Vultr Review 2026: GPU by the Hour from $0.47

Vultr's cheapest GPU is $0.471 per GPU per hour, and that number changes what is affordable — a few hours of experimentation costs a few dollars. It is also not the GPU for model inference, and the $2.50/month instance cannot run the application that calls a model at all. The question is not whether Vultr is cheap. It is which tier you actually need, and how many hours you will pay for.

The real decision is hours, not hardware

Vultr’s lowest listed GPU rate is $0.471 per GPU per hour. That is not a rounding error; it is a meter, and it runs whether or not a job is using the card.

Hourly rental changes what is affordable and nothing about what is permanent. A few hours of experimentation is cheap; a service that answers requests at three in the morning is not, because you are paying for three in the morning.

Every figure here comes from Vultr’s own pricing pages, checked on 25 September 2026. Where a number is the vendor’s marketing rather than a price, it is labelled as a claim rather than repeated as a finding.

The argument is two sentences long: the cheapest GPU is not the GPU for model inference, and the cheapest instance cannot run the application that talks to the model.

The GPU ladder, quoted per GPU per hour

The “per GPU” is the part buyers miss. The A16 is listed in 1, 2, 4, 8 and 16-GPU configurations and the rate is per card, so an 8-GPU A16 instance bills $3.767 per hour. The A40 is $1.712 for one GPU and $6.849 for four; the PCIe A100 is $2.397 for one and $19.178 per hour for eight.

What the cheapest line is for. The A16 is the lowest entry on the page and is aimed at virtual desktops and light graphics, not at large language model inference. For a remote desktop or a modest rendering job the rate is genuinely low; for a model it is the wrong tier, and buying it because it is cheapest is the predictable mistake.

What the inference lines are. Vultr positions the GH200 at $1.990 per GPU per hour and the PCIe A100 at $2.397 as the compute for model inference — four to five times the A16’s rate, per GPU, per hour. Above them sit multi-GPU-only lines quoted per GPU per hour from $1.85 to $3.50 — AMD’s MI300X at $1.850, NVIDIA’s HGX B300 at $3.500. Those sell as multi-GPU instances, so the per-GPU figure is a unit price, not a bill.

One hour is cheap; every hour is not. An 8-GPU PCIe A100 instance left up bills $19.178 for every hour it exists, job or no job. Hourly rental is for trying a pipeline or clearing a bounded batch, not for hosting an inference service cheaply.

The CPU side, where the app that talks to the model lives

Plan	vCPU	RAM	Storage	Bandwidth	Monthly	Hourly
Cloud Compute — Regular Performance	1 shared	0.5 GB	10 GB	0.5 TB	$2.50	$0.004
Cloud Compute — High Performance	1 shared	1 GB	25 GB NVMe	2 TB	$6.00	$0.009
Optimized — General Purpose	1 dedicated	4 GB	30 GB NVMe	4 TB	$30.00	$0.045
Optimized — Storage Optimized	1 dedicated	8 GB	150 GB NVMe	4 TB	$75.00	$0.112

The $2.50 tier is 1 vCPU, 0.5 GB of RAM, 10 GB of storage and 0.5 TB of bandwidth. It is the cheapest way onto the platform and it is not an application server: half a gigabyte is below what a runtime, a package manager and a reverse proxy consume between them, before your code arrives. It fits a DNS resolver, a tunnel endpoint or an SSH jump box.

The $6 High Performance tier is where this stops being an experiment. One GB is still small, but it holds a small application server, a queue worker or a static site with a light process behind it, and storage moves to NVMe. Rent a GPU for a few hours and this is the machine to put beside it, not the $2.50 tier.

The $30 General Purpose tier is the first with dedicated vCPU, so the processor is not shared with other customers, plus 4 GB of RAM. That is the first configuration worth calling a production baseline, because an application, a small database and a worker can share it without competing for someone else’s CPU. The $75 Storage Optimized tier trades CPU for disk — still 1 vCPU, but 8 GB of RAM and 150 GB of NVMe — where capacity, not compute, is the constraint.

Two lines go far past 1 vCPU. VX1 starts at 2 vCPU, 8 GB and 5 TB of block storage at $0.060/hour and reaches 192 vCPU, 768 GB and 10 TB at $5.760/hour; Memory Optimized starts at 2 vCPU, 16 GB and 5 TB at $0.080/hour and reaches 192 vCPU, 1536 GB and 10 TB at $9.258/hour. Vultr’s own claim for VX1 — “up to 82% better performance per dollar compared to leading hyperscalers” — is a vendor claim, not a measurement made here.

Hourly billing: right for experiments, wrong for anything permanent

Each plan carries both an hourly and a monthly price: the $2.50 instance at $0.004/hour and $2.50/month, the $75 instance at $0.112/hour and $75.00/month.

Hourly is the right frame when the work is bounded — a migration, a build that runs twice a day, a benchmark campaign, a training run you can stop. Hours are what you buy and the price follows them.

It is the wrong frame when the workload never stops. A website, an API or an inference service exists 24 hours a day, and then the hour count stops being something you control.

The tempting middle case is leaving a GPU instance up “just in case”: a stopped instance does no work, a running one is paid for. Either the card is in use or it is a subscription to unused capacity.

The base price is not the bill

Charge	Price	What it covers
Automatic backups	+20% of the base fee	The instance’s own backups, billed as a percentage of what you already pay
Snapshot storage	$0.05/GB/month	Stored snapshots, billed by capacity
DDoS protection	+$10/month	10 Gbps of mitigation
Windows licence	Not included in the base price	Running Windows instead of a Linux image

The structure matters more than any single line. Backups at 20% of the base fee turn the cheapest instance into $3.00/month before anything else is added. Snapshot storage at $0.05/GB/month means 100 GB of snapshots costs $5.00/month — twice the price of the $2.50 instance they are snapshots of. DDoS protection is $10/month for 10 Gbps of mitigation, four times the cheapest instance, for a level a small site may never need. A Windows licence is not in the base price.

All of it is subject to one line in Vultr’s own pricing: pricing varies by data centre location. Build the estimate from the region you will actually deploy into, add 20% if you want backups, then decide whether the tier is affordable.

What Vultr manages for you

These are not the reason to pick Vultr, but they change what you operate. Kubernetes Engine has a free control plane, so you pay only for the worker nodes and their resources. Managed MySQL and PostgreSQL make the database an endpoint rather than a package you patch. Load balancers put several instances behind one address. Block Storage, in NVMe and HDD tiers, attaches capacity separately, and S3-compatible Object Storage keeps files off the instance’s disk.

No prices are quoted for those services here: their rates could not be verified from the pages checked on 25 September 2026.

What it does well

Renting a GPU by the hour changes what is affordable. The A16 is $0.471 per GPU per hour, the A40 $1.712, the GH200 $1.990 and the PCIe A100 $2.397. A few hours of experimentation costs a few dollars, with no card to buy, power or cool — and no discovering after the purchase that the pipeline does not run.

Every plan lists an hourly price beside the monthly one. $0.004/hour and $2.50/month on the entry tier, $0.112/hour and $75.00/month on the storage tier. A migration, a build that runs twice a day or a benchmark campaign can start hourly and commit later.

The $6 tier is where this stops being an experiment, and it costs $3.50 more than entry. One GB of RAM is still small, but it holds a small application server, a queue worker or a static site with a light process behind it, storage moves to 25 GB of NVMe and bandwidth to 2 TB. Rent a GPU for a few hours and this is the machine to put beside it, not the $2.50 tier.

Dedicated vCPU starts at $30, so shared CPU is a choice rather than the only option. The General Purpose tier gives 1 dedicated vCPU, 4 GB of RAM, 30 GB of NVMe and 4 TB of bandwidth — the first configuration worth calling a production baseline, because an application, a small database and a worker can share it without competing for someone else’s CPU time.

The $75 tier makes storage its own product. Storage Optimized is still 1 vCPU but carries 8 GB of RAM and 150 GB of NVMe — five times the General Purpose tier’s storage at the same 4 TB of bandwidth — built for datasets and busy databases where capacity, not compute, is the constraint.

The ceiling is high and priced on the same hourly meter. VX1 starts at 2 vCPU, 8 GB and 5 TB of block storage at $0.060/hour and reaches 192 vCPU, 768 GB and 10 TB at $5.760/hour; Memory Optimized runs from 2 vCPU, 16 GB and 5 TB at $0.080/hour to 192 vCPU, 1536 GB and 10 TB at $9.258/hour. Needing one large machine for a short while does not mean changing vendors.

The platform is scriptable, and the managed surface removes some of the ops. A large public API and documentation surface is what lets cloud resources be wired into an existing pipeline, Kubernetes Engine’s control plane is free, and the managed databases, load balancers and object storage sit in the same account (below).

Who this is not for

The alternatives, stated honestly

The two hosting entries already on this site are both cheaper than Vultr for equivalent specifications.

InterServer’s shared hosting is for a website with no server administration: a control panel, one-click installs and someone else responsible for the stack. If the goal is a site that stays online while you do something else, paying Vultr for a server you then have to run is the more expensive route.

VPS.org is for the cheapest root-access server, and it bills hourly too. Choose it for a small self-hosted app, a staging box or a build runner, when you do not need dedicated vCPU or a GPU. Its verified pricing is on the tool page.

Vultr is the most expensive of the three for equivalent specs, which is why it sits third. It earns the premium only when you need what the others do not sell: hourly GPU access, or a tier with dedicated vCPU.

How it scores

Features is a 5. Five product lines — Cloud Compute, Optimized Cloud Compute, VX1, Memory Optimized and Cloud GPU — plus block storage, object storage, managed MySQL and PostgreSQL, Kubernetes Engine with a free control plane, and load balancers, from $2.50/month to 192 vCPU and 1536 GB of RAM.

Value is a 4. The cheapest entry is competitive and the hourly GPU rates are the reason to be here. The mid tiers and the add-ons are where it stops being a bargain.

Ease of use is a 4. There is a large public API and documentation surface, which is what makes the platform automatable — but it is still a cloud console, not a hosting panel.

Support is a 4. This is scored on the documented automation surface and the managed services. I performed no first-hand support test — no ticket, no chat, no response-time measurement — so nothing here promises how fast anyone answers.

The four scores average 4.25, which is why the rating is 4.2 rather than 4.3 — a platform that covers an unusually wide range, sold at a price that only makes sense when you need the part of the range the cheap hosts do not have.

Common questions

What is the real entry price?

$2.50/month for 1 vCPU with 0.5 GB of RAM, listed with an hourly price of $0.004. It is the cheapest way onto Vultr and not an application server. The first plan with usable RAM is the $6 High Performance tier.

Is the cheapest GPU enough to run a model?

No. The A16 at $0.471 per GPU per hour is positioned for virtual desktops and light graphics. The lines Vultr positions for inference are the GH200 at $1.990 and the PCIe A100 at $2.397 per GPU per hour.

Can I run my app on the $2.50 tier?

0.5 GB of RAM is below what an application server needs — a runtime, a package manager and a reverse proxy exhaust it before your code is deployed. For an app that talks to a model, start at $6 or above.

What does a GPU cost if I leave it running?

It bills per GPU per hour for every hour the instance exists. An 8-GPU A16 instance is $3.767/hour and an 8-GPU PCIe A100 instance is $19.178/hour, whether or not a job is running.

What is not included in the price?

Backups add 20% to the base fee, snapshot storage is $0.05/GB/month, DDoS protection is $10/month for 10 Gbps, and a Windows licence is extra. Pricing also varies by data centre location.

How do I avoid a surprise bill?

Decide the hours before you deploy, because that is what the meter measures. Add 20% if you want backups. Build the estimate from the region you will actually use, and treat the base fee as the floor.

Bottom line

Vultr is worth paying for when you need something the cheaper hosts here do not sell: a GPU by the hour, a tier with dedicated vCPU, or a large specification you can spin up, use and destroy. For a weekend of experimentation, a bounded batch job or a migration, the hourly meter is the right instrument.

It is the wrong buy as a cheap permanent host: 0.5 GB of RAM is below an application server’s needs, the cheapest GPU line is not built for inference, and backups at 20%, snapshots at $0.05/GB/month, DDoS protection at $10/month and a Windows licence all land on top of the base fee.

Buy it for the GPU hours or the dedicated-vCPU tiers, and only when the work is bounded enough that the meter works in your favour. If you want a website you do not administer, buy shared hosting; if you want the cheapest root server, VPS.org is the cheaper entry. Everything here is a snapshot checked on 25 September 2026, so verify the rates before you order.

Check the current Vultr GPU and instance rates on the official pricing page.