Independent advisory for colocation, GPU and data center projects

A reply within one business day

GPU Rental
GPU as a Service (GPUaaS), Dedicated GPU Servers and Clusters

Reserved H100, H200, B200 and B300 servers and clusters, compared across providers and negotiated for you. Tell us the GPU, count and term, and you'll hear back within 24 hours, at no cost to you.

Price My GPU Requirement
For AI teams renting GPU capacity Reserved terms negotiated for you Paid by the provider, never by you

Prices as of September 2026, updated monthly

GPU rental, also called GPU as a service (GPUaaS), is renting graphics processors for AI and other accelerated computing instead of buying them: by the hour, by the month, or reserved for a term, from a hyperscale cloud, a specialist GPU cloud known as a neocloud, or a provider of dedicated GPU servers; together these are the GPU as a service providers most teams compare. For most AI companies it is the fastest way to get capacity for inference, fine-tuning and training without committing capital to hardware the next generation will outrun. The hard part is not finding a price list. It is finding capacity that exists, on the hardware you need, on terms that hold up.

GPU CLOUD MEDIAN
$3.25
per H100 GPU-hour on demand, across 58 specialist providers
HYPERSCALERS
About $6.88
per H100 GPU-hour on demand at AWS, roughly twice the specialist median
RESERVED
25 to 45%
below on demand, typical for 3 to 12 month terms at cluster scale
Reserving a 64-GPU H100 fleet instead of renting on demand saves about $46,000 a month.
That is roughly $550,000 a year. We find the providers that can deliver it and negotiate the reservation, so the savings come with the delivery dates and terms that protect them, at no cost to you.
Price My GPU Requirement

Rent by GPU and Format

Each GPU and rental format has its own guide, with current prices, availability and the terms that matter. Start with the one that fits your workload.

Rent H100 GPUs

The most widely deployed data center GPU, 80 GB per GPU. Median about $3.25 per GPU-hour on demand. Also covers A100s, if you want to rent A100 servers for less.

H100 rental →

Rent H200 GPUs

141 GB per GPU for large models and long context. About $3.00 to $4.50 per GPU-hour on demand.

H200 rental →

Rent B200 and B300 GPUs

Current-generation Blackwell. B200 about $4.00 to $6.50, B300 median about $7.89 per GPU-hour.

B200 and B300 rental →

Rent GB200 and GB300 NVL72

Rack-scale Blackwell: 72 GPUs in one liquid-cooled rack. GB200 from about $10.50 per GPU-hour; GB300 by reservation.

GB200 and GB300 NVL72 →

GPU cluster rental

Servers joined by InfiniBand for training, priced by size tier and term.

GPU cluster rental →

Dedicated GPU servers

Bare metal GPU servers assigned to you alone, rented monthly or reserved.

Bare metal GPU servers →

Neocloud providers

The GPU-first clouds, compared on price, delivery and terms.

Neocloud providers →

Own instead of renting

GPU servers you buy, placed in colocation built for their density.

AI and GPU colocation →

GPU Rental Prices: What Each GPU Costs to Rent

Per GPU-hour unless noted, September 2026

GPUSpecialist cloud, on demandHyperscaler, on demandReserved, 3 to 12 months (est.)8-GPU server per month, on demandMemory per GPUMinimum termTime to live
H100Training and inference workhorseMedian about $3.25About $6.88About $1.80 to $2.45About $19,00080 GB1 month1 to 2 weeks
H200Large models, long contextAbout $3.00 to $4.50Varies by cloudAbout $2.05 to $2.80About $21,900141 GB1 month4 to 8 weeks
B200Current-generation trainingAbout $4.00 to $6.50From about $11.28About $2.90 to $3.95About $30,700180 GBAbout 3 months4 to 8 weeks
B300Largest models, most memoryMedian about $7.89From about $17.80About $4.35 to $5.90About $46,100288 GBAbout 6 months4 to 8 weeks
GB200 and GB300 NVL72Rack-scale, liquid-cooledQuoted per rackQuoted per rackNegotiated per rackRented as 72-GPU racksUp to 288 GB6 to 12 monthsPhased delivery
On-demand price per GPU-hour, specialist clouds
Midpoint of the ranges above
H10080 GB
$3.25
H200141 GB
$3.75
B200180 GB
$5.25
B300288 GB
$7.89

Sources: GetDeploying (H100, September 2026) and GetDeploying (B300); CloudZero (AWS H100, August 2026); GPUPerHour (H200, B200 and reserved discounts); hyperscaler B200 and B300 floors from FastGPU's September 2026 tracker. Reserved figures are Metro Colo Advisory estimates applying the typical 25 to 45 percent discount to the on-demand midpoint. Monthly figures assume 8 GPUs running around the clock. Prices move weekly; treat them as a starting point, not a quote.

Need GPUs in the next 90 days?
Send the GPU model, count, term, location and workload. You'll hear back within 24 hours from the person who will run your search. No signup, no call needed, and one contact instead of a sales team.
Price My GPU Requirement

What Does GPU as a Service Cost?

GPU as a service is priced in five main formats, and the gap between them is large, whether it is GPU server rental by the month or capacity by the hour. Current prices for each GPU are in the table above.

FormatHow it worksPrice versus on demandFits
On demandBy the hour or month, no commitmentFull priceExperiments, bursts, short projects
ReservedA fixed amount of capacity for a term, commonly 3 to 36 monthsTypically 25 to 45 percent lessSteady inference and planned training
Spot or interruptibleSpare capacity that can be reclaimed at short noticeOften around halfWork that can stop and resume from checkpoints
Dedicated GPU serverA whole physical server assigned to you, rented monthly or reservedMonthly rate, often close to reserved pricingSteady inference and fine-tuning that need consistent performance
GPU clusterServers joined by a high-speed network, reserved for a termPriced by size tier; steps down past 64 and 256 GPUsTraining and very large models

What the hourly rate leaves out

Storage

Datasets and checkpoints, often priced per terabyte on top of the GPUs.

Data transfer

Moving data out of the provider, which can rival the GPU bill for busy inference.

Cluster networking

InfiniBand between servers for training, usually priced into cluster rates.

Support and setup

Support tiers and onboarding fees that vary widely between providers.

On a large reservation, those lines can move the total as much as the headline rate, which is why every comparison we bring back shows them side by side.

Renting a GPU Cluster

A GPU cluster is nodes joined by a fabric fast enough that the GPUs spend their time computing, not waiting. At 64 GPUs and up, the fabric, delivery date, ramp and exit terms move more money than the hourly rate. Our GPU cluster rental guide covers cost by size, what to check in the quote, the terms to negotiate and when owning beats renting.

Planning to own the cluster and house it yourself? See AI and GPU colocation.

When Renting Through an Advisor Makes Sense

Short, on-demand capacity is often quickest to set up yourself; most GPU clouds let you sign up and launch a few GPUs the same day. For anything you will rent by the month or reserve, even a single server, send it to us. The value of an advisor grows with the size and length of the commitment, and it costs you nothing either way.

Often fine on your own

A few GPUs on demand for days or weeks: experiments and short jobs. Self-serve signup is fast, and there is little to negotiate.

Where we save the most

Around 64 GPUs and up, on reserved terms of three months or more: production inference fleets and training clusters, where delivery dates, contract terms and price over the term decide the outcome.

Single servers, rented monthly

Even one dedicated GPU server rented by the month is worth sending us. We compare providers and terms, and you pay nothing for it. The same goes for any requirement with a hard deadline, a compliance need or a choice between renting and owning.

GPU Rental Marketplace or Advisor?

A GPU rental marketplace lists on-demand and spot capacity from many sellers, so you can compare hourly prices and launch quickly. That works well for short jobs. Reserved servers and clusters are different: the listed price is a starting point, availability changes weekly, and the terms that matter, such as delivery dates, upgrade rights and data transfer, are negotiated rather than listed. That is the work an advisor does, and the provider pays for it.

Why Rent GPUs Through Us Instead of Going Direct

Going directThrough us
PriceThe provider's rate cardNegotiated against what the market actually pays
Cost to youNothing extraNothing extra; the provider pays us
Options comparedOnly the ones you find and callThe options that fit your requirement, compared for you
Contract termsNegotiated alone, against a deal teamNegotiated by someone who knows what providers accept
Getting startedWeeks of sales callsYou'll hear back within 24 hours
Sales contactA sales team for every provider you callOne contact
After you signAn account rep who works for the providerSomeone on your side through delivery and renewal
IN THE PRESS
Data Center Knowledge (August 2026) drew on our AI capacity searches in its reporting on usable high-density space, and Data Center POST (October 2026) published our analysis of why vacant capacity is often not usable for AI. Published availability and delivered availability are different things, and knowing the difference is most of the job.

How We Get You a Better Price

Market knowledge

A provider quoting a buyer alone quotes its rate card. A buyer who knows what comparable deployments sign for gets a different number.

Term structure

Prices step down at set term lengths, commonly 3, 6 and 12 months. We price each so you can see where the breaks are.

Size tiers

One major provider's published reserved H100 cluster rates fall from $6.16 to $5.85 to $5.54 per GPU-hour as clusters pass 64 and 256 GPUs, according to GPUPerHour. A few more GPUs can cost almost nothing extra.

Idle inventory

Providers with capacity sitting unused price it more sharply than providers who are sold out. We know who has what this month.

The terms that cost money later

Data transfer, storage, ramp schedules that bill servers only as delivered, and upgrade rights to newer GPUs. Often worth more than the headline rate.

Want these levers working on your quote?
Send the GPU, count, term and date. You'll hear back within 24 hours from the person who will run your search, and the provider you choose pays us.
Price My GPU Requirement

What GPU Providers Ask For

Providers screen buyers the way landlords do, and more as the commitment grows. On-demand capacity and small monthly rentals usually need only a card or prepayment. Reserved clusters and multi-month terms usually involve a review of financials, and newer companies may be asked for a deposit, a letter of credit or prepayment. Having that ready is often the difference between capacity next month and capacity next quarter. The AI and GPU colocation guide sets out what operators ask for at each deployment size.

What Should a GPU Rental Contract Include?

TermWhy it matters
A delivery date, with remedies if it slipsCapacity promised is not capacity delivered
The exact hardware, and upgrade rights during the termA year is a long time in GPU generations
Network and storage performanceTraining stalls on slow interconnect and slow storage
Bandwidth and data transfer chargesMoving datasets and serving traffic can rival the GPU bill
Uptime and hardware replacement commitmentsA failed GPU in an eight-GPU server can idle the whole machine
Data location and security attestationsCustomer data and model weights may carry obligations
Term, renewal price and exitThe next contract is negotiated from this one

Neocloud Providers or Hyperscalers?

A neocloud is a cloud built mainly to rent GPUs for AI, rather than a general-purpose cloud with GPUs added. Neocloud providers such as CoreWeave, Lambda, Crusoe and Nebius typically offer current-generation GPUs sooner, at lower prices for committed capacity, with networking built for training. Hyperscalers offer the surrounding services and one bill. Many AI companies use both: a hyperscaler for the application and a neocloud or dedicated provider for the GPUs, joined by a private connection. See neocloud providers and CoreWeave competitors and alternatives.

When Is Renting GPUs the Wrong Answer?

When a large fleet will run steadily for three years or more, owning the hardware, bought outright or leased, and placing it in colocation often costs less over its life. When a workload depends on one cloud's managed AI services, moving the GPUs elsewhere can cost more in rework than it saves. And for occasional experiments, hourly capacity or cloud credits are simpler than any contract. In each case, we will say so. The cloud repatriation guide covers the economics of moving steady workloads off public cloud.

HOW WE WORK
Placement, paid by the provider: we build the requirement, run the comparison, negotiate the contract and stay on the account after you sign. Consulting, paid by you, only if you want it: a fixed-fee rent-versus-own analysis for companies planning large fleets. See how a data center broker works and data center consulting.

Frequently Asked Questions

What is GPU rental?

GPU rental is paying for GPU computing capacity instead of buying the hardware: by the hour, by the month, or reserved for a term of 3 to 36 months. It ranges from a single GPU in a cloud to dedicated eight-GPU servers and multi-server training clusters. Metro Colo Advisory compares GPU rental options for your workload and negotiates the terms, at no cost to you.

What is GPU as a service (GPUaaS)?

GPU as a service is the cloud model of GPU rental: GPUs delivered as a service you rent rather than hardware you own, sold by hyperscale clouds, specialist GPU clouds known as neoclouds, dedicated GPU server providers and marketplaces. It lets AI companies get capacity quickly without committing capital to hardware.

How much does it cost to rent an H100?

Independent price trackers put the median specialist GPU cloud at about $3.25 per H100 GPU-hour on demand in September 2026, about $19,000 a month for an eight-GPU server running continuously, while hyperscalers charge roughly twice that. Reserved capacity on 3 to 12 month terms typically runs 25 to 45 percent lower. The rent H100 guide has the detail.

How much does it cost to rent an H200, B200 or B300?

On demand, H200 rental runs about $3.00 to $4.50 per GPU-hour, B200 about $4.00 to $6.50, and B300 at a market median near $7.89, according to independent price trackers in September 2026. Reserved terms cost less, and B200 and B300 supply is tighter, so delivery dates matter as much as price. See rent H200 and rent B200 and B300.

What is a dedicated GPU server, or bare metal GPU?

A whole physical server with its GPUs assigned to you alone, with no virtualization layer and no other tenants. It suits steady inference and fine-tuning where consistent performance and control matter, and it is usually rented monthly or reserved for a term. An eight-GPU H100 server runs about $19,000 a month on demand at the specialist median.

What is the difference between on-demand and reserved GPU capacity?

On-demand capacity is billed by the hour or month with no commitment and costs the most. Reserved capacity commits you to a fixed amount for a term, commonly 3 to 36 months, for a discount that is typically 25 to 45 percent. Reserved suits steady workloads; on-demand suits experiments and bursts.

Should I use a GPU rental marketplace or an advisor?

It depends on the size and length of the commitment. Marketplaces list on-demand and spot capacity from many sellers and work well for short jobs you can start yourself. For reserved servers and clusters, where delivery dates, contract terms and the price over a year matter, an advisor who knows what providers actually sign for usually saves more than a marketplace listing shows.

Can I lease GPUs instead of renting them?

Yes, but it is a different deal. Renting means paying for capacity a provider owns and runs. Leasing usually means financing GPU servers you will own or control, paid over a term, which then need a home in colocation. Leasing can make sense for large, steady fleets; renting is simpler for most teams.

Do GPU providers check credit?

For reserved capacity, yes, and more as the commitment grows. On-demand and small monthly rentals often need only a card or prepayment. Reserved clusters and multi-month terms usually involve a review of financials and may require a deposit, a letter of credit or prepayment, so preparing that early speeds things up.

What is a neocloud?

A cloud provider built mainly to rent GPUs for AI, such as CoreWeave, Lambda, Crusoe or Nebius, rather than a general-purpose cloud. Neoclouds often offer current-generation GPUs sooner and cheaper for committed capacity, with networking built for training. See neocloud providers.

Should we rent GPUs or buy them?

Rent when the workload is uncertain, bursty or short-lived, or when you need the newest hardware without a capital commitment. Owning, placed in colocation, often wins for large fleets that will run steadily for three years or more. The answer depends on fleet size, utilization and financing, and we can price both for your fleet.

How do I find GPU capacity when providers are sold out?

Take the requirement to more providers and more provider types, stay flexible on GPU generation and location where you can, and ask for committed delivery dates rather than listed availability. Older generations and secondary locations often have capacity when the newest hardware does not. Running that search is what we do, at no cost to you.

What should a GPU rental contract include?

A delivery date with remedies, the exact hardware and upgrade rights, network and storage performance, bandwidth charges, uptime and replacement commitments, data location and security attestations, and renewal and exit terms. Those terms decide whether a reservation is worth having.

Can you help with GPU clusters for training?

Yes. For training clusters we compare providers on the cluster network, storage throughput and delivery date as well as price, since a cluster is only as fast as its slowest link. Very large clusters have fewer providers and longer lead times, so starting early matters. See GPU cluster rental.

How do you get a better price than going direct?

By knowing what comparable deployments actually sign for, quoting each term length to find the price breaks, sizing to the provider's volume tiers, finding providers with idle inventory, and negotiating the terms that cost money later, such as data transfer, storage and upgrade rights. You do not pay more for it, because the provider pays us from its own sales budget.

How quickly will I hear back, and do I need a call?

You'll hear back within 24 hours from the person who will run your search, by email, and you do not need a call to get started. How fast providers can quote depends on the configuration; large clusters and rack-scale Blackwell systems take longer because fewer providers can deliver them.

Do I pay anything to use Metro Colo Advisory for GPU capacity?

No. The provider you choose pays us through its channel program, and you are under no obligation to use us. If you want a formal rent-versus-own analysis for a large fleet, that is a separate fixed-fee engagement, and only if you ask for it.

Tell Us What You Need to Run

The GPU model or performance target, how many, for how long, where, and what the workload is. That is enough to start, and you'll hear back within 24 hours from the person who will run your search. If renting is the wrong answer, we will tell you that.

Send it before you call a provider.
Once you have contacted a provider directly, it usually will not work through an advisor on your deal. The provider you choose pays us, so it costs you nothing.
Price My GPU Requirement