7e.2xlarge, us-east-140 x $3.36312$134.52
Google Cloud g4-standard-48, Iowa40 x $4.49993$180.00
Azure NC144lds, East US40 x $5.50$220.00
MassiveGRID 96 GB serverfixed$1,656.15

Result: at 40 hours a month, paying by the hour costs less. The fixed monthly price is 7.5 to 15.4 times the on-demand bills in this table. The break-even against these five prices lies between 301 hours a month (Azure, East US) and 616 hours (RunPod Secure Cloud). Rent by the hour, keep your environment in a container image so a fresh instance is usable in minutes, and look at a dedicated server again when the hours grow. Our own rent versus buy comparison reaches the same conclusion for occasional use.

Scenario 2: 300 hours a month for a team

A team runs a 96 GB model during working hours, about 14 hours a day on 22 days: roughly 300 hours a month.

OptionCalculationMonthly cost
RunPod Secure Cloud300 x $2.69$807.00
Akamai Cloud300 x $3.00$900.00
AWS g7e.2xlarge, us-east-1300 x $3.36312$1,008.94
Google Cloud g4-standard-48, Iowa300 x $4.49993$1,349.98
Azure NC144lds, East US300 x $5.50$1,650.00
MassiveGRID 96 GB serverfixed$1,656.15
AWS g7e.2xlarge, Frankfurt300 x $5.71915$1,715.75

Result: it depends on the region and on whether you really stop the instance. At 300 hours the dedicated server costs about the same as Azure in East US ($1,650.00), less than AWS in Frankfurt ($1,715.75), and more than the other four on-demand bills, whose break-even points are at 368 hours or above. Below the break-even, paying by the hour costs less, and that saving depends on discipline: someone or something has to stop the instance every evening and start it every morning, the model has to reload each time, and an instance left running over two weekends adds about 96 hours. If the workload cannot be stopped, because users in other time zones or scheduled jobs need it at night, this scenario becomes the next one.

Scenario 3: an inference endpoint that runs 24/7

A product calls a self-hosted model at any hour, so the GPU stays loaded for all 730 hours.

Option96 GB, one GPU24 GB, one GPU
MassiveGRID dedicated server$1,656.15 ($1,794.16 in year one with setup)$449.99 ($478.86 in year one with setup)
RunPod Secure Cloud$1,963.70$503.70 (L4)
Akamai Cloud$2,190.00-
AWS, us-east-1$2,455.08$587.50 (L4), $734.38 (A10G)
Google Cloud, Iowa$3,284.95$515.99 (L4)
Oracle Cloud, list price$3,285.00$1,460.00 (A10)
Azure, East US$4,015.00 (NC144lds)$2,336.00 (A10)
AWS, Frankfurt$4,174.98$734.67 (L4)

Result: the dedicated server costs less than every on-demand price compared here. Against AWS us-east-1 the 96 GB server saves $798.93 a month, and against Frankfurt $2,518.83, before AWS storage and egress. Against Google Cloud in Iowa it saves $1,628.80 a month and against Azure in East US $2,358.85. The gap is smallest against Akamai, $533.85 a month, and RunPod Secure Cloud, $307.55 a month or $169.54 in the first year. The dedicated server also gives a bill that does not change with usage or traffic and a whole machine that no other tenant shares, with billing terms of up to three years. These are on-demand prices, without commitment discounts.

With four GPUs the gap widens: Google Cloud's g4-standard-192 costs $13,139.80 for 730 hours in Iowa, AWS g7e.24xlarge $12,097.68 in us-east-1 and Akamai's four-GPU plan $8,760.00, against $5,966.44 for the 4x RTX PRO 6000 server.

How to choose between monthly and hourly

  1. Count GPU-hours honestly. Take last month's real usage, not the plan. Under about 300 hours, rent by the hour.
  2. Check whether you can stop. An endpoint with users, a queue that runs overnight or a model that takes minutes to load pushes real usage towards 730 hours.
  3. Add storage and traffic. Multiply your stored gigabytes and your monthly egress by the provider's rates before comparing.
  4. Decide what the fixed bill is worth. If procurement, privacy review or capacity planning matter more than the last few hundred dollars, a single-tenant server is the simpler thing to explain. The GPU cloud alternative page sets out that case, and private AI hosting covers the data side.
  5. Use both. A dedicated server for the steady load and hourly instances for bursts and large training runs is often the sensible split. H100-class cards rented by the hour remain the right tool for large-scale training.

If the endpoint serves an LLM, compare against API pricing too: the self-hosted LLM versus API break-even works out how many tokens a month it takes before a server costs less than premium API list prices.

Frequently Asked Questions

How much does a GPU server cost per month?

On 16 November 2026, at the providers compared here, a 24 GB GPU running on demand for a full 730-hour month costs from $515.99 (Google Cloud g2-standard-4, Iowa) to $2,336.00 (Azure NV36ads_A10_v5, East US), a 96 GB RTX PRO 6000 Blackwell from $1,963.70 (RunPod Secure Cloud) to $5,741.45 (Azure NC144lds, West Europe), and a single H100 from $2,766.70 (Lambda) to $7,300.00 (Oracle's list price). MassiveGRID's dedicated GPU servers are $449.99 for 24 GB, $1,656.15 for 96 GB and $5,966.44 for four 96 GB cards, plus a one-time setup fee.

Is a dedicated GPU server cheaper than AWS?

Above a certain number of hours, yes. The 96 GB server at $1,656.15 costs less than an AWS g7e.2xlarge on demand once the instance would run more than about 492 hours a month in us-east-1 or 290 hours in Frankfurt, before AWS storage and its $0.09 per GB egress. Below that many hours a month, paying by the hour costs less.

Is there a setup fee on a dedicated GPU server?

On ours, yes, and it is stated in the store: $346.48 on the RTX PRO 4000 server and one month's price on the three RTX PRO 6000 servers. The first month therefore totals $796.47, $3,312.30, $4,970.10 or $11,932.88. Hourly clouds have no setup fee, which is one reason paying by the hour suits short projects.

What does an RTX PRO 6000 Blackwell cost per hour at Azure, Google Cloud and Oracle?

On 16 November 2026 the on-demand price for one 96 GB card was $5.50 an hour at Azure in East US (NC144lds), $4.49993 at Google Cloud in Iowa (g4-standard-48) and $4.50 at Oracle's list price. Over 730 hours that is $4,015.00, $3,284.95 and $3,285.00, between 1.98 and 2.42 times the $1,656.15 monthly price of our 96 GB server, which adds a one-time setup fee of the same amount. Those instances can be stopped when idle.

A GPU server with a fixed monthly bill

MassiveGRID GPU dedicated servers are single-tenant bare metal with NVIDIA RTX PRO Blackwell cards: 24 GB for $449.99 per month, 96 GB from $1,656.15, or 384 GB across four cards for $5,966.44, each with a one-time setup fee stated in the store. Every server has a 1 Gbps port with unlimited traffic, a dedicated IPv4 address and root access, with no per-hour metering and no egress charges.

Compare with GPU Cloud

Further Reading