Why the Peaks are higher than Berget
Berget AI sells a per-token API; we rent you the GPU. Per token suits small or bursty volume; a GPU of your own suits steady volume, isolation, custom stacks and fine-tuning. The table, the break-even and the questions below say where the line goes.
Side by side
The comparison.
Berget AI sells a per-token API; we rent you the GPU.
| Per-token API vs your own GPU | Per-token API | Your own GPU |
|---|---|---|
| What you buy | Tokens | The GPU |
| Who else runs on the hardware | A shared pool | Nobody during your term |
| Models | Their shelf | Any |
| Access | API only | Root + API on your box |
| Pricing | Per token, prepaid plan required | Per month, published |
| When it wins | Small or bursty volume | Steady volume, isolation, custom stacks, fine-tuning |
When each wins
Pick by volume and control.
Two short lists: when tokens are the better buy, and when the box is.
A per-token API wins when
- Volume is small or bursty: you pay per token and run nothing.
- A model on its shelf is exactly what you need.
- You want no ops: nothing to keep alive, nothing to configure.
Your own GPU wins when
- Volume is steady: above the break-even, tokens cost more than the machine.
- You need any model: anything on Hugging Face, fine-tuned or not.
- Isolation matters: nobody else on the hardware during your term.
- You run a custom stack: your vLLM settings, embeddings and speech on one box.
- You want root access and a fixed monthly price to budget.
Break-even
Where the line goes.
One dedicated 5090 at 8 000 SEK/month costs about the same as 1.5 billion output tokens at a typical Swedish per-token API’s list price. Below that, tokens are cheaper. Above it, or when you need isolation, a custom stack or fine-tuning, the GPU wins.
List price used: €0.50 per million output tokens, the published price of a 31B open model at a Swedish per-token API on 23 Sep 2026; 8 000 SEK/month converted at about 11 SEK per euro. Ex VAT.
Tokens
A month, priced both ways.
What one box makes in a month at full load, and what that month costs at their price.
| Model | A month at full load, one box | That month at Berget | Obsidian Peaks |
|---|---|---|---|
| Kimi-K3about 20 boxes of memory, in one chassis · €3.00 · €15.00 per M tokens | — | — | on request |
| Qwen3.8-27B-FP8one box, tight fit · €0.40 · €3.00 per M tokens | 1.3 B tokens (~500/s) | about 43 000 SEK | 8 000 SEKtwo boxes, room for context: 16 000 SEK |
| GLM-5.3-Flashsize not public · €0.25 · €0.50 per M tokens | — | — | — |
| Gemma-4-31Bone box, 4-bit · €0.25 · €0.50 per M tokens | 2.6 B tokens (~1 000/s) | about 14 000 SEK | 8 000 SEKtwo boxes, FP8: 16 000 SEK |
| Mistral-Small-3.2-24Bone box · €0.30 · €0.30 per M tokens | 2.6 B tokens (~1 000/s) | about 8 600 SEK | 8 000 SEK |
| KB-Whisper (speech-to-text)one box · €0.12 per hour of audio | 43 200 hours of audio (~60× realtime) | about 56 000 SEK | 8 000 SEK |
| E5 embeddingsone box · €0.03 per M tokens | 51.8 B tokens (~20 000/s) | about 17 000 SEK | 8 000 SEK |
| FLUX.1 (image generation)one box · not on their list | 1.0 M images (~24 a minute) | not on their list | 8 000 SEK |
The month’s output is our estimate for one 5090 kept busy around the clock, output tokens only; measured figures from our own boxes will replace it. Send less than that and their API is cheaper on price alone. Send more, or want what the box adds, and the box wins.
Their prices: api.berget.ai model list, 23 Sep 2026; 11 SEK per euro; bought through a prepaid plan from €25 a month, credits non-refundable, usage past the credit at list price. Kimi-K3 is a trillion-parameter-class model: about 600 GB at 4-bit, twenty boxes’ worth of memory in one chassis; for Kimi, their API is the sane buy.
On top
What the box adds.
What a GPU of your own adds that no token can.
Yours alone
One tenant per GPU. Prompts, files and weights stay on your box; traffic to it passes through our relay in Helsinki (EU).
Any model
Text, vision, speech, image generation, your own fine-tunes: whatever fits in 32 GB.
Agents that never stop
Coding agents and pipelines in Docker, around the clock, no meter. Their coding seats are €150–500 per person a month, rate-limited; one box serves the whole team.
Root access
SSH in, install what you like, keep the state between jobs.
One price
No per-token bill, no rate limit, no prepaid plan. Locked for your term.
Questions
Questions and answers.
- What is the difference between a per-token API and renting a GPU?
- With a per-token API you buy output tokens from a shared pool of hardware running the provider’s shelf of models. With us you rent the GPU itself: nobody else runs on it during your term, you run any model, and you have root access and an API endpoint on your own box.
- When is a per-token API the better choice?
- Small or bursty volume, a model that is on its shelf, and no need for isolation: you pay only for the tokens you use and run nothing yourself.
- When does renting the GPU win?
- Steady volume, a model that is not on the API’s shelf, fine-tuning, a custom stack, or when isolation matters: one machine, one customer, during your term.
- Where is the break-even?
- One dedicated 5090 at 8 000 SEK/month costs about the same as 1.5 billion output tokens at a typical Swedish per-token API’s list price (€0.50 per million output tokens on 23 Sep 2026). Below that, tokens are cheaper; above it, the GPU is.
- Can I get an OpenAI-compatible API from a rented GPU?
- Yes. The dedicated inference box serves the model you pick on an OpenAI-compatible endpoint (vLLM) behind your own key, so anything that speaks the OpenAI API can point at your box.
- How do I pay, and can I stop?
- A monthly price under a signed agreement, monthly or on a 12-month term, for companies and ex VAT, invoiced by Obsidian Peaks Technologies AB.
- Is my data in Sweden?
- Your workload data is: it stays on the box, which we own and run in the Stockholm region, Sweden, 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). The contract, the invoice and the support come from Obsidian Peaks Technologies AB under Swedish law.
