AI inference from Germany

Use AI models without your own GPU – billed per token

Language models for your application – no hardware, no base fee, no minimum term. You only pay for what is actually processed: from €0.12 per 1M tokens. Processing runs in our data center, optionally with a location guarantee.

Request access View pricing
Price
from €0.12 / 1M tokens
Location
Germany • own data center
Data protection
GDPR • DPA available
Processing exclusively in data centers in Germany, no transfer to third countries. With the itp-mini-local tier, guaranteed to stay in our own data center.

What you get with us

Language models, billing and data protection from a single source – without a detour via hyperscalers.

Use cases Send an inquiry
Four model tiers
Selectable per use case, switchable at any time.
  • An affordable tier for bulk tasks, a stronger one for complex requests
  • A dedicated tier with a location guarantee for sensitive data
  • Embeddings for search, RAG and classification
Ready to go
No setup, no operations, no maintenance.
  • Access via API – integrates directly into your application
  • No base fee and no contract term
Billing per token
For inference, without running your own GPU.
  • You pay for what is actually processed
  • Input and output tokens itemized separately
Data protection & operations
The difference from the big providers.
  • Processing exclusively in Germany – optionally guaranteed to stay in our own data center
  • Data processing agreement under Art. 28 GDPR on request
  • Your inputs are not analysed and not used to train models – neither by us nor by third parties
  • Optionally operated with our managed services

Usage-based billing

You pay for processed tokens – no hardware, no base fee, no minimum term.

How billing works
We count what the model actually processes.
  • Input tokens (your text sent to the model) and output tokens (the response) are counted separately
  • Prepaid: top up your balance and reload flexibly at any time
  • No base fee, no minimum purchase, no contract term
  • Usage visible at any time – per tier and per period
  • With a fixed monthly commitment the price drops – package discounts
A tier per use case
You're not locked into one model.
  • An affordable tier for bulk tasks, a stronger tier for complex requests
  • Location guarantee selectable wherever it's needed
  • Switch at any time

Why not your own server?

A GPU only pays off once it is permanently utilized. As long as demand fluctuates or is still being tested, usage-based billing means you only pay for actual usage – no purchase, no operations, no idle costs. Once you have a steady base load, you can move to your own hardware: dedicated GPU systems are planned from Q2 2027.

Models & token pricing

Four tiers – from the affordable entry level to the location guarantee. You choose per use case.

Billing models Request access
Tier Best for Input Output Example 70/30 Processing
itp-mini Standard tasks, high volume, most affordable entry €0.14 €0.30 €0.19 Our data center, fallback capacity in Germany
itp-mini-local on request Like itp-mini, but with a location guarantee €0.18 €0.38 €0.24 Exclusively our data center
itp-standard More demanding tasks, longer contexts €0.35 €0.75 €0.47 Our data center, fallback capacity in Germany
itp-embed Embeddings for search, RAG and classification €0.12 – €0.12 Data center in Germany

All prices per 1M tokens, incl. 19% VAT. Input and output are counted separately. The Example 70/30 column shows the blended price at a typical ratio of 70% input to 30% output – for orientation only; billing is always based on actual usage. Embeddings have no output tokens.

Package discounts

The selected package lowers the token price on all tiers.

Prepaid • €0/monthList price
Starter • €19/monthList price
Business • €79/month−10%
Enterprise • €399/month−20%

What is a token?

A token is the smallest processing unit of a language model. As a rule of thumb, one token corresponds to roughly four characters of text – a full A4 page comes to roughly 600 to 800 tokens. One million tokens therefore equals about 1,300 to 1,600 pages.

Packages

Token prices are the same in every package – the package determines keys, connectivity, support and the discount applied to them.

Prepaid

Self-service entry
€0/ month
Token price: list price
  • 1 API key
  • Top-ups from €20
  • Self-service
  • All four model tiers
Top up balance
No base fee and no term.

Starter

For first team projects
€19/ month
Token price: list price
  • 5 API keys with budgets
  • Dashboard & monthly invoice
  • Data processing agreement included
  • Ticket support by the next business day
  • Minimum purchase of €19 token credit per month
Choose Starter
Billed monthly via the customer area.

Business

Recommended
For production use
€79/ month
Token price: −10% on all tiers
  • 25 API keys
  • Private connectivity from our network
  • Cost center reports
  • Priority support within 4 h
  • Minimum purchase of €800 token credit per month
Choose Business
The discount applies to all four tiers.

Enterprise

For critical applications
€399/ month
Token price: −20% on all tiers
  • Unlimited API keys
  • Custom aliases
  • Fallback routing
  • Dedicated contact person
  • SLA up to backend level
  • Minimum purchase of €1,995 token credit per month
Discuss Enterprise
We define scope and SLA together.
All four model tiers are available in every package (itp-mini-local on request) – the packages only differ in keys, connectivity, support and discount. All package prices incl. 19% VAT. The discount applies to the token price of all tiers – example: in the Business package, itp-mini costs €0.126 instead of €0.14 per 1M input tokens.
Planned for Q2 2027

Our own GPU hardware – in preparation

Today we bill exclusively per token. For customers with a steady base load, we are building two additional models.

Dedicated GPU systems from Q2 2027
Your own card, your own stack, predictable fixed costs.
  • Exclusive GPU, no sharing
  • Root access, free choice of driver, CUDA and framework versions
  • Combinable with NVMe storage and 10G connectivity
GPU on demand from Q2 2027
Compute time by the hour, for training runs.
  • Billed by actual runtime
  • Not tied to a specific card

Register your interest

Tell us which models you want to run and what base load you expect – we'll get in touch as soon as capacity is available. Until then, you can cover your needs with per-token billing.

Register interest

Typical use cases

What our customers use GPU capacity for.

Training & fine-tuning
Your own models on your own data.
Inference in production
Assistants, classification, text processing – with predictable latency.
Document processing with data protection requirements
Contracts, HR records, patient data – processing stays in-house.
Rendering & HPC
Simulation, batch jobs, image processing.

Frequently asked questions

Still unsure what fits your project? Drop us a line – we'll give you a concrete answer.

Ask a question

A token is the smallest processing unit of a language model – roughly speaking, one token corresponds to about four characters of text. You are billed for what is actually processed, split into input tokens (your text sent to the model) and output tokens (the response). You don't run your own GPU and don't pay for base load.

On a prepaid basis: you top up your balance and use it for all four model tiers – without separate contracts or bookings per model. Only the tokens actually processed are deducted. You can top up at any time and in any amount. There is no base fee, no minimum purchase and no contract term – if no credit is used, no costs are incurred.

No. Credit you've topped up remains valid – there is no deadline by which it must be used and no expiry at the end of the year. So you don't unexpectedly run out, we notify you by email as soon as your balance gets low – leaving enough time to top up before requests start failing.

Currently we offer per-token billing only. Dedicated GPU systems and hourly capacity are planned for Q2 2027. Once a GPU is permanently utilized, your own hardware is usually cheaper than usage-based billing – plus you get full control over drivers, frameworks and model versions. Until then, you can cover your needs with per-token billing.

Processing takes place exclusively in data centers in Germany – there is no transfer to third countries. Your inputs are not analysed and not used to train models – neither by us nor by third parties. If you additionally need the assurance that your data uses only our own data center, choose the itp-mini-local tier. We conclude a data processing agreement under Art. 28 GDPR on request, including information on the subprocessors used.

Yes. On dedicated systems you have root access and can freely choose drivers, CUDA version and frameworks – as well as your own or open-source models. On request, we handle setup and operation as part of our managed services.

Yes, and they can also be combined. A typical path: start with per-token billing to understand your demand, later switch to a dedicated GPU once the base load is established – supplemented by on-demand capacity for training phases.
Want to work through your GPU needs together?

Tell us briefly which models you use and what utilization looks like – we'll get back to you with a recommendation on which billing model is the most cost-effective for you.

Request without obligation Dedicated servers