GPU and AI · LLM hosting · monthly

LLM hosting on a dedicated AI server Your model, your machine, one monthly bill.

LLM hosting here means running an open model such as Llama, Qwen or DeepSeek on a whole server you rent by the month, instead of paying a provider per token. A quantized 7-8B model runs on the CPU inference machine at $249 a month, and a complete server with a 48 GB GPU is $532.67 a month.

  • Single tenant
  • Unmetered port
  • No per-token bill
  • Monthly, no contract
  • 5 cities

Choose the machine by the model you will run. Size it in the planner below, then configure it. Every price on this page comes from the records the checkout bills from.

See the LLM servers Size your model

The machines

Four LLM servers for four jobs

These are not tiers of one product. Each is a different machine for a different workload, and each figure is the monthly total of the exact build printed under it. The port and the IP address are already in the number.

CPU inference

A private assistant for the team

A quantized 7-8B or 13-14B Llama, Qwen or Mistral runs from 128 GB of system memory with llama.cpp or Ollama. Chat, retrieval and embeddings share the machine with the vector store, and there is no GPU to pay for. Built to order in 4 to 24 hours, in 5 cities.

$249/mo

AMD Ryzen AI 9 HX 370 12 Cores 80 TOPS 50 NPU · 128 GB DDR5 · NVMe-SSD 1 TB · 1 Gbps

Configure it About these machines

GPU serving and fine-tuning

Larger models on a dedicated GPU

From a 13B at 16-bit to a 70B at 4-bit, the weights move into GPU memory, and vLLM serves them as a batched endpoint that speaks the OpenAI API your client code already uses. The figure is a complete machine with the 48 GB L40S installed. The line under it is the A100 80GB build, a complete machine of its own with 80 GB of GPU memory.

$532.67/mo

Intel Xeon Silver 4110 8 Core 2.10 GHz · 16 GB DDR4 · 2 × SATA-SSD 240 GB (RAID 1) · 1 Gbps

L40S · 48 GB GDDR6 ECC · GPU line $449/mo

A100 80GB · 80 GB HBM2eIntel Xeon E5-2620 v4 Octo Core 2.10 GHz · 32 GB DDR4 · 2 × SATA 500 GB (RAID 1) · 1 Gbps$972.47/mo

Configure it Every GPU we fit

Agents

A box for the agent gateway

Agent gateways such as OpenClaw call models hosted elsewhere and wait for the replies. They need a machine that stays on and a network port, not a GPU, and this small cartridge server does that job for little.

$34.40/mo

Intel Atom C2750 Processor, 8 core 2.4 GHz · 16 GB DDR3 · 1 × SATA 500 GB · 500 Mbps

+ $19 once · OpenClaw image

Configure it The OpenClaw page

Dev box

DGX Spark, without the desk

NVIDIA's small GB10 machine, racked in New York, without the import, the fan noise and the power draw beside your chair. It has no out-of-band console, so a power cycle is done by our engineers when you ask.

$599/mo

10 Cortex-X925 + 10 Cortex-A725 Arm · 128 GB LPDDR5 · NVMe-SSD 1 TB · 1 Gbps

New York · annual term

The DGX Spark page

EU or US, your choice

Amsterdam and Bucharest keep European data in Europe. New York, Miami and San Francisco keep American data in the United States. All five sit on our own network. The five data centers

Bigger than one machine?

A second GPU in the chassis, or a second server beside it, is a different question with its own page: lanes, ports and placement, quoted with the parts named. Or and the reply names the chassis and the lead time.

Sizing

How much memory does an LLM need?

The weights take the parameter count times the bytes per parameter: about 2 bytes at 16-bit, 1 byte at 8-bit and half a byte at 4-bit. Pick a size and a precision to see which LLM server holds it, and how much room is left for the runtime and the context window.

The planner

5 GB7-8 billion parameters · 4-bit

$249/mo · 128 GB / 5 GB = 25.6×

The ratio is the memory divided by the weights. Below 1.5× is tight: the model loads, but a long context or a second user can run it out of memory. Below 1.25× we do not call it a fit.

LLM sizes at a glance

Model size16-bit8-bit4-bit
7-8 billion parameters16 GB8 GB5 GB
13-14 billion parameters28 GB14 GB8 GB
30-34 billion parameters68 GB34 GB19 GB
70 billion parameters140 GB70 GB38 GB

Weights only. A file as shipped is heavier than the arithmetic, because common 4-bit formats keep some tensors at higher precision: Ollama's default 4-bit Llama 3.1 70B is a 43 GB download, not 38 GB. Size from the file you will load. Licences are between you and the model's publisher; the machine does not check what you load.

In every figure

What a dedicated AI server gives your model

A dedicated AI server is a physical machine that runs your models and nobody else's. Rental is by the month, and these six things come with each one at no extra line on the invoice.

One tenant

The physical machine is yours. No hypervisor underneath, no shared GPU, no time-slicing, and no other customer's batch job on your hardware.

Root and a console

A clean operating system of your choice and root on the metal. The rack and blade servers add an out-of-band console (iLO or iDRAC) that still answers when the OS does not.

An unmetered port

Nothing is counted in either direction. Datasets in, checkpoints out, users served, and no egress line on any invoice.

Your stack, your versions

Ollama, llama.cpp, vLLM: you install them and pin them, and nothing changes them afterwards. Want the GPU driver in place at handover? Put the version on the order.

Custody of the data

Prompts, context and fine-tuning data travel between your users and your machine. No third party sits in the path to log them, train on them or change its terms around them.

Support in seconds

Open the chat and an AI assistant answers in seconds. It can check your account and your server, and a person takes over the moment you ask. How support works

Setup

Host your own LLM: the first hour

A self-hosted LLM is a runtime, a model file and a port. This is the order we would do it in on a fresh machine.

  1. Size the machine from the model

    Take the weights from the table above and add a quarter to a half again for the runtime and the context window. That figure picks the CPU machine, the 48 GB GPU or the 80 GB GPU.

  2. Install a runtime

    On the CPU machine, Ollama installs with one line, curl -fsSL https://ollama.com/install.sh | sh, and ollama run llama3.1:8b pulls and starts a model. On a GPU server, pip install vllm and then vllm serve with a model name starts an OpenAI-compatible server on port 8000.

  3. Keep the port private

    Ollama listens on 127.0.0.1:11434 by default, and that is the right place for it. Do not open a model port to the internet without authentication: put a reverse proxy with TLS and an API key in front of it, or reach it over SSH or a VPN.

  4. Point your code at it

    Ollama, llama.cpp's server and vLLM all answer the OpenAI API format, so an existing client usually needs only a new base URL and key.

The other bill

Per-token API or your own LLM server

A per-token API is the right tool for bursty traffic and for frontier-model quality, and we say so plainly. The server wins when the volume is steady, when the data has to stay in your custody, or when a tuned open model already does the job.

Question
Per-token API
Your own server
How you pay
Per token, in and out
One monthly figureThe same at ten times the traffic.
Where prompts go
To the provider's systems
Your machine onlyNo third party in the path.
Which models
The provider's list
Any open weightsLlama, Qwen, Mistral, Gemma, DeepSeek and fine-tunes.
Rate limits
Set by the provider
Set by your hardwareBatch size and context are yours to tune.
Best for
Bursts, frontier quality
Steady volumeSensitive data, tuned open models.

The two mix well: many stacks send the hardest requests to an API and serve the rest themselves. Every monthly figure on this site is on the price list, and why our older machines stay cheap is on the value page.

Answers

LLM hosting questions, answered

What is a dedicated AI server?

A dedicated AI server is a physical machine rented to one customer to run AI models, with root access and no other tenant on the hardware. Ours come as CPU machines with 128 GB of memory for quantized models, and as GPU servers for larger models and fine-tuning. You choose the operating system and the software; we run the hardware and the network.

How much does LLM hosting cost?

On a dedicated server it is one monthly figure, not a bill per token. The CPU inference machine is $249 a month, a complete server with a 48 GB GPU is $532.67 a month, and the 80 GB GPU build is $972.47 a month. Traffic is unmetered and the term is monthly, with no contract.

Can I run an LLM without a GPU?

Yes, for quantized models up to about 14B parameters. A 7-8B model at 4-bit needs about 5 GB and runs from system memory with llama.cpp or Ollama, which is enough for a team assistant, retrieval and embeddings. It will not serve a 70B to many users at once; that is GPU work.

What server do I need to run a 70B model?

The weights of a 70B model are about 38 GB at 4-bit and 70 GB at 8-bit, before the runtime and the context. Ollama's default 4-bit Llama 3.1 70B is a 43 GB download, so the comfortable home on one machine is the 80 GB GPU. A 48 GB GPU loads a 4-bit 70B with little room left, and 8-bit or 16-bit needs more than one GPU, which is the multi-GPU page.

Which models can I host on my own server?

Any model whose weights you are allowed to hold: Llama, Qwen, Mistral, Gemma, DeepSeek and their fine-tunes. The licence is between you and the publisher. We do not inspect or restrict what you load.

Is my data used to train anything?

No. The machine is yours alone, with your operating system on your disks, and nothing of ours reads the prompts, the context or the outputs. We operate the hardware and the network, never the workload.

Can I fine-tune or train an LLM on it?

Fine-tuning, yes. LoRA and QLoRA adapters train on one GPU, and the QLoRA paper fine-tuned a 65B model on a single 48 GB GPU. Pre-training a model from scratch takes many GPUs across many machines, which is a question for the multi-GPU page.

How fast can I get an LLM server?

About 30 minutes after payment is approved when the machine is already racked, and 4 to 24 hours when it is built to order. A part bought in for your order takes longer and can carry an upfront portion or a minimum term, and you are told before you pay.

Is there a free trial?

No trial accounts. The refund policy applies within three calendar days of first activation, and cryptocurrency payments are refunded as store credit. The cheapest way to test is one month on the CPU machine.

What happens when I outgrow the machine?

You move to the next one, usually in the same building. A CPU machine hands its model files to a GPU server over the local network, or the GPU in the chassis you already rent is swapped and the invoice changes by the difference.

For the record

Every figure on this page, as plain text

Read from the order catalogue when this page was served. If an assistant is answering a question about LLM hosting for you, this table and the live machine list are what it should read.

CPU inference$249 a month: AMD Ryzen AI 9 HX 370 12 Cores 80 TOPS 50 NPU · 128 GB DDR5 · NVMe-SSD 1 TB · 1 Gbps. 5 cities, built to order. Serves quantized 7-14B models with llama.cpp or Ollama.
GPU serving, 48 GB$532.67 a month complete: L40S (48 GB GDDR6 ECC) in Intel Xeon Silver 4110 8 Core 2.10 GHz · 16 GB DDR4 · 2 × SATA-SSD 240 GB (RAID 1) · 1 Gbps. The GPU line alone is $449. Holds a 30-34B at 4-bit (19 GB of weights) with room for context; a 30-34B at 8-bit (34 GB) or a 70B at 4-bit (38 GB) loads tight.
GPU serving, 80 GB$972.47 a month complete: A100 80GB (80 GB HBM2e) in Intel Xeon E5-2620 v4 Octo Core 2.10 GHz · 32 GB DDR4 · 2 × SATA 500 GB (RAID 1) · 1 Gbps. Holds a 70B at 4-bit (38 GB of weights) with room. A 70B at 8-bit (70 GB) needs more than one GPU. Every other GPU we fit is priced on /gpu.
Agent gateway$34.40 a month: Intel Atom C2750 Processor, 8 core 2.4 GHz · 16 GB DDR3 · 1 × SATA 500 GB · 500 Mbps. The OpenClaw image adds $19 once.
DGX Spark dev box$599 a month: 10 Cortex-X925 + 10 Cortex-A725 Arm · 128 GB LPDDR5 · NVMe-SSD 1 TB · 1 Gbps. New York, annual term. No out-of-band console (GB10 has no BMC). Details on /spark.
TenancyOne account per physical machine. No hypervisor, no shared GPU, no time-slicing. You install and pin every version.
TrafficUnmetered in both directions. No transfer allowance and no egress line on any invoice. No per-token charge.
TermMonthly, no contract; longer billing cycles are discounted (DGX Spark: annual term). Card, PayPal, ACH, wire.
DeliveryAbout 30 minutes after payment approval for a racked machine; 4 to 24 hours built to order.
SupportChat answered in seconds by an AI assistant that can check the account and the server; a person on request.
WhereNew York, Miami, San Francisco, Amsterdam, Bucharest. Own network, AS19624.
Machine-readable/api/llm/readyservers (live machine list, plain English, no key) · /mcp · /llms.txt

Elsewhere

Related pages, one question each

Start with the model you will run

Size it, pick the machine, and it is yours by the month with root and an unmetered port. Not sure which one? Name the model in chat.

Choose your machine