Dedicated servers · AMD Ryzen AI · monthly
Yes, you can rent one: a Ryzen AI dedicated server here is a whole AMD machine, either the Ryzen AI Max+ 395 at $399 a month or the Ryzen AI 9 HX 370 with 128 GB of memory at $249 a month. You get root on the hardware, an unmetered 1 Gbps port and no contract.
The CPU cores, the integrated Radeon and the NPU sit in one package and read from one pool of memory. A model that fits in that memory runs with no discrete GPU in the machine. That is the case for renting one: AI inference without a GPU, on hardware nobody else is scheduled onto.
The builds
Both are fixed builds. The processor and the memory come as a pair, so the choice is between speed and room.
Faster package
$399/mo
Take it when the model fits in 64 GB and you want it to answer quickly. The larger Radeon is the reason to pay more.
Configure this buildMore memory
$249/mo
The cheaper machine holds twice the memory. Pick it when the weights you load are bigger than the other build can hold.
Configure this build| Spec | AMD Ryzen AI Max+ 395 | AMD Ryzen AI 9 HX 370 |
|---|---|---|
| Processor | AMD Ryzen AI Max+ 395 | AMD Ryzen AI 9 HX 370 |
| CPU cores | 16 | 12 |
| AI throughput | 126 TOPS | 80 TOPS (50 on the NPU) |
| Graphics | AMD Radeon 8060S | AMD Radeon 890M |
| Memory | 64 GB DDR5 | 128 GB DDR5 |
| Storage | 1 TB NVMe, one slot | 1 TB NVMe, one slot |
| Network | 1 Gbps unmetered | 1 Gbps unmetered |
| OS | Fedora 43, root | Fedora 43, root |
| Per month | $399 | $249 |
| Setup | $29 once | $29 once |
Prices are per month before sales tax. Orders in Bucharest and Amsterdam are billed the same figure in euros.
Inside the package
On a GPU server the model is copied into the GPU's own memory before it runs. Here the GPU has no separate memory, so there is no copy.
The CPU cores, the Radeon and the NPU share the system memory. A model loaded once is visible to all three, so you can run it on the cores with llama.cpp, offload layers to the Radeon through llama.cpp's Vulkan backend, or test what your toolchain can push to the NPU.
The trade-off is speed. Memory bandwidth on a laptop-class package is far below what a datacenter GPU has, so tokens arrive more slowly and fewer users can be served at once. For one team, a private assistant or a batch job that runs overnight, that is usually fine. When it stops being fine, a dedicated GPU server is the next step.
Model sizing
Weights take about half a gigabyte per billion parameters at 4-bit quantisation and about one gigabyte at 8-bit. Leave roughly a quarter of the memory for the OS and the context cache.
| Model size | Weights at 4-bit | Weights at 8-bit | Ryzen AI Max+ 395 | Ryzen AI 9 HX 370 |
|---|---|---|---|---|
| 8B | about 5 GB | about 9 GB | fits at 8-bit | fits at 8-bit |
| 14B | about 9 GB | about 15 GB | fits at 8-bit | fits at 8-bit |
| 32B | about 19 GB | about 34 GB | fits at 8-bit | fits at 8-bit |
| 70B | about 40 GB | about 75 GB | 4-bit only | fits at 8-bit |
| 120B | about 65 GB | about 125 GB | too large | 4-bit only |
Rough figures for planning, weights only. Long contexts and many parallel requests need more. The LLM server planner works it out for a specific model.
Delivery
4–24 h
We do not keep these racked and waiting. Each is built in the city you choose, and the configurator quotes 4 to 24 hours. Because it is a custom build, the first invoice carries a one-time $29 setup fee.
root
It arrives with a clean Fedora 43 and root over SSH. No inference service is preinstalled. Ollama, llama.cpp, PyTorch or your own build go on at the versions your code was tested with.
1 Gbps
The gigabit port is unmetered: no transfer allowance and no overage line on the invoice. One IPv4 address is included and IPv6 is available at checkout.
Built in New York, Miami, San Francisco, Bucharest and Amsterdam, at the same monthly figure in each. Billing is monthly; 3, 6 or 12-month cycles take 5, 10 and 15 percent off.
Included and optional
Basic management is part of the price: 12-hour response and failed hardware replaced within 24 hours. The paid tiers shorten both and add monitoring and alerts.
| Drive | Per month |
|---|---|
| 1 TB NVMe | included |
| 2 TB NVMe | +$16.40 |
| 3.84 TB NVMe | +$65.40 |
| Tier | Response | Hardware swap | Per month |
|---|---|---|---|
| Basic | 12 h | 24 h | included |
| Bronze | 4 h | 4 h | +$19 |
| Silver | 2 h | 3 h | +$39 |
| Gold | 1 h | 2 h | +$69 |
| Platinum | 30 min | 2 h | +$99 |
Encrypted backup plans are offered at checkout too. Support starts on live chat with an AI assistant that answers in seconds and can check your account and your server; a person takes over whenever you ask.
Other routes
Answers
Yes. We rent two as dedicated servers: the AMD Ryzen AI Max+ 395 with 64 GB of memory and the Ryzen AI 9 HX 370 with 128 GB. Each is a whole machine with root access, billed monthly with no contract.
The Ryzen AI Max+ 395 build is $399 a month and the Ryzen AI 9 HX 370 build is $249 a month, each with a 1 TB NVMe and an unmetered 1 Gbps port. Both are built to order, so a one-time $29 setup fee is added to the first invoice.
Choose by memory first. The HX 370 has 128 GB, twice the Max+ 395, so it holds models the faster machine cannot. Choose the Max+ 395 when your model fits in 64 GB and response time matters more.
Yes, within its memory. The model loads into system memory, which the CPU cores, the integrated Radeon and the NPU share. A 4-bit 70B model is about 40 GB of weights and fits on either build. It answers more slowly than on a dedicated GPU, so heavy multi-user traffic belongs on a GPU server.
A whole machine. You are the only tenant, with root on the hardware and no hypervisor between you and the processor. The memory, the Radeon and the NPU are not split with anyone.
It uses the same processors that AI mini PCs are built around. The difference is where it lives: ours sits in our datacenter on our network, with power, cooling, a gigabit port and hardware replacement handled for you.
No. It ships with a clean Fedora 43 and root access. You install the inference stack you use, at the versions you have tested.
One NVMe slot: 1 TB is included, and the configurator can swap it for 2 TB or 3.84 TB. There is no second slot, so for RAID or several drives ask us for a custom build.
Yes. Servers ordered in Bucharest or Amsterdam are billed in euros, at the same figure the US cities bill in dollars.
4 to 24 hours after payment is approved. Each machine is built for its order rather than taken from a shelf. If your date is tight, ask on chat before ordering and we will tell you where the queue stands.
No. Our Ryzen line is these two Ryzen AI machines. For a general-purpose server, the price list shows every line we sell.
Pick the build and the city. The configurator shows the full monthly total and the setup fee before you pay.