Dedicated GPU servers, sold at cost.

We publish what each server costs us — hardware, power, rack, network, operations — and that is the monthly price. B300 to RTX 4090, 13 cards, no setup fee, no contract. Crypto only, no identity check.

See the servers Read the cost sheet

An eight-GPU accelerator baseboard with NVLink bridges under indigo light

Per server · per month · at cost

H100 SXM 80 GB · 1 card $1,134 cheapest found $1,230
B200 180 GB · 1 card $1,408 cheapest found $3,723
RTX 4090 24 GB · 1 card $139 cheapest found $274

Competitor prices are published list prices, checked 2 September 2026

0%margin on the server, by construction

13GPU models, B300 to RTX 4090

$0setup fee, egress, minimum term

4regions on renewable power

99.9%uptime SLA, credited ×10

Catalogue

Every server, priced from its cost sheet.

Per server, per month. Multiply by the number of cards and nothing else changes: the chassis share is already in each card’s cost. Open a cost sheet to see the six lines that make the price, or compare every card side by side.

Cards

Sort by

13 of 13 cards · single-card server · every region · no setup fee

A multi-card server costs the per-card price times the card count; NVLink and InfiniBand are included on SXM nodes. Prepay 3, 6 or 12 months and the per-month price drops 3, 6 or 10%: a longer term costs us less to run — one provisioning, no re-imaging between tenants, a firmer power forecast — and the difference is passed on. “Cheapest found” is the lowest published monthly price for the same card at a provider with public pricing, checked 2 September 2026; hourly clouds are converted at 730 hours.

How we price

Six lines, added up, rounded up. That is the price.

No “from”, no promotion, no quote. The cost sheet of every card is public, the assumptions are below, and the same numbers produce every price on this page.

01Hardware, over 36 months

The card at our invoiced purchase price plus its share of the chassis — CPUs, RAM, NVMe, NICs, baseboard — straight-line over 36 months. An H100 SXM: $27,000 plus $8,750 of HGX node, $993 a month.
02Power, at our contract rate

TDP plus the chassis share, at 70% average load, times a PUE of 1.15, times 730 hours, at $0.052 per kWh — geothermal, hydro and wind in the Nordics. The same H100: 529 kWh, $27.
03Rack, network, spares, operations

Rack space and cooling per card class, transit and one IPv4, a 2% per year reserve for failed cards, and the people who keep it running. Then the total is rounded up to the next dollar — never down.

Where the money comes from, then

A server at cost has to be paid for somewhere. These are the four places, and none of them is the server.

Options with a normal margin. Extra NVMe, extra IPv4 addresses, 25 Gbit/s uplinks, private VLANs. Priced below, optional, never bundled.
Power bought at scale. Long-term contracts in Iceland and Finland at rates a single tenant could not get. The cost sheet uses our rate, not yours.
Full racks. Hardware we own earns nothing while it waits for a buyer at a markup. A rack full of tenants at cost keeps the next rack funded.
No card network, no chargebacks, no sales team. Crypto settlement costs a fraction of a percent and never reverses. Nobody calls you; nobody’s pilot is subsidised with your invoice.

Price a server

GPU Cards per server Servers Months prepaid

The whole term is taken from your balance and renewed for the same term at the locked rate. Turn auto-renewal off and the server stops at the end of the paid term; nothing else to cancel.

Per month

$1,134

$1,134 × 1 card × 1 server
Cheapest competitor found: $1,230 per month.
Due now: $1,134

Order this server

Options — where the margin lives

Extra NVMe$6per TB per monthEnterprise NVMe added to the chassis, RAID-1 on request
Extra IPv4$3per address per monthAnnounced from our own space, rDNS included
25 Gbit/s uplink$60per server per monthWhere the standard port is 10 Gbit/s
Private VLAN$25per server per monthLayer-2 between your servers in a region

Why DediGPU

The machine you would buy. Without buying it.

A whole server, its cards, its NVMe and its interconnect, for the cost of running it — one month at a time.

The price is the cost

Every server carries its cost sheet: hardware over 36 months, power at our contract rate, rack, network, spares, operations. Add them up and you have the price. We earn on options, not on the server.
Dedicated, never shared

One tenant per server, one tenant per card. No vGPU, no time-slicing, no noisy neighbour. nvidia-smi shows the whole card because you have the whole card.
NVLink where it counts

SXM nodes ship with the full NVLink domain: 1.8 TB/s per B200, 900 GB/s per H100. Eight-card nodes talk to each other over 3.2 Tbps InfiniBand.
Monthly, no contract

One month at a time, renewed from your balance. No setup fee, no notice period. Prepay 3, 6 or 12 months and pay 3, 6 or 10% less per month: a longer term costs us less to run, and that saving is yours.
Your stack, pre-installed

Ubuntu, Debian, Rocky or Windows Server, then PyTorch, TensorFlow, vLLM, SGLang, TGI, Ollama, ComfyUI, Forge, Blender, Jupyter, Docker, k3s or Slurm on top — with the model you name pulled before you log in. Root SSH and IPMI from the first minute.
$0 egress, always

Public bandwidth is in the network line of the cost sheet. Move datasets and weights as often as you like; we do not meter bytes.

Templates, ready at first boot

Ubuntu · Debian · Rocky · Windows Server
PyTorch 2.7
TensorFlow + JAX
vLLM
SGLang
TGI
Ollama
ComfyUI
Forge
Blender render node
Jupyter Lab
Docker
k3s
Slurm
ports and first commands

How it works

Three steps. No forms, no phone calls, no waiting.

01Create an account

An email address and a password. No documents, no phone call, no waiting for approval. Thirty seconds.
02Top up in crypto

Bitcoin, Monero, Ethereum, USDT, Litecoin, TRON or Solana. Your balance is credited as soon as the network confirms.
03Order the server

Pick the card, the count, the region and a template. The first month is taken from your balance; root SSH is live in under 5 minutes.

Create an account

Thirty seconds, an email address and a password. Nothing else.

What will it run?

The right card for the job, named.

Four workloads, one recommendation each, and the cards that also fit. Every price is the one from the catalogue.

LLM training & pre-training

8-GPU NVLink nodes with 3.2 Tbps InfiniBand between them. Checkpoints land on local NVMe at 14 GB/s.
We recommend B200 180 GB $1,408/mo
Fine-tuning & inference

LoRA, QLoRA and full fine-tunes on 7B to 70B. vLLM and SGLang templates serve the result from the same server.
We recommend H100 SXM 80 GB $1,134/mo
Image & video generation

Flux, SDXL, Wan and HunyuanVideo on GDDR7 cards that were built for exactly this. ComfyUI is one click away.
We recommend RTX 5090 32 GB $163/mo
Rendering & simulation

96 GB of ECC memory for scenes that do not fit anywhere else. Blender, Octane, Redshift and Houdini out of the box.
We recommend RTX PRO 6000 96 GB $493/mo

Regions

4 regions. Cheap, clean power — it is in the cost sheet.

Every card on this page is deployable in every region below, at the same price. Pick by latency: the latency tables.

ReykjavíkEU-North · REY
- 100% geothermal + hydro
- Free-air cooling year-round
- 18 ms London · 40 ms New York
HelsinkiEU-Nordic · HEL
- 100% wind + hydro
- Free-air cooling 10 months a year
- 27 ms Frankfurt · 36 ms London
AmsterdamEU-West · AMS
- 100% wind
- Rear-door liquid cooling
- 7 ms Frankfurt · 12 ms London
SingaporeAPAC · SIN
- Grid + certified RECs
- Direct-to-chip liquid cooling
- 38 ms Tokyo · 45 ms Sydney

Compared

Against dedicated GPU hosts, and against the clouds.

The two places people rent GPUs by the month today, and where a published cost sits between them.

Criterion	DediGPU	Dedicated GPU hosts	GPU clouds
1× H100, per month	$1,134 — the cost sheet	$1,230 – $2,099	$2,460 (median on-demand ×730 h)
1× RTX 4090, per month	$138	$274 – $323	$307 (median on-demand ×730 h)
How the price is set	published cost, zero margin	cost plus margin, unpublished	market rate, changes weekly
Setup fee	$0	$0 – $99	$0
Minimum term	one month, no notice	one to twelve months	none
GPU dedicated to you	yes, whole server	yes	often shared hosts
Egress	$0	metered above a quota	$0.05 – $0.09 per GB
Identity check	none	ID + card, usually	full KYC, credit card
Payment	crypto balance	card, wire, PayPal	card

Dedicated host figures are the published monthly prices of Latitude.sh, GPU-Mart and HostKey for a single-card server on 2 September 2026; cloud figures are the median on-demand rate tracked by getdeploying.com the same day, times 730 hours. What the hourly price hides.

Questions

Straight answers.

What does “sold at cost” actually mean?

Each server’s monthly price is the sum of six published lines: the card and its share of the chassis amortised straight-line over 36 months, electricity at our contract rate of $0.052 per kWh with a PUE of 1.15 and a 70% average load, rack space and cooling, network and IP, a 2% per year reserve for failures, and operations. The sum is rounded up to the next dollar and that is the price. Every line is on the catalogue, per card.

Then how do you make money?

On what is optional: extra NVMe, extra IPv4 addresses, faster uplinks and private VLANs. Those carry a normal margin and are priced on this page. The server itself does not. We fill racks at cost because a full rack of paying tenants is worth more to us than an empty one at a markup.

Is the server really dedicated to me?

Yes. You rent a physical server: its cards, its CPUs, its RAM and its NVMe. No other tenant, no vGPU profiles, no time-slicing. On SXM nodes you get the complete NVLink fabric between your cards. IPMI is yours too.

Why is a longer term cheaper if you sell at cost?

Because it costs us less. The one-month price is the raw cost sheet. A 3, 6 or 12-month term means one provisioning instead of several, no re-imaging between tenants, a firmer power forecast and less idle time between rentals — so the operations and spares lines shrink. We pass that on: 3% off for 3 months, 6% for 6, 10% for 12, on the server and its options alike. The rate is locked for as long as the server renews.

How does term billing work with a crypto balance?

You top up a balance in crypto. When you order, the whole term — 1, 3, 6 or 12 months — is taken from it and the server is provisioned. At the end of the term the same term is taken again automatically, at the locked price. If the balance is short, the server stays online for 72 hours while you top up, is then powered off with its data kept for 14 days, and is wiped after that. You can turn auto-renewal off at any time and the server simply stops at the end of the paid month.

Which payment methods do you accept?

Bitcoin, Monero, Ethereum, USDT (ERC-20 and TRC-20), Litecoin, TRON and Solana. Unused balance can be refunded to the address it came from.

Do I need to verify my identity?

No. An email address and a password open an account, and a crypto top-up activates it. We never ask for documents, a phone number or a card, at any spend level. What we ask for instead is that you respect the acceptable-use policy.

How long does provisioning take, and what is on the image?

A single-card server is live in under 5 minutes; an eight-card node in under fifteen. Pick the OS — Ubuntu 24.04 or 22.04, Debian 12, Rocky Linux 9, Windows Server 2022 on workstation cards — then a template: bare driver, PyTorch, TensorFlow + JAX, vLLM, SGLang, TGI, Ollama, ComfyUI, Forge, a Blender render node, Jupyter Lab, Docker, a k3s node or a Slurm worker. Serving templates pull the model you name before first boot, and the order form checks it fits in memory. You get root over SSH, an IPMI console and an API token.

Can I get multi-node clusters?

Up to eight cards per server are on the catalogue. For 16 to 512 cards we place nodes on the same InfiniBand fabric and hand you a Slurm or Kubernetes cluster; the per-card price is the one printed on this page.

What is the uptime commitment?

99.9% monthly availability per server, measured from our side and published on the status page. When we miss it, the shortfall is credited automatically, ten times over, without a ticket. Failed cards are swapped from local spares; that reserve is one of the six lines of the cost sheet.

Start with one server. Scale to a rack.

Same cost sheet from one card to five hundred. Same account, same balance, same month-by-month billing.

See the servers Create an account

Dedicated GPU servers, sold at cost.

Every server, priced from its cost sheet.

Six lines, added up, rounded up. That is the price.

01Hardware, over 36 months

02Power, at our contract rate

03Rack, network, spares, operations

Where the money comes from, then

Price a server

Options — where the margin lives

The machine you would buy. Without buying it.

The price is the cost

Dedicated, never shared

NVLink where it counts

Monthly, no contract

Your stack, pre-installed

$0 egress, always

Three steps. No forms, no phone calls, no waiting.

01Create an account

02Top up in crypto

03Order the server

The right card for the job, named.

LLM training & pre-training

Fine-tuning & inference

Image & video generation

Rendering & simulation

4 regions. Cheap, clean power — it is in the cost sheet.

Against dedicated GPU hosts, and against the clouds.

Straight answers.

Start with one server. Scale to a rack.