Opening checkout

Halloween deal 80% off your first month

Ends in

GPU Dedicated & Virtual Servers for AI and ML Workloads

Train models, fine-tune LLMs, run inference and render on NVIDIA GPUs. Take the whole machine as a dedicated server, or a virtual server at a lower starting estimate.

See GPU pricing
  • 6-month estimates from $290/mo
  • 1 to 8 GPUs per server
  • 6-month minimum term
Memory per GPU1 to 8 GPUs per server
  1. RTX 409024 GB GDDR6X
  2. RTX 509032 GB GDDR7
  3. RTX PRO 6000 Blackwell96 GB GDDR7 ECC
  4. L40S48 GB GDDR6 ECC
  5. A10080 GB HBM2e
  6. H10080 or 94 GB HBM3
  7. H200141 GB HBM3e
  8. B200180 GB HBM3e

Choose your GPU

NVIDIA GPU server pricing. Choose your GPU.

Each price is an estimate for one GPU with its included host, on the 6-month minimum term. No setup fee, and longer terms cost less: 10% off for 12 months and 15% off for 24 months.
  • Ada Lovelace
  • GeForce RTX

NVIDIA RTX 4090

Best for inference, image generation and rendering.

A GeForce card with 24 GB GDDR6X and FP8 Tensor Cores. It holds a 7B model at FP16 (14 GB) or a 14B model at 8-bit (14 GB), with room for context.

The GPU

GPU memory
24 GB GDDR6X
Up to 192 GB in an 8-GPU server
Memory bandwidth
1,008 GB/s
Per GPU
Tensor Cores
FP8, FP16, BF16
4th gen, 16,384 CUDA cores

Included with each GPU

CPU
16 cores
8 vCPU on virtual
RAM
64 GB
DDR5 ECC, up to 2 TB
NVMe storage
1 TB
Up to 32 TB
Network
1 Gbps
Unmetered, up to 25 Gbps

The CPU is AMD EPYC or AMD Ryzen, your choice on request.

Dedicated GPU server

From$410/mo

1 GPU with its included host, on the 6-month term.

Starting estimates a month for the NVIDIA RTX 4090, by term
TermDedicatedVirtual
6 months$410/mo$290/mo
12 months10% off$360/mo$260/mo
24 months15% off$340/mo$240/mo
  • No setup fee
  • 6-month minimum term

Estimates. Our team confirms the final price and availability in your quote.

  • Blackwell
  • GeForce RTX

NVIDIA RTX 5090

Best for inference, image and video generation, and rendering.

Blackwell brings FP4 Tensor Cores and 32 GB GDDR7. One card holds a 14B model at 8-bit (14 GB) or a 32B model at 4-bit (16 GB).

The GPU

GPU memory
32 GB GDDR7
Up to 256 GB in an 8-GPU server
Memory bandwidth
1,792 GB/s
Per GPU
Tensor Cores
FP4, FP8, FP16
5th gen, 21,760 CUDA cores

Included with each GPU

CPU
16 cores
8 vCPU on virtual
RAM
64 GB
DDR5 ECC, up to 2 TB
NVMe storage
1 TB
Up to 32 TB
Network
1 Gbps
Unmetered, up to 25 Gbps

The CPU is AMD EPYC or AMD Ryzen, your choice on request.

Dedicated GPU server

From$540/mo

1 GPU with its included host, on the 6-month term.

Starting estimates a month for the NVIDIA RTX 5090, by term
TermDedicatedVirtual
6 months$540/mo$380/mo
12 months10% off$490/mo$340/mo
24 months15% off$460/mo$320/mo
  • No setup fee
  • 6-month minimum term

Estimates. Our team confirms the final price and availability in your quote.

  • Blackwell
  • RTX PRO

NVIDIA RTX PRO 6000 Blackwell

Best for large-model inference, fine-tuning and rendering.

96 GB of ECC memory on one Blackwell card, with RT cores for rendering. It holds a 70B model at 8-bit (70 GB) and leaves room for context.

The GPU

GPU memory
96 GB GDDR7 ECC
Up to 768 GB in an 8-GPU server
Memory bandwidth
1,597 GB/s
Per GPU
Tensor Cores
FP4, FP8, FP16
5th gen, 24,064 CUDA cores

Included with each GPU

CPU
16 cores
16 vCPU on virtual
RAM
128 GB
DDR5 ECC, up to 2 TB
NVMe storage
2 TB
Up to 32 TB
Network
1 Gbps
Unmetered, up to 25 Gbps

The CPU is AMD EPYC or AMD Ryzen, your choice on request.

Dedicated GPU server

From$1,300/mo

1 GPU with its included host, on the 6-month term.

Starting estimates a month for the NVIDIA RTX PRO 6000 Blackwell, by term
TermDedicatedVirtual
6 months$1,300/mo$900/mo
12 months10% off$1,150/mo$810/mo
24 months15% off$1,100/mo$770/mo
  • No setup fee
  • 6-month minimum term

Estimates. Our team confirms the final price and availability in your quote.

  • Ada Lovelace
  • Data center GPU

NVIDIA L40S

Best for inference, fine-tuning, rendering and video.

A data center card with 48 GB of ECC memory, RT cores and video encoders. It holds a 70B model at 4-bit (35 GB) or a 14B model at FP16 (28 GB).

The GPU

GPU memory
48 GB GDDR6 ECC
Up to 384 GB in an 8-GPU server
Memory bandwidth
864 GB/s
Per GPU
Tensor Cores
FP8, FP16, BF16
4th gen, 18,176 CUDA cores

Included with each GPU

CPU
16 cores
16 vCPU on virtual
RAM
128 GB
DDR5 ECC, up to 2 TB
NVMe storage
2 TB
Up to 32 TB
Network
1 Gbps
Unmetered, up to 25 Gbps

The CPU is AMD EPYC or AMD Ryzen, your choice on request.

Dedicated GPU server

From$770/mo

1 GPU with its included host, on the 6-month term.

Starting estimates a month for the NVIDIA L40S, by term
TermDedicatedVirtual
6 months$770/mo$540/mo
12 months10% off$690/mo$490/mo
24 months15% off$650/mo$460/mo
  • No setup fee
  • 6-month minimum term

Estimates. Our team confirms the final price and availability in your quote.

  • Ampere
  • Data center GPU

NVIDIA A100

Best for training, fine-tuning and HPC.

80 GB HBM2e and FP64 Tensor Cores for HPC. It has no FP8, so it trains and fine-tunes in BF16 or TF32. MIG splits one GPU into up to 7 instances.

The GPU

GPU memory
80 GB HBM2e
Up to 640 GB in an 8-GPU server
Memory bandwidth
1,935 GB/s
Per GPU
Tensor Cores
FP16, BF16, TF32
3rd gen, 6,912 CUDA cores

Included with each GPU

CPU
16 cores
16 vCPU on virtual
RAM
128 GB
DDR5 ECC, up to 2 TB
NVMe storage
2 TB
Up to 32 TB
Network
1 Gbps
Unmetered, up to 25 Gbps

The CPU is AMD EPYC or AMD Ryzen, your choice on request.

Dedicated GPU server

From$1,200/mo

1 GPU with its included host, on the 6-month term.

Starting estimates a month for the NVIDIA A100, by term
TermDedicatedVirtual
6 months$1,200/mo$860/mo
12 months10% off$1,100/mo$770/mo
24 months15% off$1,050/mo$730/mo
  • No setup fee
  • 6-month minimum term

Estimates. Our team confirms the final price and availability in your quote.

  • Hopper
  • Data center GPU

NVIDIA H100

Best for large-scale training and LLM inference.

Hopper's Transformer Engine trains and serves models in FP8. Builds of 1 or 2 GPUs get the H100 NVL (94 GB). Builds of 4 or 8 get the H100 SXM (80 GB), so an 8-GPU server has 640 GB of GPU memory.

The GPU, by build

NVIDIA H100 NVL

1 or 2 GPUs

PCIe cards, an NVLink bridge on each pair

GPU memory
94 GB HBM3
Bandwidth
3.9 TB/s
NVLink
600 GB/s bridge
Per server
Up to 188 GB
NVIDIA H100 SXM

4 or 8 GPUs

SXM boards on an HGX baseboard

GPU memory
80 GB HBM3
Bandwidth
3.35 TB/s
NVLink
900 GB/s
Per server
Up to 640 GB
Tensor Cores, both cards
FP8, FP16, BF16
4th gen, Transformer Engine

Included with each GPU

CPU
16 cores
16 vCPU on virtual
RAM
256 GB
DDR5 ECC, up to 2 TB
NVMe storage
2 TB
Up to 32 TB
Network
1 Gbps
Unmetered, up to 25 Gbps

The CPU is AMD EPYC or AMD Ryzen, your choice on request.

Dedicated GPU server

From$1,750/mo

1 GPU with its included host, on the 6-month term.

Starting estimates a month for the NVIDIA H100, by term
TermDedicatedVirtual
6 months$1,750/mo$1,200/mo
12 months10% off$1,600/mo$1,100/mo
24 months15% off$1,500/mo$1,050/mo
  • No setup fee
  • 6-month minimum term

Estimates. Our team confirms the final price and availability in your quote.

  • Hopper
  • Data center GPU

NVIDIA H200

Best for the largest models and long-context inference.

The Hopper GPU of the H100 with 141 GB HBM3e at 4.8 TB/s, on the NVL card and the SXM board alike. One GPU holds a 70B model at 8-bit (70 GB) with room for long context.

The GPU, by build

NVIDIA H200 NVL

1 or 2 GPUs

PCIe cards, an NVLink bridge on each pair

GPU memory
141 GB HBM3e
Bandwidth
4.8 TB/s
NVLink
900 GB/s bridge
Per server
Up to 282 GB
NVIDIA H200 SXM

4 or 8 GPUs

SXM boards on an HGX baseboard

GPU memory
141 GB HBM3e
Bandwidth
4.8 TB/s
NVLink
900 GB/s
Per server
Up to 1,128 GB
Tensor Cores, both cards
FP8, FP16, BF16
4th gen, Transformer Engine

Included with each GPU

CPU
16 cores
16 vCPU on virtual
RAM
256 GB
DDR5 ECC, up to 2 TB
NVMe storage
2 TB
Up to 32 TB
Network
1 Gbps
Unmetered, up to 25 Gbps

The CPU is AMD EPYC or AMD Ryzen, your choice on request.

Dedicated GPU server

From$2,400/mo

1 GPU with its included host, on the 6-month term.

Starting estimates a month for the NVIDIA H200, by term
TermDedicatedVirtual
6 months$2,400/mo$1,650/mo
12 months10% off$2,150/mo$1,500/mo
24 months15% off$2,050/mo$1,400/mo
  • No setup fee
  • 6-month minimum term

Estimates. Our team confirms the final price and availability in your quote.

  • Blackwell
  • Data center GPU
  • 8-GPU servers only

NVIDIA B200

Best for frontier-scale training and inference.

A Blackwell data center GPU with 180 GB HBM3e and FP4 Tensor Cores. It comes in 8-GPU servers, with 1,440 GB of GPU memory in one server.

The GPU

GPU memory
180 GB HBM3e
1,440 GB in one 8-GPU server
Memory bandwidth
About 8 TB/s
Per GPU
Tensor Cores
FP4, FP8, FP16
2nd-gen Transformer Engine

Included with each GPU

CPU
16 cores
16 vCPU on virtual
RAM
256 GB
DDR5 ECC, up to 2 TB
NVMe storage
2 TB
Up to 32 TB
Network
1 Gbps
Unmetered, up to 25 Gbps

The CPU is AMD EPYC or AMD Ryzen, your choice on request.

Dedicated GPU server

On quote

Our team prices every B200 build in your quote.

  • No setup fee
  • 10% off for 12 months and 15% off for 24 months
  • 6-month minimum term

Our team confirms the final price and availability in your quote.

Custom or multi-server builds

Another card, several servers or a build outside the form. Tell us what you need, and our team prices it.

Available configurations

Available configurations. 1 to 8 GPUs per server.

Each build includes a standard host for every GPU. The prices are the request form's monthly estimates on the 6-month term. Choose a build to open the form on it.

NVIDIA RTX 4090

24 GB GDDR6X per GPU. Included per GPU: 16 cores, 64 GB RAM, 1 TB NVMe.

NVIDIA RTX 5090

32 GB GDDR7 per GPU. Included per GPU: 16 cores, 64 GB RAM, 1 TB NVMe.

NVIDIA RTX PRO 6000 Blackwell

96 GB GDDR7 ECC per GPU. Included per GPU: 16 cores, 128 GB RAM, 2 TB NVMe.

NVIDIA L40S

48 GB GDDR6 ECC per GPU. Included per GPU: 16 cores, 128 GB RAM, 2 TB NVMe.

NVIDIA A100

80 GB HBM2e per GPU. Included per GPU: 16 cores, 128 GB RAM, 2 TB NVMe.

NVIDIA H100

80 or 94 GB HBM3 per GPU. Included per GPU: 16 cores, 256 GB RAM, 2 TB NVMe.

NVIDIA H200

141 GB HBM3e per GPU. Included per GPU: 16 cores, 256 GB RAM, 2 TB NVMe.

NVIDIA B200

180 GB HBM3e per GPU. 8-GPU servers only. Included per GPU: 16 cores, 256 GB RAM, 2 TB NVMe.

Locations

US
Chicago, New York
Europe
London, Dublin, Amsterdam, Frankfurt
Asia
Tokyo, Singapore

Choose a preferred city in the request form. Our team confirms where your build can go.

Prices and terms

Monthly estimates on the 6-month term. 10% off for 12 months and 15% off for 24 months. No setup fee.

The host shown is for a dedicated server; a virtual server has vCPU in place of cores. Our team confirms the final price and availability in your quote.

Use cases

GPU servers for AI and rendering. Which GPU fits your workload.

Training, reinforcement learning, inference and fine-tuning each lean on a different part of the GPU. Find your work, see what matters for it, and open its GPUs.

LLM training

Pre-train a model or continue its training on new data. GPUs exchange data at every step, so training needs the most compute, memory bandwidth and fast links between GPUs.

What matters

  • Memory bandwidth
  • FP8 Tensor Cores
  • NVLink

GPUs that fit

Large runs use 8-GPU servers with NVLink. Cards without it link over PCIe and suit smaller models.

Opens the request form on 8 × H100 SXM.

Reinforcement learning

RLHF and other RL post-training keep several models in memory at once: PPO holds a policy, a critic, a reward model and a reference model. Robotics RL trains in simulation, such as NVIDIA Isaac Sim.

What matters

  • GPU memory
  • RT cores
  • CPU cores and NVMe

GPUs that fit

Isaac Sim needs RT cores, so robotics simulation runs on the RTX PRO 6000, L40S or RTX cards. The A100, H100, H200 and B200 have none.

Opens the request form on the H200 NVL.

Inference and serving

Serve a model to your users or your own apps. GPU memory holds the weights and the context of every live request, and memory bandwidth sets the tokens per second.

What matters

  • GPU memory
  • Memory bandwidth
  • FP4 Tensor Cores
  • MIG

GPUs that fit

MIG splits one A100, H100, H200 or B200 into up to 7 instances, each for a small model.

Opens the request form on the RTX 5090.

Fine-tuning: full and LoRA

Adapt an open model to your data. LoRA trains small adapters on a frozen model, so one GPU is often enough. Full fine-tuning with Adam needs about 16 bytes of GPU memory per parameter.

What matters

  • GPU memory
  • FP8 Tensor Cores
  • NVLink

GPUs that fit

QLoRA keeps the base model in 4-bit: its authors fine-tuned a 65B model on one 48 GB GPU.

Opens the request form on the RTX PRO 6000.

Image and video generation

Generate images and video with diffusion models. FLUX.1 [dev] has 12 billion parameters, about 24 GB in BF16. Long, high-resolution video clips need more memory.

What matters

  • GPU memory
  • FP4 Tensor Cores
  • Video engines

GPUs that fit

Opens the request form on the RTX 5090.

3D rendering

Render scenes and animation with GPU renderers. Each GPU holds its own copy of the scene, so the memory on one card sets the largest scene.

What matters

  • RT cores
  • GPU memory
  • ECC memory

GPUs that fit

The A100, H100, H200 and B200 have no RT cores, so they are not rendering GPUs.

Opens the request form on the RTX PRO 6000.

Data science and HPC

Run simulations, GPU dataframes and classic machine learning. Double-precision (FP64) work needs the HBM GPUs, and large datasets need memory and bandwidth.

What matters

  • FP64
  • Memory bandwidth
  • ECC memory

GPUs that fit

Opens the request form on the A100.

Computer vision

Train and run detection, segmentation and video analytics models. Most fit 24 to 48 GB cards, and more memory allows larger batches and images.

What matters

  • GPU memory
  • Video engines
  • CPU cores and NVMe

GPUs that fit

Loading and decoding images often limits training speed, so CPU cores and NVMe matter too.

Opens the request form on the L40S.

How much GPU memory does your model need?

Its weights take the number of parameters times the bytes per parameter: 2 at FP16, 1 at 8-bit and 0.5 at 4-bit. Leave room on top for the context (the KV cache) and the runtime. Training needs far more: about 16 bytes per parameter with the Adam optimizer.

Compare the GPUs side by side
Model weights in GPU memory, by model size and precision
Model sizeFP168-bit4-bit
7B parameters14 GB7 GB3.5 GB
13B parameters26 GB13 GB6.5 GB
34B parameters68 GB34 GB17 GB
70B parameters140 GB70 GB35 GB

AI agents and coding tools that call a hosted model need no GPU of their own. Run them on an AI agent VPS.

Hardware

GPU server hardware. NVIDIA’s figures, side by side.

Compare the GPUs by the parts your work leans on: memory and bandwidth, Tensor Core formats, NVLink and MIG. Choose a GPU to open its card and its prices.
NVIDIA GPU specifications: memory, compute, links, power and form factor of each GPU in the request form, with the H100 and H200 NVL and SXM cards apart
SpecificationNVIDIA RTX 4090Ada LovelaceNVIDIA RTX 5090BlackwellNVIDIA RTX PRO 6000BlackwellNVIDIA L40SAda LovelaceNVIDIA A100AmpereNVIDIA H100HopperNVIDIA H200HopperNVIDIA B200Blackwell
H100 NVL1 or 2 GPUsH100 SXM4 or 8 GPUsH200 NVL1 or 2 GPUsH200 SXM4 or 8 GPUs
Memory
GPU memory24 GBGDDR6X32 GBGDDR796 GBGDDR7 ECC48 GBGDDR6 ECC80 GBHBM2e94 GBHBM380 GBHBM3141 GBHBM3e141 GBHBM3e180 GBHBM3e
Memory bandwidthPer GPU1,008 GB/s1,792 GB/s1,597 GB/s864 GB/s1,935 GB/s3.9 TB/s3.35 TB/s4.8 TB/s4.8 TB/sAbout 8 TB/s
Compute
CUDA cores16,38421,76024,06418,1766,912Not listed16,896Not listedNot listedNot listed
Tensor CoresGeneration4th gen5th gen5th gen4th gen3rd gen4th gen4th gen4th gen4th gen2nd-gen Transformer Engine
Tensor formatsFP8, FP16, BF16FP4, FP8, FP16FP4, FP8, FP16FP8, FP16, BF16FP16, BF16, TF32FP8, FP16, BF16FP8, FP16, BF16FP8, FP16, BF16FP8, FP16, BF16FP4, FP8, FP16
FP16 Tensor, denseTFLOPS per GPU165.2209.5Not listed3623128359898359892,250
FP8 Tensor, denseTFLOPS per GPU330.3419Not listed733No FP81,6701,9791,6701,9794,500
RT cores128170188142NoneNoneNoneNoneNoneNone
Links, power and form
NVLinkGB/s per GPUNoNoNoNo600Bridge600Bridge900900Bridge9001,800
MIGInstances per GPUNoNoUp to 4NoUp to 7Up to 7Up to 7Up to 7Up to 7Up to 7
Max power450 W575 W600 W350 W300 W400 W700 W600 W700 W1,000 W
Form factorPCIe Gen 4 cardPCIe Gen 5 cardPCIe Gen 5 card, passivePCIe Gen 4 card, passivePCIe Gen 4 cardPCIe Gen 5 card, dual-slotSXM5 on an HGX boardPCIe card, dual-slotSXM on an HGX boardSXM on an 8-GPU HGX board
  • Tensor figures are dense, without sparsity. NVIDIA prints most of them with sparsity, at twice these rates.
  • RTX 4090 and RTX 5090: FP16 and FP8 with FP32 accumulate, the mode most training frameworks use.
  • RTX PRO 6000: NVIDIA lists peaks of 1 PFLOPS FP16 and 2 PFLOPS FP8 and does not say whether they use sparsity, so the table shows no figure.
  • H100 and H200: a build of 1 or 2 GPUs gets the NVL card, a PCIe card with an NVLink bridge on each pair. A build of 4 or 8 GPUs gets SXM boards on an HGX baseboard, with NVLink between all GPUs.
  • A100: the NVLink Bridge links the GPUs in pairs. B200: per GPU of an 8-GPU HGX B200 board, and NVIDIA marks its MIG figure as preliminary.
  • Source: NVIDIA product pages, datasheets and architecture papers, accessed October 2026.

The host: CPU, RAM, NVMe and network.

Each GPU price includes a standard host for that GPU. The host runs AMD EPYC or AMD Ryzen, your choice on request, with DDR5 ECC RAM. More CPU, RAM, storage or port speed costs extra. Choose a preferred city from our locations. For the processors behind our trading VPS plans, see the hardware page.
GPUs per server
1, 2, 4 or 8
The prices on the GPU cards are for 1 GPU. The B200 comes in 8-GPU servers only.
CPU, dedicated
16 to 192 cores
Included: 16 cores per GPU
CPU, virtual
8 to 192 vCPU
Included: 8 vCPU per GPU on the RTX 4090 and RTX 5090, 16 vCPU on the others
CPU family
AMD EPYC, AMD Ryzen or no preference
AMD Ryzen hosts take 1 or 2 RTX 4090, RTX 5090, RTX PRO 6000 or L40S GPUs, with up to 16 cores or 32 vCPU. Our team confirms the CPU model in your quote
RAM
64 GB to 2 TB of DDR5 ECC
Included: 64 GB per GPU on the RTX 4090 and RTX 5090, 128 GB on the RTX PRO 6000 Blackwell, L40S and A100, 256 GB on the H100, H200 and B200
NVMe storage
1 to 32 TB
Included: 1 TB per GPU on the RTX 4090 and RTX 5090, 2 TB on the others
Network port
1 Gbps unmetered, 10 Gbps with 100 TB a month, 10 Gbps unmetered or 25 Gbps unmetered
Included: 1 Gbps unmetered. The 25 Gbps port is priced in your quote.
Location
No preference, or one of our eight cities
A preference: our team confirms where your build can go
Operating system
Ubuntu 24.04, Ubuntu 22.04, Windows Server 2025, Windows Server 2022
Choose one in the request form
Term
6, 12 or 24 months
6-month minimum. 10% off for 12 months and 15% off for 24 months. No setup fee

Uptime and facilities.

Our SLA and our datacenter standards apply to GPU servers, as to every QuantVPS server.

Uptime SLA

99.999%

A monthly uptime target. If we miss it, you are eligible for a service credit.

Read the SLA
Datacenters
Tier IV
Eight Tier IV datacenters, from Chicago to Tokyo.
Power
N+1
Fault-tolerant N+1 utility and UPS feeds, with diesel generators and 48+ hours of fuel on site.
Cooling
N+1
Concurrently maintainable N+1 cooling holds rack temperatures through equipment failures.
Compliance
SOC 2 Type II
SOC 2 Type II facilities with biometric access, mantrap entry, and 24/7 security staff.

Dedicated or virtual

Dedicated GPU server or GPU VPS. The same GPUs, two ways to run them.

A dedicated GPU server is a whole physical machine, used only by you. A GPU VPS is a virtual GPU server: a virtual machine with each GPU you choose passed through to it in full. You choose either one in the same request form.

Need a dedicated server without a GPU? See dedicated trading servers.

Dedicated GPU server

The server
A whole physical server, used only by you
The GPUs
Every GPU in the server, used only by you
Operating system
Runs on the hardware, with no hypervisor
CPU
Physical cores: 16 per GPU, up to 192
Starting estimate
From $410/mo, RTX 4090
Choose it for
Heavy, steady workloads that use the whole machine

Virtual GPU server

The server
A virtual machine with the GPUs you choose
The GPUs
Each GPU passed through to your virtual machine in full
Operating system
Runs in a virtual machine on a hypervisor
CPU
vCPU: 8 per GPU on the RTX 4090 and RTX 5090, 16 on the others, up to 192
Starting estimate
From $290/mo, RTX 4090
Choose it for
The same GPU at a lower starting estimate

On both: The same 8 GPUs, 1 to 8 per server, the same operating systems and locations, and the 6-month minimum term.

How a build request works

How a build request works. From your build to a quote.

There is no instant deploy for GPU servers. To rent one, you send the build, and our team replies with a quote.

Send your build

Choose the GPU, the server type and the parts in the request form. It shows an estimated price as you choose.

We review it

Our team checks availability and the price for your build.

You get a quote

Your quote confirms the final price and availability. It comes by email, to the address you give.

Build and handover

When you accept the quote, we build your server and hand it over to you.

Questions, answered

GPU server FAQ.

A GPU dedicated server is a whole physical server with one or more NVIDIA GPUs, used only by you. Your operating system runs on the hardware, with no hypervisor, so every GPU, its memory and the server’s CPU, RAM and storage are yours.

A dedicated GPU server is a whole physical server, and your operating system runs on the hardware. A GPU VPS (a virtual GPU server) is a virtual machine: each GPU you choose is passed through to it in full, so you do not share it. A GPU VPS has a lower starting estimate for the same GPU: the RTX 5090 from $380 a month, against $540 a month dedicated.

For large models, the H100, H200 or B200. They have HBM memory, FP8 Tensor Cores and NVLink between GPUs (900 GB/s per GPU in 4 or 8-GPU H100 and H200 servers, 1.8 TB/s on the B200), which multi-GPU training needs. The A100 has no FP8, and its starting estimate is lower. For small models and experiments, an RTX PRO 6000 (96 GB) or an L40S (48 GB) works well.

Start from the model’s memory. A model of up to 14B (14 GB at 8-bit) runs on an RTX 4090 (24 GB) or an RTX 5090 (32 GB). A 70B model needs about 70 GB at 8-bit or 35 GB at 4-bit, plus room for its context: an RTX PRO 6000 (96 GB) or an H200 (141 GB) holds it on one GPU at 8-bit. Memory bandwidth sets the tokens per second, so the H200 (4.8 TB/s) and the B200 serve the most users.

About 140 GB at FP16, 70 GB at 8-bit and 35 GB at 4-bit, for the weights alone: the parameters times the bytes per parameter. The context (the KV cache) adds memory for each token, about 2.5 GB for one 8,192-token request on Llama 3.1 70B at FP16. This is general arithmetic, so measure your own model.

Yes, with LoRA or QLoRA. LoRA trains small adapters on a frozen model, so a 7B or 8B model fits one 24 to 48 GB GPU. QLoRA keeps the base model in 4-bit: its authors fine-tuned a 65B model on one 48 GB GPU. Full fine-tuning with Adam needs about 16 bytes per parameter, so a 7B model needs more than one 80 GB GPU.

For RLHF and other RL post-training of an LLM, choose GPUs with large memory, such as the H100, H200 or B200: PPO keeps a policy, a critic, a reward model and a reference model in memory. For robotics simulation with NVIDIA Isaac Sim, choose a GPU with RT cores: the RTX 4090, RTX 5090, RTX PRO 6000 or L40S. Isaac Sim does not support GPUs without RT cores, such as the A100 and H100.

Each GPU on this page has a starting estimate for one GPU with its included host, on the 6-month term: from $290 a month for a virtual RTX 4090. The request form shows a price range for your build as you choose. Every price is an estimate: our team confirms the final price and availability in your quote.

6 months. You can also choose a 12- or 24-month term in the request form. Longer terms cost less a month than the 6-month price: 10% off for 12 months and 15% off for 24 months. The estimate shows the total for the term you choose.

No. A GPU server has no setup fee. Your quote shows the monthly price for your build and term.

A standard host. CPU: 16 cores per GPU on a dedicated server; on a virtual server, 8 vCPU per GPU on the RTX 4090 and RTX 5090, 16 vCPU on the others. The CPU is AMD EPYC or AMD Ryzen, your choice on request: choose one in the request form, or no preference. AMD Ryzen hosts take 1 or 2 RTX 4090, RTX 5090, RTX PRO 6000 or L40S GPUs, with up to 16 cores or 32 vCPU. RAM: DDR5 ECC, 64 GB per GPU on the RTX 4090 and RTX 5090, 128 GB on the RTX PRO 6000 Blackwell, L40S and A100, 256 GB on the H100, H200 and B200. NVMe storage: 1 TB per GPU on the RTX 4090 and RTX 5090, 2 TB on the others. Network: a 1 Gbps unmetered port. You can add more in the request form, or choose 10 Gbps with 100 TB a month, 10 Gbps unmetered or 25 Gbps unmetered. The 25 Gbps port is priced in your quote.

1, 2, 4 or 8. The B200 comes in 8-GPU servers only. The A100, H100, H200 and B200 support NVLink between GPUs. The A100 and the H100 and H200 NVL cards link in pairs, and 4 or 8-GPU H100 and H200 servers link all their GPUs. The RTX 4090, RTX 5090, RTX PRO 6000 and L40S connect over PCIe only. For another card or a larger build, describe it in the request form.

It depends on the GPU count. Builds of 1 or 2 GPUs get the NVL card: a PCIe card, with an NVLink bridge on each pair. The H100 NVL has 94 GB of HBM3 at 3.9 TB/s, and the H200 NVL 141 GB of HBM3e at 4.8 TB/s. Builds of 4 or 8 GPUs get SXM boards on an HGX baseboard, with NVLink at 900 GB/s per GPU. The H100 SXM has 80 GB of HBM3 at 3.35 TB/s, and the H200 SXM 141 GB of HBM3e at 4.8 TB/s.

Ubuntu 24.04, Ubuntu 22.04, Windows Server 2025 or Windows Server 2022. Choose one in the request form. You get administrator access on Windows or root access on Linux, as on every QuantVPS plan, so you can install compatible frameworks and tools.

Choose a preferred city from our eight cities: Chicago, New York, London, Dublin, Amsterdam, Frankfurt, Tokyo or Singapore. You can also choose no preference. Our team confirms where your build can go in your quote.

Yes. GPU servers have the 99.999% monthly uptime target of our SLA, as our VPS and dedicated servers do. If we miss it, you are eligible for a service credit. QuantVPS deploys in SOC 2 Type II certified, Tier IV facilities with fault-tolerant power and cooling, 24/7 on-site security, and biometric access controls. Read the SLA.

Your GPU server.
Built to your request.

Choose the GPU and the parts, and see the estimate as you go. Our team checks availability and the price, then sends your quote by email.

Talk to our team