Apply

Managed GPU ownership · Iceland

Own the compute.
We run the giant.

AI demand is outrunning compute supply, and most capital rents that compute — owning nothing at the end of it. Ymir sells you an 8-GPU NVIDIA server, titled and serialised to you, then installs, hosts, operates and rents it on 100%-renewable Icelandic power.

In continuous operation
Since 0
Measured occupancy
0%
Fleet in production
0H200 GPUs
Renewable power
0%
What you own

A machine,
not a position.

Not a token. Not a share of a pool. Not a claim on somebody else's hardware. A serialised NVIDIA server with your name on the title, in a rack you can walk up to.

01 — Title

The hardware is legally yours

Title, serial numbers and documented transfer. Ymir operates the machine under a separate agreement — the ownership and the service are two distinct things, and you can end one without losing the other.

02 — Liquidity

Sellable whenever you decide

Enterprise GPUs trade on an active global secondary market. Title can transfer without the machine leaving the facility, or it ships to your buyer. Continuing with Ymir stays a monthly choice, never a lock-in.

03 — Residual

Whatever it is worth is yours

Residual value belongs to the owner, on top of every month the machine has already earned. Resale happens at prevailing market prices, which move with demand and generation cycles.

How it works

Four stages.
One accountable operator.

From procurement to the statement in your inbox each month, one company is answerable for the machine — and it is the company that installed it.

NVIDIA HGX 8-GPU server
01 — Acquire

You buy the server

Ymir procures at scale and invoices you directly. Title and serial numbers transfer to you on delivery.

Server received at the Iceland facility
02 — Deploy

We ship and install

Import, transport, racking, power, cooling and network commissioning, carried out by our own crews.

Fleet monitoring in the operations centre
03 — Operate

We run and rent it

Round-the-clock monitoring, maintenance, and continuous routing of your capacity to the best-paying channel.

NVIDIA H200 accelerator
04 — Report

You are paid monthly

Your machine's rental revenue less your machine's costs, with a machine-level statement behind every line.

Why Iceland

The largest running cost of AI,
reduced to a rounding error.

Electricity decides whether AI hardware earns or bleeds. Iceland runs on geothermal and hydro at industrial rates, and the climate does much of the cooling for nothing.

The operating campus on the Icelandic lava field, geothermal plant beyond
Operating site · Iceland
≈2%
Electricity, as a share of rental revenue
  • Grid100% renewable — geothermal & hydro
  • CoolingFree-air assisted, year-round ambient
  • ConnectivityRedundant subsea fibre to Europe
  • JurisdictionEEA · stable regulatory regime
The placement engine

Capacity is never parked
on one platform.

Most owned-hardware programmes list the machine on a single marketplace and hope. We route each machine's GPUs continuously to whichever channel pays best at that moment.

Structured cabling on a live rack
Live capacity routing

It is the capability we built in live operation, and the reason the fleet realises materially above-median rates at high occupancy rather than repeating a beginner's ramp.

Ymir realised $2.91 / GPU-h
Market median $1.94 / GPU-h
  • Measured occupancy88.4% — rolling window
  • Revenue streamsGPU rental · storage · bandwidth
  • Direct tenancyChannel objective — not yet contracted
We would rather under-claim and over-prove.
In commercial operation since early 2025
Hardware

Three builds.
One operation behind them.

Every configuration runs in the same halls, on the same power contract, under the same discipline. Only one is the machine our live fleet is built from, and we label it plainly.

Entry
8 × H100
$300,000 USD
  • GenerationNVIDIA Hopper
  • DeploymentIceland, Ymir-operated
  • Data statusIndicative

The market's most competitive segment — thinner margins, more exposed to incoming Blackwell supply.

Live fleet Flagship
8 × H200
$400,000 USD
  • Memory141 GB HBM3e per GPU
  • DeploymentIceland, Ymir-operated
  • Data statusMeasured — this exact build

The configuration our operating fleet is made of. Every operational figure we publish is measured on these machines.

Next generation
8 × B200
$600,000 USD
  • GenerationNVIDIA Blackwell
  • DeploymentIceland, Ymir-operated
  • Data statusIndicative — allocation-dependent

Densifies income on the same racks and fibre. Subject to allocation and confirmed lead time.

  • Management & hosting$2,000 per server, per month
  • Electricity & networkPassed through at cost
  • Minimum allocationOne server

Prices are per server, excluding VAT and duties where applicable. Unit economics and the full operating record are set out in the memorandum, released to qualified applicants.

Operations

What “fully managed”
actually means.

No part of the operation is handed to a party you have no contract with. Ymir crews receive the hardware, rack it, power it, cool it, watch it, and sell its capacity.

A GPU server being installed into the rack

Installation

Machines are received, inventoried against serial numbers, racked and commissioned by Ymir staff.

Power and cooling plant

Power & cooling

Redundant power paths and a plant engineered for sustained accelerated-compute load, not office IT.

Liquid cooling manifolds at the rack

Thermal management

Liquid cooling where density demands it, so hardware runs inside spec and holds its resale value.

Structured cabling and optics

Network

Diverse carrier entries and redundant fabric, with low-latency routes to European demand.

Telemetry dashboards in the operations centre

Monitoring

Continuous telemetry on power, thermals and utilisation, with alerting and on-site remote hands.

NVIDIA H200 accelerator

Monetisation

Capacity placed across rental channels and, where terms are better, contracted directly to tenants.

The operating facility at dusk
The facility · Iceland
Evidence, not projection

Measured, not modelled.

There is a fleet, it has been renting capacity commercially since early 2025, and every operational number we publish is measured on it — occupancy from sampled logs, costs from source invoices, rates from settled rentals. Where a figure is indicative rather than measured, we say so on the line where it appears.

  • Fleet configuration8 × H200 per server
  • Commercial operationContinuous, since early 2025
  • Cost basisSource invoices · ECB reference FX
  • Occupancy basisSampled logs, rolling window
  • Owner reportingMonthly, per machine
Risk

Ownership carries
real risk.

Four variables decide the outcome. We would rather you weigh them now than discover them later — and we walk the downside through before anyone commits.

01

Utilisation

Income depends on how much of your capacity is rented. Occupancy is measured continuously and has been high, but it is not contractually guaranteed.

02

Market pricing

GPU rental rates move with supply and demand, and fell sharply from peak in the previous generation. Rates can fall as new capacity enters.

03

Technology cycles

Accelerator generations succeed one another. Competitiveness and residual value change across the holding period.

04

Operations

Hardware faults, power events and connectivity incidents occur. Redundancy and maintenance mitigate them; they do not eliminate them.

Apply

The full numbers are one conversation away.

Allocations start at one server. We review every application before releasing the memorandum — the economics, the operating record and the downside scenarios are only meaningful with a conversation around them.

  • 01You apply. We review and respond within two business days.
  • 02Qualification call — objectives, horizon, jurisdiction.
  • 03Memorandum and full operating record released.
  • 04Numbers walkthrough, contracts, and a site visit if useful.

Ymir Compute does not accept funds through this website. Applications are reviewed manually; we do not share applicant information with third parties.