Obsidian Peaks
Pilot

30-day pilot

Thirty days. One RTX 5090. Your numbers decide.

A dedicated RTX 5090 in the Stockholm region with a managed endpoint, OpenAI-compatible or ComfyUI, set up for your workload. One fixed fee for the 30 days, credited against month one if you continue.

The pilot

30 days

4 000 SEK · paid once ex VAT · 5 000 incl. VAT

  • Credited against month one
  • Managed endpoint
  • For companies, ex VAT

Within one working day you get the pilot agreement to sign and the three day-0 questions. The endpoint and keys follow within three working days of your answers.

Paid by card through Stripe. Swedish companies pay 25% VAT on top; an EU company enters its VAT number and is reverse-charged. The receipt comes from Stripe, the agreement from us.

What you get

One machine, one endpoint, one operator.

The fee covers the 30 days and everything it takes to run your workload on the box:

  • One whole machine with an RTX 5090 (32 GB), root access, for you alone
  • Your model, any open model on Hugging Face, served on an OpenAI-compatible endpoint (vLLM) or as a ComfyUI queue
  • A fixed hostname, your own API key, and the ports you need
  • A report at day 26–28: cost per 1 000 outputs, p50 and p95 latency, and quality agreement against your current route
  • A named operator who set the box up and answers within 4 hours on weekdays 08–18
  • Electricity, storage and traffic

The 30 days

Four steps, then a decision.

  1. Day 0–2

    Scope

    One workload: an endpoint, a queue or a fine-tune. We agree the quality bar and export a cost baseline for that step.

  2. Day 3–5

    Stand up

    The box is provisioned, the model loaded, the endpoint and keys issued to you.

  3. Day 6–25

    Shadow run

    The same jobs go to your current route and to the box. We log cost per 1 000 outputs, latency and quality side by side.

  4. Day 26–30

    Decide

    The numbers, side by side. Continue on the monthly box with the fee credited, or walk away with the report and your data.

Day 0

Three answers start the clock.

Reply to the confirmation email with these and a sample of your real inputs and prompts. We provision from there.

  1. 1

    How much

    Outputs, answers or jobs per day today, and what you expect in three months.

  2. 2

    What runs it now

    Which model or API does the work today, and whether it is on credits or paid.

  3. 3

    The bar

    The number you will judge us on: agreement rate against your current output, a latency, a cost. A number, not a description.

After the pilot

Continue, or walk away with the report.

Continue on the same box at the published dedicated price, with the pilot fee credited against month one. Or stop: you keep the report and your data, access closes, and the disk is kept for 14 days and then erased.

Said plainly

What a pilot is not.

  • One box is not high availability. Keep your current route as the fallback while you measure.
  • An RTX 5090 has 32 GB of VRAM. A model that does not fit in that, even quantised, needs a larger GPU.
  • There are no backups. The disk is local and single; copy anything you cannot lose off the box.
  • We hold no SOC 2 or ISO 27001 certificate. The box is our own hardware at our own site in the Stockholm region.
  • Public traffic exits through our server in Helsinki (EU), so the internet sees a Finnish address.
  • For companies only. The pilot needs a company name and an org.nr or EU VAT number at checkout.

Questions

Questions and answers.

What happens after I pay?
Within one working day you get an email from us with the pilot agreement to sign and the three day-0 questions. The endpoint and keys follow within three working days of your answers. Nothing is provisioned before the agreement is signed.
What does “managed endpoint” mean?
We install your model, any open model on Hugging Face, and serve it on an OpenAI-compatible URL (vLLM) or as a ComfyUI queue on your box, behind your own key. Root access stays yours. During the pilot we keep it running and swap the model if the first choice misses your bar.
What does the pilot cost, and is the fee refundable?
4 000 SEK ex VAT, paid once when you start. It is credited in full against the first month if you continue. If we cannot provision your machine, we refund it in full; otherwise it buys the 30 days.
Which machine do I get?
One whole machine with an RTX 5090 (32 GB), an AMD Ryzen 7 9800X3D, 32 GB of RAM or more and about 1.4 TB of NVMe, owned by us and run in the Stockholm region, Sweden. The box is named in the agreement and the handover email.
Can we skip the pilot and start on the monthly box?
Yes. Reserve a machine or book a call, and we start on the monthly agreement directly: 8 000 SEK/month ex VAT, or 7 200 SEK/month on a 12-month term.
Where is my data during the pilot?
Your workload data stays on the box and transits our relay server in Helsinki (EU). Account, billing and support data go to our subprocessors: Stripe, Cloudflare, Resend and Google (US, under the EU Standard Contractual Clauses or the EU–US Data Privacy Framework), Hetzner (Finland) and Telegram (the operator’s support inbox).
Who is the pilot for?
Companies with one workload that runs all month: an inference endpoint, an image or video queue, a fine-tune. Not for multi-GPU jobs, models over 32 GB of VRAM, or bursty jobs under a day.

Rather talk first?

Twenty minutes about your workload, with the person who runs the machines. Then start the pilot, or not.