The full GPU, not a slice
Every GPU is passed straight to your operating system. No hypervisor, no MIG partition, no vGPU profile and no time-slicing, so your drivers and frameworks see the hardware as it is.
Dedicated AI servers · training, serving, rendering
Dedicated AI server hosting from one GPU to a full cluster
A dedicated AI server is a whole physical machine with its GPUs passed straight to your own operating system, so the model you train or serve shares the hardware with nobody. When one machine is not enough, you can add GPUs to it or put more machines beside it. This page shows what each path costs, how fast the machines can talk to each other, and which of our 5 cities holds which platform.
Monthly, no contract. Each machine's port is yours alone at the switch, and nothing meters what goes through it.
GPU platforms4
Fastest port100 Gbps
Memory per machine1 TB
Longer-term discount15%
01What every machine comes with
What you get with a dedicated AI server here
The same four things on every node, from one machine for a fine-tuning job to several for a training run.
A port that is yours
The network port is unshared at the access switch and nothing meters it, in either direction. Copying a dataset or a checkpoint between your nodes adds nothing to the bill.
Month to month
Each machine runs on its own one-month term with no contract. Paying for three, six or twelve months takes up to 15% off.
Answers in seconds
Our AI chat replies in seconds and can look at your account and your servers while it helps. Ask for a person whenever you want one.
02Scaling past one GPU
More GPUs in one server, or more servers?
They sound like the same decision, but they cost differently, arrive on different dates and run out for different reasons. Work out which one you need before you compare anything else.
AGrow the machine
Add GPUs to the server you have
Pick this when you are short of GPU memory or GPU speed and the rest of the machine is fine. Several GPUs in one chassis is a build we quote rather than a box you tick at checkout: how many fit depends on how wide the GPU you choose is and how much power the chassis can feed it.
- What decides it
- The GPU you pick. Name it on chat and the reply names the chassis, how many fit and the date.
- Worth knowing
- Two GPUs are not one bigger GPU. Their memory only adds up when your software splits the model across both (tensor or pipeline parallelism, which the common training and serving tools support), and some speed goes on the traffic between them. If your work is many separate jobs, two GPUs simply get through twice as many.
BGrow the cluster
Add machines beside the first
Pick this when the machine itself is full: out of processors, out of memory, or out of room for the next job. The largest platform here stops at 2 processor sockets and 1 TB of memory, and past that the answer is another server.
- What decides it
- The city and the port. Machines that work together belong in the same building, and the link between them is the network port on each one, paid per machine.
- Worth knowing
- There is no cluster product and no minimum node count. Each server is ordered, billed, upgraded and replaced on its own, so you can start with two and add a third next month.
03The platforms
Pick your machine learning server by what limits you
Every platform we fit a GPU to, compared on what matters once you go past one machine. Read the two middle columns together: the platform with the widest path to the GPU has the slowest top port, and the fastest port sits on an older bus.
| Platform | Lanes to the GPU | Top port | Sockets | Memory, up to | Cities |
|---|---|---|---|---|---|
| Intel Xeon Silver / Gold | 48 lanes per socketPCIe 3.0 | 100 Gbps1 Gbps incl. | 2 | 1 TB | 5 |
| Intel Xeon E5-2600 v3/v4 | 40 lanes per socketPCIe 3.0 | 40 Gbps1 Gbps incl. | 2 | 1 TB | 5 |
| Intel Xeon E5-2600 v1/v2 | 40 lanes per socketPCIe 3.0 | 20 Gbps1 Gbps incl. | 2 | 256 GB | 3 |
| AMD EPYC | 128 lanes per socketPCIe 4.0 | 20 Gbps1 Gbps incl. | 2 | 1 TB | 2 |
- Lanes to the GPUHow fast data moves between the processor and the GPU. It matters when you stream training batches or a render feeds scene data every frame, and hardly at all once a model is loaded and stays in GPU memory.
- Top portHow fast one machine talks to the next. A single server doing its own work never notices it. For a job split across several machines, it is the whole question.
- SoA job that fits in one machine and leans on the bus wants the widest lanes. A job spread over several machines wants the port and can live with an older bus. Both at once is the one mix we do not have, and you should hear that before you pay.
Which GPU goes in each of these, and what it adds per month, is on GPU dedicated servers, every price shown against the machine it belongs to. This page prices no GPU, so a figure never lands next to the wrong server.
04The link between machines
A multi-node GPU cluster pays for the port on every node
The network link is part of each server, so four machines at 10 Gbps means four 10 Gbps ports every month. That line is the one that surprises people who have only ever priced a single server, so the table below does the multiplying for you.
| Port | One server | Two | Four | Platforms |
|---|---|---|---|---|
| 1 Gbps | incl. | incl. | incl. | 4 / 4 |
| 2 Gbps | $189 | $378 | $756 | 4 / 4 |
| 3 Gbps | $299 | $598 | $1,196 | 4 / 4 |
| 5 Gbps | $459 | $918 | $1,836 | 4 / 4 |
| 10 Gbps | $529 | $1,058 | $2,116 | 4 / 4 |
| 20 Gbps | $1,629 | $3,258 | $6,516 | 4 / 4 |
| 40 Gbps | $2,799 | $5,598 | $11,196 | 2 / 4 |
| 100 Gbps | $3,999 | $7,998 | $15,996 | 1 / 4 |
A wider cluster can cost less than a faster one. In places on that table, two machines one step down cost less than one machine one step up, and you end up with an extra server. That only pays off if your work splits across machines.
What we do not sell between machines. Our catalogue lists Ethernet ports and nothing else: no InfiniBand fabric and no NVLink bridge. If your build needs either, ask before you order. The answer is a quote, and it may be no. The smallest multi-node setup we rent is the NVIDIA DGX Spark, where up to four units wire together over ConnectX-7.
The port on a single machine, on its own, is covered on unmetered bandwidth.
05Where it can live
For a GPU cluster, choose the city first
Machines that work together belong in one building, and our data centres do not all hold the same platforms. Pick the platform first and you may find the city you need does not carry it. New York and Bucharest hold every one.
- New York4 / 4Intel Xeon Silver / Gold · Intel Xeon E5-2600 v3/v4 · Intel Xeon E5-2600 v1/v2 · AMD EPYC
- Bucharest4 / 4Intel Xeon Silver / Gold · Intel Xeon E5-2600 v3/v4 · Intel Xeon E5-2600 v1/v2 · AMD EPYC
- Miami3 / 4Intel Xeon Silver / Gold · Intel Xeon E5-2600 v3/v4 · Intel Xeon E5-2600 v1/v2
- San Francisco2 / 4Intel Xeon Silver / Gold · Intel Xeon E5-2600 v3/v4
- Amsterdam2 / 4Intel Xeon Silver / Gold · Intel Xeon E5-2600 v3/v4
Each building and what it connects to is described on data centres. If your setup has to span two cities, ask support to measure the latency between them before you order.
06Lead times
How soon your AI servers are running
A build like this is often several machines and parts that are not on a shelf, so you get a date before you pay, read off the same steps as our custom server builds.
- ~30 min
Already racked as it is
Nothing is fitted or moved. Instant servers lists the machines ready to go, worth a look first if your nodes are ordinary ones.
- 4–24 hours
Parts we hold
The usual case. Technicians fit the GPUs, build the disks and install the operating system on each machine, in parallel rather than one after another.
- 5–10 working days
Parts we buy in
Something in the specification is not a part we keep on hand, so we order it and then build. With several GPUs in one machine, the chassis is usually what takes longest.
- 10–15 working days
A scarce part
Current-generation accelerators ship on allocation, and the date is the distributor's. Where we buy hardware for one customer, we ask for three months up front.
Each step starts once the order clears payment and fraud review. Most clear within a couple of hours; a large first order from a new account can take longer, and a wire or stablecoin payment often clears faster than a card payment.
Marketplace
Multi-GPU machines listed by their owners
Owners list machines with several GPUs, and linked pairs, on our marketplace at a monthly price they set. Listings held by our sister company, Server Room, are rented on serverroom.net.
Own a GPU server that sits idle?
List it on the Primcast marketplace and rent it to AI teams by the month, at a price you choose.
07Questions
Questions about dedicated AI servers
01What is a dedicated AI server?
It is a physical server rented to you alone and fitted with one or more GPUs for training, fine-tuning or serving models. Here the GPUs are passed straight to your own operating system, with no hypervisor and nobody else on the machine, and you pay by the month with no contract.
02Can I rent a bare metal GPU server for deep learning?
Yes. Every GPU server we rent is bare metal: you get root on the whole machine and install the driver, CUDA version and framework your code expects. The GPUs and their monthly prices are on GPU dedicated servers, each shown against the machine it fits.
03Can one server hold more than one GPU?
Yes, as a quoted build rather than a checkout option. How many fit depends on the width of the GPU and the power the chassis can deliver, so tell us which GPU you want on chat and the reply names the chassis, the count and the date.
04How do I build a GPU cluster with you?
Order the machines in the same city, on a platform that reaches the port speed you need, and pick that speed on each one. They stay separate servers on separate invoices, so you can start with two and add more later. Tell us the node count and the job first and we will check the city and the platform before you pay.
05Do you offer InfiniBand or NVLink between servers?
Not as a listed option. Machines connect over Ethernet ports up to 100 Gbps. If your build needs InfiniBand or an NVLink bridge, ask before ordering: it would be a quote, and the answer may be no.
06Is a dedicated server better than cloud GPUs for machine learning?
It depends on how steadily you use it. A dedicated AI server costs the same each month whether it runs flat out or sits idle, so it suits long training runs and models served around the clock. Traffic is not metered here, so moving datasets between nodes or sending results to users adds nothing. For a job that runs a few hours a month, renting by the hour can cost less.
07What is an HPC server, and do you rent them?
HPC means high-performance computing: simulations, rendering, scientific code and model training spread over many processors or GPUs. We rent the parts, meaning machines with up to 2 sockets and 1 TB of memory, GPUs inside them and ports between them. You set up the scheduler yourself, Slurm or Kubernetes or whatever your team already uses.
08How much does a dedicated AI server cost?
The machine and its GPU are priced together on GPU dedicated servers. A port faster than the included one is the table on this page, per machine and per month. Paying for three, six or twelve months takes up to 15% off, and a custom build carries a one-time setup fee that the configurator shows before you pay.
09How long does it take to get one?
About four to twenty-four hours when we hold the parts, five to ten working days when they have to be bought in, and ten to fifteen working days for a scarce part. A machine that is already racked as you want it can run in about 30 minutes.
10Do you rent CMP 170HX servers for AI?
Yes. They have their own page, CMP 170HX servers, with the builds and their monthly prices.
11Where are your GPU servers?
In New York, Miami, San Francisco, Amsterdam and Bucharest. Not every platform is in every city; New York and Bucharest hold all of them, which makes them the easiest places to grow a cluster.
12What help do I get if something breaks?
Our AI chat answers in seconds and can check your account and your servers while it helps, and a person takes over whenever you ask. Response times and hardware replacement are on support.
08Related pages
Where to go next
- GPU dedicated serversEvery GPU we fit and what each costs on its machine.
- LLM hostingWhat a model of a given size needs to run well.
- DGX Spark hostingSmall units that wire together, up to four at a time.
- Custom server buildsA full specification priced and dated before you commit.
- Unmetered bandwidthPort speeds on one machine, and what each costs.
- Data centresThe five buildings and what each one holds.
Plan it with us
Tell us what your model has to do
The node count, whether the work splits, and what it has to talk to. Our AI chat answers in seconds, a person joins when you ask, and if one machine would do the job, we will say so.