The full GPU
PCIe passthrough into your operating system. No hypervisor, MIG partition, vGPU profile or time-slicing sits between your code and the GPU.
GPU servers for training and inference
Dedicated AI server Full GPU. Your OS. A port per node.
A dedicated AI server is a whole physical machine with its GPU passed through to your own operating system, rented by the month. Ours start at $169.93 a month with Instinct MI50 (16 GB HBM2), and they scale two ways: more GPUs in one chassis, or more nodes on one switch.
- Single tenant
- PCIe passthrough
- No egress fee
- 5 cities
- No contract
This page is for the step after the first GPU. It compares every platform we fit a GPU in on PCIe lanes, port speed and city, prices the network for a cluster per node, and gives the delivery dates before you order.
- 4platforms that take a GPU
- 100 Gbpstop port, per node
- 2sockets, one chassis
- 1 TBRAM, one chassis
- 5cities
Included on every node
What every dedicated AI server includes
The same four things on a single GPU server for deep learning or on a cluster of twenty.
Your own stack
A clean operating system with root. You pin the driver, the CUDA or ROCm release and the framework builds your code was tested against.
A port with no meter
Unmetered in both directions. Datasets in, checkpoints out and replication between nodes never show up as an egress line.
Answers in seconds
Our AI chat replies in seconds and can check your account and your servers while it works. Ask for a person and one takes over.
Scaling past one GPU
Deep learning server: add a GPU or add a machine
They fix different shortages, cost differently and arrive on different dates. Find where the limit is before you compare prices.
Multi-GPU server
More GPUs in one chassis
Choose this when the GPU is the limit: the model no longer fits in GPU memory, or work queues on the GPU while the rest of the machine idles. More than one GPU per chassis is quoted, not ticked, because slot clearance, PCIe lanes and the power supply set the count for each GPU model.
- Decided by
- The chassis. Name the GPU and the quote names the chassis, the count and the date.
- Know this
- Two GPUs keep two memory pools. One large job has to be split with tensor or pipeline parallelism and pays for the traffic between the GPUs. Independent jobs simply run twice as many at once.
GPU cluster
More machines on one switch
Choose this when the machine is the limit. The largest GPU platform here stops at 2 sockets and 1 TB of RAM. Past that, the next unit of capacity is another server.
- Decided by
- The city and the port. Nodes that work together go in one building, and each one pays for its own port.
- Know this
- A cluster here is a set of separate servers, each ordered, billed, upgraded and replaced on its own. There is no cluster SKU, so there is no cluster premium and no minimum node count.
Not sure which side you are on? Send the model size, the dataset and the deadline in . If one machine covers it, we say so.
Live from the catalog
Machine learning server platforms, side by side
Every platform we fit a GPU in, on the figures that size a cluster. The PCIe generations match the GPU page. Every other figure comes from the order catalog each time this page loads.
| Platform | PCIe to the GPU | Port ceiling | Sockets | RAM ceiling | Cities |
|---|---|---|---|---|---|
| Intel Xeon Silver / Gold | 48 lanes per socketPCIe 3.0 | 100 Gbps1 Gbps included | 2 | 1 TB | 5 |
| Intel Xeon E5-2600 v3/v4 | 40 lanes per socketPCIe 3.0 | 40 Gbps1 Gbps included | 2 | 1 TB | 5 |
| Intel Xeon E5-2600 v1/v2 | 40 lanes per socketPCIe 3.0 | 20 Gbps1 Gbps included | 2 | 256 GB | 3 |
| AMD EPYC | 128 lanes per socketPCIe 4.0 | 20 Gbps1 Gbps included | 2 | 1 TB | 2 |
- PCIe to the GPU limits host-to-GPU transfer: streaming training batches, loading scenes, anything that spills out of GPU memory. A model that stays loaded hardly uses it.
- Port ceiling limits node-to-node transfer. One machine never notices it. A job split across two machines depends on it.
- Cities comes first for a cluster. A platform your city does not hold is not an option.
- The trade: the widest bus, AMD EPYC on PCIe 4.0, tops out at 20 Gbps. The 100 Gbps port is on Intel Xeon Silver / Gold, on an older bus. No platform has both, so decide which one your job needs.
Which GPU goes in each platform, and its monthly price beside the machine it belongs to, is on GPU dedicated servers.
Cluster networking
GPU cluster ports, priced per node
A port is bought per machine, so the network under a cluster is the same line once per node, every month. The table does that multiplication for you.
| Port | 1 node | 2 nodes | 4 nodes | 8 nodes | Platforms |
|---|---|---|---|---|---|
| 1 Gbps | included | included | included | included | all 4 |
| 2 Gbps | $189 | $378 | $756 | $1,512 | all 4 |
| 3 Gbps | $299 | $598 | $1,196 | $2,392 | all 4 |
| 5 Gbps | $459 | $918 | $1,836 | $3,672 | all 4 |
| 10 Gbps | $529 | $1,058 | $2,116 | $4,232 | all 4 |
| 20 Gbps | $1,629 | $3,258 | $6,516 | $13,032 | all 4 |
| 40 Gbps | $2,799 | $5,598 | $11,196 | $22,392 | 2 of 4 |
| 100 Gbps | $3,999 | $7,998 | $15,996 | $31,992 | 1 of 4 |
- Steps are not linear. At several points, more nodes one step down cost less than fewer nodes one step up. If the job partitions well, that buys an extra server.
- The top steps are narrower. The fastest ports exist on fewer platforms, and not on the one with the widest bus.
- Ethernet, stated plainly. The link between two of our machines is an Ethernet port. No InfiniBand line and no GPU-to-GPU bridge between machines is on any order form. If your job needs one, ask before ordering: it is a quote, and the answer can be no.
Every speed is unmetered in both directions with no transfer allowance. For one port on one machine, see unmetered bandwidth.
Placement
Pick the city before the platform
Nodes that share a job belong in one building, and our five cities do not all hold the same platforms. New York and Bucharest hold all of them. Check the city first, then configure.
- 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
Power, carriers and access for each building are on data centers. If latency between two of our cities matters to your design, ask support to measure it between those rooms.
From order to racked
Delivery dates come with the quote
A multi-node build can involve several machines and a purchase order for parts. The quote names the slowest part before you pay, using the same lead times as build to order.
- ~30 min
Racked, taken as it is
The machine already exists and is handed over. Instant servers lists what is racked now, and it is the first place to look for the plain nodes of a cluster.
- 4–24 h
Built from our stock
GPUs fitted, disks set up, operating system installed. Several nodes are built in parallel.
- 5–10 days
Parts bought in
A part we do not keep on the shelf is ordered first. For a multi-GPU chassis this is the usual case, and the chassis is usually the slow part.
- 10–15 days
GPUs on allocation
Current accelerators ship when the distributor ships them. Hardware bought in for one order is paid three months up front.
Working days in the last two rows. Every clock starts once payment and fraud review clear. On a large first order, a bank wire usually clears review faster than a card payment.
Answers
Dedicated AI server questions
What is a dedicated AI server, exactly?
A physical server rented to one customer, with one or more GPUs passed through to the operating system that customer installs. No hypervisor sits between your code and the GPU, and nothing on the machine is shared: the cores, memory, disks, the full GPU and the network port are yours, with root access.
What does a dedicated AI server cost per month?
The lowest complete price with a training-class GPU is $169.93 a month: Instinct MI50 (16 GB HBM2) in Intel Xeon E5-2630L Hex Core 2.00 GHz · 32 GB DDR3 · 2 × SATA 500 GB (RAID 1) · 500 Mbps. Larger GPUs cost more, and every one is listed with its complete monthly price on GPU dedicated servers. A build with a GPU fitted to order adds a one-time setup of $29. In a cluster, each node also pays for its own port above the included speed.
Which GPU is best for a deep learning server?
Start from GPU memory, not speed. The weights, activations and optimizer state have to fit, and a job that does not fit is not fixed by a faster GPU. We list 15 training-class GPU models, with up to 80 GB of memory on one GPU. For training that streams batches from the host, also check the PCIe column; for serving a model that stays loaded, it matters much less.
Can I put several GPUs in one dedicated AI server?
Often, and it is quoted rather than ticked at checkout. Slot clearance, the width of the GPU, the PCIe lanes per socket and the power supply set the count for a given chassis. Tell us the GPU in chat and the quote comes back with the chassis, the count and a date. Two GPUs do not pool their memory unless your framework splits the model across them.
Do you offer GPU clusters with InfiniBand?
Not as a catalog item. Nodes connect over Ethernet, one port per node, from 1 Gbps included up to 100 Gbps. InfiniBand, or a GPU-to-GPU bridge between machines, is a quote, and the answer can be no. Ask before you plan a job around it.
How fast can two of my servers talk to each other?
As fast as the slower of their two ports. The top port is 100 Gbps, available on Intel Xeon Silver / Gold. Keep the nodes in one city: between cities, distance adds latency that a faster port does not remove.
Is traffic between my nodes metered?
No. Every port is unmetered in both directions, with no transfer allowance and no egress charge. Copying datasets, syncing gradients and writing checkpoints between your nodes costs the port price and nothing more.
Do I share the GPU with other customers?
No. One machine, one account. The GPU is on PCIe passthrough, with no MIG partition, no vGPU profile and no time-slicing, so the benchmark you run on day one still holds on day ninety.
Which driver and CUDA version comes installed?
The ones you install. The server is handed over with a clean operating system and root access, and you choose the driver, the CUDA or ROCm release and the framework versions. If you want a driver in place at handover, name the exact version on the order.
How fast is a dedicated AI server delivered?
About 30 minutes after payment approval for a machine that is already racked and taken as it is. 4 to 24 hours when we build it from our stock, 5 to 10 working days when a part is bought in, and 10 to 15 working days for GPUs on distributor allocation. A quote for several nodes names the slowest part before you pay.
Where are your AI servers located?
New York, Miami, San Francisco, Amsterdam and Bucharest. Not every city holds every GPU platform: New York and Bucharest hold all 4. Machines in Amsterdam and Bucharest are billed in euros.
Can I rent an AI server monthly, without a contract?
Yes. Each machine runs month to month with nothing to sign. 3, 6 and 12-month cycles save up to 15%. Hardware bought in for a single order is paid three months up front.
For the record
This page in figures
Read from the order catalog when this page was served, the same rows the checkout bills from.
- Lowest price
- $169.93 a month, complete: Instinct MI50 (16 GB HBM2) in Intel Xeon E5-2630L Hex Core 2.00 GHz · 32 GB DDR3 · 2 × SATA 500 GB (RAID 1) · 500 Mbps. Every other GPU and its monthly price: GPU dedicated servers.
- Training-class GPUs
- 15 models classed ML / HPC in the catalog, up to 80 GB of GPU memory each.
- GPUs per chassis
- One is a checkbox. More than one is quoted with the GPU named, because slot clearance, lanes and the power supply decide the count.
- Nodes per cluster
- Any number. Each node is its own server, invoice and term; no cluster SKU, no cluster premium, no minimum.
- GPU platforms
- 4. Widest bus: AMD EPYC (PCIe 4.0, 128 lanes per socket), port ceiling 20 Gbps, 2 of 5 cities. Fastest port: 100 Gbps on Intel Xeon Silver / Gold.
- Link between nodes
- Ethernet, one port per node: 1 Gbps included on each, priced steps up to 100 Gbps. No InfiniBand or GPU bridge in the catalog; ask for a quote.
- One machine tops out at
- 2 sockets and 1 TB of RAM.
- Cities
- 5: New York, Miami, San Francisco, Amsterdam, Bucharest. All GPU platforms in New York and Bucharest.
- Traffic
- Unmetered in both directions at every port speed. No egress charge.
- Tenancy
- One account per physical machine. GPUs on PCIe passthrough: no hypervisor, MIG, vGPU or time-slicing.
- Software
- Clean operating system with root. You install the driver, CUDA or ROCm release.
- Delivery
- About 30 minutes for a racked machine taken as it is; 4 to 24 hours built from stock; 5 to 10 working days for bought-in parts; 10 to 15 for allocation parts.
- Term
- Monthly, no contract. 3, 6 and 12-month cycles save up to 15%. Hardware bought in for one order is paid three months up front.
- Support
- AI chat that answers in seconds and can check your account and servers; a person on request.
Listed by their owners
Multi-GPU and linked machines on the marketplace
Machines with two or more GPUs, and linked pairs, that owners list on our marketplace at their own monthly price. Primcast, our sister company, lists some; those are rented on primcast.com.
Related pages
Where the next question is answered
GPU dedicated serversEvery GPU we fit, each priced beside its machine.
LLM hostingRunning a language model on one machine you control.
NVIDIA DGX SparkA GB10 machine, and up to four of them networked together.
Unmetered bandwidthOne port on one machine, every speed above the included one.
Build to orderA whole specification, priced and dated before you pay.
Instant serversWhat is racked and powered now.
Data centersThe five buildings a cluster can live in.
PricingEvery monthly figure on the site, in one table.
Older hardware, lower priceThe cheaper end of these platforms, and why it is still racked.
SupportReplacement windows per plan, and the credits behind them.
Tell us the model, the data and the deadline
Configure the lowest-priced build now, or describe the job in chat and get a quote for the GPUs, the nodes and the date.