Dedicated GPU server vs cloud GPU: cost, privacy and speed

For steady use, a dedicated GPU server rented by the month costs less than a cloud GPU billed by the hour; for short bursts, the hourly cloud wins. The break-even point is the monthly price divided by the hourly rate: our RTX 4090 at $84.99 a month equals 85 hours at an example rate of $1.00 an hour.

This guide compares a dedicated GPU server vs cloud GPU rental on cost, availability, performance and privacy, using our own GPU server prices and example hourly rates, never another provider's quotes.

The short answer: dedicated GPU server vs cloud GPU

A cloud GPU is capacity by the hour: you start an instance, pay while it runs and stop it. A dedicated GPU server has one or more whole GPUs rented by the month, running around the clock whether you use it or not. The cloud sells flexibility; the dedicated server sells a lower price per hour of use and a card nobody else touches.

• Rent monthly when the card works most days: a model serving users, a team's coding assistant, recurring fine-tuning, render queues.
• Rent hourly for a few hours of testing, a one-off training run, or peaks far above your baseline.
• Combine both when a steady base load has occasional peaks.

GPU server vs cloud GPU cost: find your break-even hours

To decide whether to rent a GPU monthly vs hourly, divide the monthly price by the hourly rate you are offered. Above that many hours a month, the monthly server is cheaper:

break-even hours = monthly price ÷ hourly rate
share of the month = break-even hours ÷ 730

The table applies it to our RTX 4090 and H100 prices. The hourly rates are example rates, not quotes from any provider: use the rates you are offered.

So if an RTX 4090-class card costs you $0.70 an hour and you use it more than about four hours a day, the month is cheaper. Around the clock, our monthly prices work out to about $0.085 an hour for the RTX A4000, $0.116 for the RTX 4090, $0.185 for the RTX 5090, $0.488 for the A100 80 GB and $0.797 for the H100. That is before the bonus on balance top-ups (+10% from $100 up to +50% from $1,000), which pays for servers and renewals but cannot be withdrawn.

Hourly bills can also hide costs. Check what you pay for storage while an instance is stopped and for data leaving the cloud, and count the time each fresh instance spends loading weights: the 43 GB download of a 4-bit 70B model takes about six minutes even at a full 1 Gbps. A forgotten instance bills all night. A monthly server has no meter, and bandwidth on ours is unmetered under fair use.

The cheapest way to rent an H100

It depends on your hours. Below about 145 to 291 hours a month, at example rates of $4 down to $2 an hour, hourly rental costs less; above that, our H100 80 GB SXM5 server at $581.99 a month does. Quarterly billing brings it to $552.89 a month and yearly billing to $512.15, about $0.70 an hour around the clock. If 80 GB at lower bandwidth is enough, the A100 80 GB costs $355.99.

Availability: spot and on-demand capacity vs a card that is yours

Hourly GPUs come in two kinds, and neither promises the card will be there when you need it.

• On-demand instances bill the full rate but depend on free capacity. AWS, for example, documents an InsufficientInstanceCapacity error for when it "does not currently have enough available On-Demand capacity", and sets default per-region instance limits that you ask to raise.
• Spot or preemptible instances cost less because the provider can take them back: on AWS, the interruption notice comes two minutes before the instance is stopped or terminated. Training runs need frequent checkpoints, and endpoints need a fallback.

A dedicated card is yours for the period you paid: no reclaim notice, no capacity check when you restart. The limits come before delivery. Delivery takes 1 to 24 hours, stock depends on the card and the jurisdiction, and if a configuration is temporarily unavailable we tell you in the client area and either hold the order or return the payment to your balance. GPU servers are not refundable once delivered.

Performance and consistency

On a dedicated server, each GPU is a whole card that only your jobs use: no time-slicing, no partition. A cloud GPU is often a whole card too, but not always:

• MIG partitions. NVIDIA's Multi-Instance GPU splits an A100 or H100 into up to seven instances, each with its own memory, cache and compute cores: isolated, but a fraction of the card.
• Time-slicing. Several workloads take turns on one GPU, with, as NVIDIA's documentation puts it, no memory or fault isolation between them.

Memory capacity decides which models fit and bandwidth how fast they generate, as our guide on VRAM for LLMs explains, so check that a low price buys a whole GPU.

Storage next to the GPU. Our GPU servers have local NVMe, from 1 TB on the RTX 4090 to 30 TB on the 4 × H100 server, which keeps models and checkpoints between sessions. Fast local disks in the cloud can be temporary: on AWS, instance store data does not persist when the instance is stopped, hibernated or terminated, so each fresh instance loads the weights again.

Consistency. The same machine every day keeps the same driver, a warm model cache and kernels compiled on the first start. The trade-off is scale: our largest server has four H100s linked by NVLink, behind a 1 Gbps port. Our guide to self-hosting an LLM on a GPU server covers the setup.

Privacy and identity

Large GPU clouds generally want to know who you are before they hand over expensive cards: an account tied to a payment card or a bank, verification steps, default usage limits. Your prompts and datasets then sit under logging and retention rules you accept rather than set.

At OffshoreServ, an account is an email address, which can be disposable, and a password: no name, postal address, phone number or ID documents. You top up a USD balance in Bitcoin, Ethereum, Monero, Tether (USDT) or Solana through a payment gateway that never receives your email or account details; our crypto payments page lists the confirmation times.

We do not log or inspect server traffic, with no deep packet inspection and no content scanning, and we never see or log your prompts, outputs or training data. Our security log keeps a browser and OS summary and a country, never an IP address.

Privacy is not immunity. GPU servers run in Moldova, the Netherlands, Iceland or Romania, depending on the card, under local law, and in the Netherlands and Romania under the EU Digital Services Act too. Mining is allowed on GPU servers, and our acceptable use policy applies to everything you run.

When the cloud is the better choice

• Short experiments. Three hours on an H100 at an example $4 an hour cost $12; our month costs $581.99. If the card would sit idle most of the month, rent by the hour.
• More than four GPUs. Our largest server has four H100s and 320 GB of VRAM. Training across dozens of GPUs belongs on a cloud cluster; existing GPU and dedicated customers can ask us for a custom build by ticket.
• Autoscaling across regions. If traffic swings tenfold within a day, or you need capacity near users on several continents within minutes, on-demand clouds fit better: our servers take 1 to 24 hours to deliver, in four jurisdictions.

A dedicated server also leaves the operating system, serving stack and security to you. A managed service costs more but runs them for you.

Dedicated GPU server vs cloud GPU: a decision checklist

1. Count your hours. Above the monthly price divided by the hourly rate, rent monthly.
2. Size the memory. Up to 80 GB, headroom included, fits one A100 or H100; up to 320 GB, a 4 × H100 server; beyond that, a cloud cluster.
3. Check interruption tolerance. A job that cannot checkpoint, or an endpoint users rely on, should not run on spot capacity.
4. Check where data may go. If prompts or datasets must not reach a third-party platform, keep them on a dedicated server.
5. Check identity requirements. If you will not tie the work to your identity or a card, choose a provider that takes crypto without KYC.
6. Plan the start. Need a GPU in ten minutes? Use the cloud. Can you wait up to a day and commit to a month? Rent dedicated.
7. Consider both. A monthly server for the base load and hourly GPUs for peaks is often the cheaper mix.

Frequently asked questions

Is it cheaper to rent a GPU monthly or hourly?

Monthly, once you use the card for more hours than the monthly price divided by the hourly rate. Our RTX 4090 costs $84.99 a month, so at an example rate of $0.70 an hour the month wins after about 121 hours, roughly four hours a day. Below that, hourly rental costs less.

What is the difference between a GPU dedicated server and a cloud GPU?

A GPU dedicated server gives you whole GPUs for a monthly period, with local storage that keeps your data. A cloud GPU is rented by the hour and started at will; it can be a whole card, a MIG partition or a time-slice, and spot capacity can be reclaimed. The first is cheaper for steady work, the second for short bursts.

Can I rent an H100 without KYC?

Yes. At OffshoreServ, an account is an email address and a password, with no name, phone number or ID documents, and you pay in Bitcoin, Ethereum, Monero, Tether or Solana. An H100 80 GB SXM5 server costs $581.99 a month and is ready in 1 to 24 hours. Local law and our acceptable use policy still apply.

Are cloud GPUs shared?

Not always, but they can be. A cloud GPU may be a whole card, a MIG partition (up to seven isolated instances on an A100 or H100) or a time-sliced share without memory or fault isolation. Check what an offer includes before you compare prices. On our GPU servers, every card is dedicated to your server and never shared.

How many hours a month make a dedicated GPU worth it?

Divide the monthly price by the hourly rate you would otherwise pay. For our $84.99 RTX 4090, that is 212 hours at an example $0.40 an hour and 85 hours at $1.00; for our $581.99 H100, 291 hours at $2 and 145 hours at $4. Above those numbers, the month is cheaper.