GPU dedicated servers · NVIDIA Blackwell · by the month
An RTX 5090 dedicated server is a bare-metal machine rented by the month, and the full GPU is handed to your operating system by PCIe passthrough. The larger machines take up to 4 RTX 5090s, and a complete server starts at $472.47 a month.
The RTX 5080 is the 16 GB option on the same architecture, from $312.67 a month. Both prices cover processor, memory, disks, the port and one IPv4 address, with the GPU billed as its own monthly line. Choose the GPU by memory, then change the rest in the configurator.
The two GPUs
Both are GeForce Blackwell GPUs with GDDR7, 5th-generation Tensor cores that run FP4, 4th-generation RT cores and AV1 hardware encoding. The RTX 5090 has twice the CUDA cores of the 5080, twice the memory and a 512-bit bus with 1,792 GB/s. If your model or scene fits in 16 GB, the 5080 runs it for less a month, just more slowly.
32 GB GDDR7 · PCIe 5.0 x16 GPU
$472.47/mo, complete server
GPU $399/mo, plus the machine:
Intel Xeon E5-2620 v4, 8 cores, 2.10 GHz · 32 GB DDR4 · 2 × SATA 500 GB (RAID 1) · 1 Gbps
16 GB GDDR7 · PCIe 5.0 x16 GPU
$312.67/mo, complete server
GPU $229/mo, plus the machine:
Intel Xeon Silver 4110, 8 cores, 2.10 GHz · 16 GB DDR4 · 2 × SATA-SSD 240 GB (RAID 1) · 1 Gbps
Each price is the cheapest machine that takes that GPU, as the configurator opens it. The button opens that machine with the GPU already selected; faster processors, more memory, bigger disks and a faster port are priced there. Bucharest builds are billed in euros, and the configurator shows that price when you pick the city.
Host link
The RTX 5090 and RTX 5080 are built for PCIe 5.0 x16. The machine lines priced below run the GPU slot at PCIe 3.0 or PCIe 4.0, and the fastest of them is PCIe 4.0. For some jobs that decides which machine to buy, so here is what each one gives the GPU.
Inference with the weights already in GPU memory, video encoding, and rendering once the scene is loaded. Data crosses the bus at load time and then stays on the GPU, so a PCIe 3.0 link costs seconds at start-up and nothing after.
Training loops that copy a batch from system memory every step, renders that stream textures larger than GPU memory, and several GPUs splitting one model. GeForce RTX 50 GPUs have no NVLink, so multi-GPU traffic goes over PCIe too. For these, pick the machine with the faster link in the table below.
| Machine | PCIe link to the GPU | RTX 5080 | RTX 5090 |
|---|---|---|---|
| Intel Xeon E5-2600 v3/v4from $73.47/mo without a GPU | PCIe 3.0 | not offered | up to 4 |
| Intel Xeon Silver / Goldfrom $83.67/mo without a GPU | PCIe 3.0 | up to 4 | up to 4 |
| AMD EPYCfrom $217.59/mo without a GPU | PCIe 4.0 | not offered | up to 1 |
| Intel Xeon Gold 3rd gen, DDR4built to order in New York, priced in the configurator | ask us in chat | not offered | up to 2 |
| Intel Xeon Gold 4th gen, DDR5built to order in New York, priced in the configurator | ask us in chat | not offered | up to 2 |
Machine prices are the cheapest build of each line without a GPU. The newer lines built to order are priced in the configurator; ask us in chat which PCIe generation their GPU slot runs at before you order. RTX 50 servers can be ordered in New York, Miami, San Francisco and Bucharest; Amsterdam machines are not offered with these GPUs. Data centers describes each building.
Sizing for AI
For inference, memory decides first. Weights take parameters times bytes per parameter, and the GPU needs a quarter to a half more on top for the KV cache and the runtime. The table applies that rule to one RTX 5080, one RTX 5090 and a machine with several 5090s.
| Model size (weights) | RTX 508016 GB | RTX 509032 GB | 2 x RTX 509064 GB | 4 x RTX 5090128 GB |
|---|---|---|---|---|
| 7-8 billion parameters | 4-bit: Fits8-bit: Fits | 4-bit: Fits8-bit: Fits | 4-bit: Fits8-bit: Fits | 4-bit: Fits8-bit: Fits |
| 13-14 billion parameters | 4-bit: Fits8-bit: No | 4-bit: Fits8-bit: Fits | 4-bit: Fits8-bit: Fits | 4-bit: Fits8-bit: Fits |
| 30-34 billion parameters | 4-bit: No8-bit: No | 4-bit: Fits8-bit: No | 4-bit: Fits8-bit: Fits | 4-bit: Fits8-bit: Fits |
| 70 billion parameters | 4-bit: No8-bit: No | 4-bit: No8-bit: No | 4-bit: Fits8-bit: No | 4-bit: Fits8-bit: Fits |
A 30B model at 4-bit or a 13B one at 8-bit loads with context to spare, behind your own API on Ollama, llama.cpp or vLLM. The Blackwell Tensor cores also run FP4 natively, which runtimes built for it, such as TensorRT-LLM, use to serve 4-bit models faster.
Our larger machines take up to 4 of them, 128 GB of GPU memory in total, which vLLM or llama.cpp split a model across. Two hold a 70B model at 4-bit with room to spare. For one GPU with more memory, see every GPU we fit; for model-by-model plans, LLM servers.
Workloads
A model on your own machine, behind your own endpoint. Prompts and outputs stay on hardware rented to one account, and the bill is the same whether the GPU is busy all month or idle.
Blender Cycles renders on CUDA or OptiX, and the RT cores speed up ray tracing in it and in other GPU renderers. Install the operating system your renderer needs and queue frames overnight.
The RTX 5090 has three NVENC encoders and the RTX 5080 two, each handling H.264, HEVC and AV1 in hardware. Two NVDEC decoders on each take the input side off the processor.
The GeForce driver runs DirectX, Vulkan and OpenGL the way a desktop does. Host a game server, a cloud gaming box or a remote workstation with a dedicated GPU behind it.
Delivery and terms
It is installed and handed over with nothing to fit. Instant servers lists what is racked now, and those can also be rented by the hour.
A technician installs the GPU in the machine you configured, builds the disk array and installs the operating system. Extra memory or drives go in on the same visit.
We order the part, then build as above. You are told which of the three applies before you pay.
Each timeline starts once payment and fraud review clear. Most orders clear within a couple of hours; a first order on a new account can take longer.
Owners also list RTX 50 machines on our marketplace, each at a monthly price the owner sets. Listings that come through our sister company, Primcast, are rented on primcast.com.
Specifications
| GPU | Memory | Bus | Bandwidth | CUDA cores | Boost | Power | Video |
|---|---|---|---|---|---|---|---|
| RTX 5090 | 32 GB GDDR7 | 512-bit | 1,792 GB/s | 21,760 | 2.41 GHz | 575 W | 3 NVENC, 2 NVDEC |
| RTX 5080 | 16 GB GDDR7 | 256-bit | 960 GB/s | 10,752 | 2.62 GHz | 360 W | 2 NVENC, 2 NVDEC |
GPU figures as NVIDIA publishes them for the GeForce RTX 50 series. Prices, machines, cities and per-machine counts are read from our order catalogue, the one the checkout bills from.
Answers
A complete RTX 5090 server starts at $472.47 a month: the GPU at $399 plus the cheapest machine that takes it, Intel Xeon E5-2620 v4, 8 cores, 2.10 GHz · 32 GB DDR4 · 2 × SATA 500 GB (RAID 1) · 1 Gbps. A GPU fitted to order adds a one-time setup of $29, and there is no contract.
Only a machine that already has the GPU in it and is taken as it stands: those are billed by the hour, from a one-hour minimum, and instant servers lists which ones are racked now. A machine you configure here is built for you and billed by the month, with no contract.
No. NVIDIA sells the RTX 5090 only as a GeForce GPU. If you want a Blackwell GPU that NVIDIA also builds as a server edition, look at the RTX PRO 6000 Blackwell, with 96 GB of GDDR7 and ECC. We fit that 96 GB GPU too, priced on our RTX PRO 6000 servers page, and the RTX 5090 in rack servers with the full GPU passed through.
Yes. Our larger machines take up to 4 RTX 5090s, 128 GB of GPU memory in total, and the configurator shows the most each machine takes. Each GPU is its own monthly line, and the GPUs talk to each other over PCIe.
For inference and QLoRA fine-tuning, yes. Its 32 GB takes a 30B model at 4-bit or a 13B one at 8-bit with context to spare, and two of them hold a 70B model at 4-bit. It has no ECC mode, so for long training runs without checkpoints a data center GPU is the safer choice.
Pick by memory. The RTX 5080 has 16 GB and 960 GB/s, and a complete server starts at $312.67 a month; it holds an 8B model at 8-bit or a 13B model at 4-bit. The RTX 5090 has 32 GB and 1,792 GB/s, for larger models, heavier scenes and more encoding at once.
Not on the machines whose link we publish. The GPU is PCIe 5.0 x16, and the lines priced on this page run the slot at PCIe 3.0 or PCIe 4.0, the fastest being PCIe 4.0. For the newer lines built to order, ask us in chat which generation the slot runs at. Inference, encoding and most rendering barely notice; training that streams data from system memory every step does.
No. The machine is rented to one account and the GPU is passed straight through to your operating system. There is no hypervisor, no vGPU profile and no time-slicing, and you install the NVIDIA driver and CUDA version you choose.
About 30 minutes after the order is approved when a machine with the GPU is already racked. When we fit a GPU from our stock, allow 4 to 24 hours; when it has to be bought in, 5 to 10 working days.
In New York, Miami, San Francisco and Bucharest, on Server Room's own network. The machines that take RTX 50 GPUs are not offered with them in Amsterdam.
Pick the GPU and how many, then the processor, memory, disks and city. The configurator prices every change as you make it.
Every GPU we fit · RTX 4090 servers · LLM servers · Racked right now