Home/GPU Marketplace
GPU MarketplaceRent NVIDIA GPUs by the hour or by the month
GPU servers from verified providers and from Velyrix, in professional data-centre facilities. Launch a virtual machine or a container from the customer portal, pay from prepaid credit metered per second, and see every price component before you launch.
A virtual machine, a container or a whole server
Choose the isolation and the billing that suit your workload. All three run on the same servers and are ordered self-service in the customer portal.
GPU virtual machine
A KVM virtual machine with the GPUs passed through with VFIO. Your own kernel, full root access and the strongest isolation from other tenants. Billed by the hour.
- Ubuntu 22.04 and 24.04 LTS images
- Images with the NVIDIA driver and CUDA preinstalled
- SSH and a web console
GPU container
A Docker container with the NVIDIA Container Toolkit, started from a ready template or, where enabled, from your own image from an approved registry. Billed by the hour.
- CUDA, PyTorch, TensorFlow and Jupyter templates
- SSH and Jupyter through mapped ports
- Encrypted /workspace volume for your data
Whole server
One virtual machine that owns every GPU and nearly all the CPU and memory of the server. Ordered by the month and paid in advance, with an egress allowance included.
- All GPUs of the server
- Paid monthly in advance
- Egress allowance included
Data-centre GPUs in professional facilities
Every server on the marketplace is approved before it is listed. Servers are operated by independent providers that have passed verification, and by Velyrix itself. Velyrix is the seller of record: you contract with and pay Velyrix.
- Data-centre NVIDIA GPUs only: B300, B200, H200, H100, A100, L40S, L4, RTX PRO 6000 Blackwell and RTX A6000, depending on what is listed at the time.
- No GeForce: GeForce cards are never accepted, because NVIDIA's driver licence does not allow them in data centres.
- Professional facilities: servers must be in a professional data-centre facility, in an export-cleared country from the same list as the Velyrix markets.
- Everything on the listing: GPU model and count, GPU memory, vCPU, RAM, disk, bandwidth, region, benchmark results, reliability score and price.
The share of five-minute intervals over the last 30 days in which the server was online. Time inside an announced maintenance window does not count against it. You can sort and filter by it.
Marketplace instances run on single servers. For multi-node clusters with InfiniBand fabric, a signed acceptance report and contracted service levels, see dedicated GPU servers.
Which GPU models, regions and quantities are available changes as servers are listed and rented. The customer portal shows what you can launch right now.
Per-component pricing, published up front
You pay for each component separately: GPU-hours, vCPU-hours, RAM GB-hours, disk GB-months and egress GB. Ingress is free. The actual price is set per server and shown in the customer portal before you launch, and the hourly rates of an instance are locked at launch for its lifetime.
| GPU model | GPU memory | Per GPU-hour | Whole server — per GPU / month |
|---|---|---|---|
| B300 SXM | 288 GB | from $10.88 | from $7,300 |
| B200 SXM | 180 GB | from $8.63 | from $6,200 |
| H200 | 141 GB | from $4.95 | from $3,900 |
| H100 NVL / PCIe | 94 GB | from $3.53 | from $2,400 |
| H100 SXM | 80 GB | from $4.05 | from $3,000 |
| A100 80GB | 80 GB | from $2.48 | from $1,700 |
| A100 40GB | 40 GB | from $1.91 | from $1,300 |
| RTX PRO 6000 Blackwell | 96 GB | from $2.85 | from $1,500 |
| RTX A6000 | 48 GB | from $0.90 | from $620 |
| L40S | 48 GB | from $1.76 | from $985 |
| L4 | 24 GB | from $0.50 | from $355 |
Starting prices in USD, exclusive of tax: the platform's default rate for each GPU model. Individual servers may be priced differently. The price shown on each server in the customer portal is the one that applies, and an instance keeps the rates it was launched with. A whole server costs the per-GPU price multiplied by its number of GPUs.
Other components
| Component | Unit | Price |
|---|---|---|
| vCPU | per vCPU-hour | $0.03 |
| Memory (RAM) | per GB-hour | $0.006 |
| Disk — system and data | per GB-month | $0.15 |
| Egress traffic | per GB | $0.03 |
| Ingress traffic | per GB | Free |
| Dedicated public IPv4 | per IP-hour | $0.0075 |
| Dedicated public IPv4 | per IP / month | $5.40 |
| Whole server — egress allowance | per GPU / month | 2 TB included |
| Example: 1× H100 SXM, 16 vCPU, 128 GB RAM, 200 GB disk | per hour, plus egress | $5.34 |
Disk is charged per GB-month, accrued hourly, for as long as the instance exists, including while it is stopped. Egress above a whole server's allowance is charged at the egress rate. The example uses the starting rates above.
Prepaid credit, metered per second
Hourly instances run on prepaid account credit. Top up in the customer portal; to launch, your credit must cover at least the first hours of the instance, and the portal shows the amount needed.
Usage is measured per second and deducted from your credit every minute. The portal shows usage per component, and you receive a monthly usage statement.
We e-mail you when your remaining credit is expected to last about 24 hours, and again at about 2 hours, based on what your running instances use.
When your available credit reaches zero, running hourly instances are stopped. Disk continues to be charged while a stopped instance exists.
An instance stopped for lack of credit or non-payment is kept for 7 days. If you do not top up or pay in that time, it is deleted and its data is crypto-erased. Deleted data cannot be recovered.
Ordered by the month, invoiced in advance and provisioned when payment is received. Renewals are invoiced in advance; an unpaid renewal stops the server and starts the 7-day retention period.
Hourly instances are charged only for the usage the server reports, so nothing is charged for the time a server is offline. Whole-server rentals receive reliability-based service credits under the GPU Marketplace Terms.
Encrypted by default, reachable two ways
Encrypted disks and volumes
Virtual machine disks and container data volumes are encrypted with a separate key for each instance. Keys are held by the platform, delivered only to the server running the instance and never stored on its disks. Deleting an instance destroys its key.
Isolation
Virtual machines have their own kernel and GPUs assigned through VFIO. Containers run unprivileged with resource limits. Instances cannot reach the provider's network, other tenants or the host's services.
Direct or through the relay
Connect over SSH or mapped ports on the server's public IP address, or through the Velyrix relay gateway. A dedicated public IPv4 address is available as an option.
Container image layers and anything stored outside the /workspace volume are not encrypted, so keep your data in /workspace.
Four steps to your first instance
Create an account
Sign up in the customer portal and verify your e-mail address.
Accept the marketplace terms
Accept the GPU Marketplace Terms once, in the marketplace section of the portal. They cover capacity hosted by third parties, per-component billing, prepaid credit, retention and deletion, encryption and export restrictions.
Top up credit
Add prepaid credit to your account. Whole servers are paid by invoice instead.
Launch
Choose a server, pick a virtual machine image or a container template, add your SSH key, review the price breakdown and launch.
Find a GPU and launch it
Browse servers, prices and reliability scores in the customer portal. You need an account and prepaid credit to launch. Do you own data-centre GPU servers? You can list them too.
Frequently asked questions
How is the GPU Marketplace different from dedicated GPU servers?
The marketplace rents single servers by the hour or by the month, self-service, from prepaid credit, on capacity operated by verified providers or by Velyrix. Dedicated GPU servers are single-tenant Velyrix nodes on monthly or longer terms, with InfiniBand fabric options, a signed acceptance report and contracted service levels, priced in a written quotation.
Which GPUs can I rent?
Data-centre NVIDIA GPUs such as B300, B200, H200, H100, A100, L40S, L4, RTX PRO 6000 Blackwell and RTX A6000, depending on what is listed at the time. GeForce cards are never accepted. The customer portal shows what is available now.
How is usage billed?
Per component: GPU-hours, vCPU-hours, RAM GB-hours, disk GB-months and egress GB, plus a dedicated IPv4 address if you add one. Ingress is free. Usage is metered per second and deducted from your prepaid credit every minute.
Can the price change while my instance is running?
No. The hourly rates of an instance are locked when you launch it and stay the same for its lifetime. Price changes apply only to new launches.
What happens when my credit runs out?
We e-mail you at about 24 hours and about 2 hours of remaining credit. At zero, running hourly instances are stopped and disk continues to be charged. If you do not top up within 7 days, the stopped instances are deleted and their data is crypto-erased.
Is my data encrypted?
Yes. Virtual machine disks and container data volumes are encrypted with a separate key for each instance. The key is held by the platform and never stored on the provider's server at rest, and deleting the instance destroys it. Container image layers and data outside the /workspace volume are not encrypted.
How do I connect to an instance?
Over SSH or the mapped ports on the server's public IP address, or through the Velyrix relay gateway. Virtual machines also have a web console in the customer portal, and container templates with Jupyter expose it through a mapped port.