MindsMachines
EN
Use Case Call

AI Infrastructure Calculator

The AI servers needed by your company

Eight key figures, three paths compared. The calculator shows GPU, power, staff and costs for AI in your own company and discloses every assumption.

  • The calculation runs only in your browser. Nothing is stored until you request the report.
  • Recommended path and key figures straight away, without a form.
  • All assumptions with their sources in the methodology.
Step 1 of 8Size

How many employees does your company have

All locations combined.

employees
5010,000
Continue with Enter

How the calculator works

Every figure in the result follows from your key figures and the assumptions next to them. The assumptions are kept in a single file and maintained with a date.

  1. Active users are employees times share of knowledge work times usage in the first year.
  2. Concurrent requests at peak are active users times the share of concurrent use. The share falls as the number of users grows, and with several shifts the busiest shift counts.
  3. Tokens per day are active users times tokens per application, weighted by the share of users per application.
  4. Peak throughput follows from concurrent requests and the target speed per request.
  5. The model class follows from the applications. Agents and programming need larger models than writing and questions.
  6. The number of GPU is the higher value from memory requirement and throughput, plus one model instance as a reserve for failure.
  7. Connected load is GPU power plus server base load, times the PUE. The PUE depends on the data centre, and without your own server room colocation applies. Power per year at average utilisation.
  8. Staff follows from path, number of users and number of GPU.
  9. Costs over three years are purchase plus three times running costs. Excluding the OneMachine licence and implementation.
  10. The recommendation follows from protection needs, existing server room, free capacity in the IT team and costs over three years.
AssumptionValueSourceType
Usage
Usage in the first year, low20 %ZEW 2026, Bitkom 2026Survey
Usage in the first year, medium40 %UK Government 2025, ZEW 2026Survey
Usage in the first year, high60 %UK Government 2025, ZEW 2026Survey
Concurrent at peak, from 1 active users8.5 %Microsoft Azure LLM Inference Trace 2024Defined value
Concurrent at peak, from 50 active users6 %Microsoft Azure LLM Inference Trace 2024Defined value
Concurrent at peak, from 100 active users4.7 %Microsoft Azure LLM Inference Trace 2024Defined value
Concurrent at peak, from 500 active users3.1 %Microsoft Azure LLM Inference Trace 2024Defined value
Concurrent at peak, from 1,000 active users2.7 %Microsoft Azure LLM Inference Trace 2024Defined value
Several shifts, users in the busiest shift60 %MindsMachines 2026Defined value
Working days per year, one shift205IAB Arbeitszeitrechnung 2025measured
Working days per year, several shifts300IAB Arbeitszeitrechnung 2025Defined value
Tokens per day, Writing and questions20,000, 100 % of usersMicrosoft Azure LLM Inference Trace 2024Defined value
Tokens per day, Document search60,000, 60 % of usersMicrosoft Learn, Azure OpenAI On Your Data 2026per vendor
Tokens per day, Agents and automation200,000, 30 % of usersAnthropic 2025per vendor
Tokens per day, Programming150,000, 10 % of usersZiegler et al. (GitHub) 2022measured
Tokens per day, Coding agents1,000,000, 5 % of usersAnthropic, Claude Code 2026per vendor
Target speed per request30 tokens per secondBrysbaert 2019, MLCommons 2025Defined value
Models
small open model, around 8 billion parameters8 GB, 2,500 tokens per second per GPUMeta Llama 3.1, NVIDIA NIMper vendor
medium open model, 30 to 70 billion parameters70 GB, 250 tokens per second per GPUSemiAnalysis InferenceX 2025measured
large mixture of experts model, over 100 billion parameters400 GB, 250 tokens per second per GPUSemiAnalysis InferenceX 2026measured
Large model from this number of active users1,500MindsMachines 2026Defined value
GPU, power and prices
Memory per GPU141 GBNVIDIA H200per vendor
Power per GPU700 WNVIDIA H200per vendor
Base load per server2,500 WLawrence Berkeley National Laboratory 2024measured
Purchase per GPU including server share40,000 eurosLenovo 2026, Supermicro 2026per vendor
Network and storage, one-time50,000 eurosDELTA Computer 2026, Geizhals 2026per vendor
Maintenance per year10 % of purchaseLenovo 2026Defined value
PUE, own data centre1.5EnEfG § 11Legal standard
PUE, colocation1.4Bitkom 2025Forecast
Average utilisation45 %Patel et al. 2024, Wilkins et al. 2026Defined value
Industrial electricity price21 cents per kWhEurostat 2026measured
Path A, per 1 million tokens3 eurosAzure, AWS, Google, Mistral 2026per vendor
Path B, per GPU hour2.75 eurosOpen Telekom Cloud 2026per vendor
Staff
Path A, positions0.3 plus 0.2 per 1,000 users plus 0 per GPUMindsMachines 2026Defined value
Path B, positions1 plus 0.2 per 1,000 users plus 0.03 per GPUMindsMachines 2026Defined value
Path C, positions2 plus 0.2 per 1,000 users plus 0.06 per GPUMindsMachines 2026Defined value
Cost per full-time position93,000 eurosDestatis 2025, Hessisches Statistisches Landesamt 2025measured
Freely available share of the IT team10 %MindsMachines 2026Defined value

Assumptions as of 26.10.2026.

OneMachine

A tailored solution on each of the three paths

The tailored solution runs in your tenant or data centre, with your roles, with logging and without vendor access to content. Updates, monitoring and security patches are included in the licence.

The documented operating path is Microsoft Azure, other environments are supported. You clarify which path fits your company in the Architecture and Security Call.

Path A

Models via EU API

OneMachine in your environment, models hosted in the EU with a data processing agreement.

Path B

GPU in a German cloud

Dedicated GPU from a provider in Germany, open models.

Path C

GPU in your own data centre

Your own GPU servers on premises, full control over hardware and models.

Your next step

Go through your result with an architect

In the Architecture and Security Call, you check your calculation result against your environment. The outcome is clarity on operation in your own tenant, access model, roles and logging.

  • Architecture and Security Call, 30 minutes, online
  • Operation in your own tenant or data centre
  • Access model, roles and logging
Book an Architecture and Security Call

The platform

OneMachine runs in your tenant

Data and logs stay in your environment, on paths B and C the models as well. MindsMachines has no access to content.

Johannes Heister at Future Tech Fest 2026
Johannes HeisterManaging Director, MindsMachines AI GmbH
+49 211 86942524 johannes@mindsmachines.ai
AI Infrastructure Calculator | MindsMachines