PRIVATE AI · ON-PREMISE LLM · GDPR

Private AI like ChatGPT, on your own server

Private AI is a language model on your own or a dedicated EU server that answers from your documents and cites the source. A free test shows answers from 50 of your documents within 5 working days; a 14-day pilot starts at €3,200 excl. VAT.

Get the free 50-document testUnder NDA · results within 5 working days
Quick and free:AI cost calculatorReferencesOlingo Medical · Rowan Legal

Leave your number, I'll call you

Krystof Olik · founder and CEO

Same working day (CET), otherwise the next working morning.

Or email me directly at info@ollsoft.ai

We use your details only for this request. Privacy
ChatGPT Business
$20 per user per month, billed annually (OpenAI, 23 Oct 2026)
Open model gpt-oss-120b
runs on a single 80 GB GPU (OpenAI model card)
NVIDIA RTX PRO 6000
96 GB GPU memory, 600 W max (NVIDIA)
AI literacy
measures required since 2 Feb 2025 (Art. 4 AI Act)
Official sourcesFull GDPR text (EUR-Lex)Full AI Act text (EUR-Lex)EU-US Data Privacy Framework decision
PricingTest: freePilot: from €3,200Operation: from €490 per monthFull offer

Why Ollsoft?

We are engineers who build and run private AI ourselves, from the GPU server to the answer from your documents.

  • we design, wire up and deliver the GPU server turnkey
  • without public cloud, our AI already runs on the servers of two university hospitals and of the law firm Rowan Legal
  • we can fine-tune the model on your data; the adapter is yours
An Ollsoft engineer connecting a power cable in an open server packed with GeForce RTX graphics cards

What does private AI do?

Private AI is a language model with document search (RAG) that runs on a server in your building or on a dedicated EU server. It answers questions from your documents: it first finds the relevant passages, then the language model writes the answer from them. Model and search index run on hardware you control.

  • questions on internal policies, contracts and technical documentation
  • summaries of long files and comparison of document versions
  • draft replies to customers from your knowledge base
  • data that must not go into public AI: health data, personal data, trade secrets

Why is cloud ChatGPT for business not enough?

It is fine for general text, but every request goes to the provider, personal data and trade secrets included.

  • the provider acts as your processor under Art. 28 GDPR
  • OpenAI states it does not train on Business data by default; the question is who can access the data and under which law
  • transfers to the US rely on the Commission adequacy decision of 10 July 2023 (EU-US Data Privacy Framework)
  • the EU Court of Justice struck down both predecessors: Safe Harbor in 2015, Privacy Shield in 2020
  • under the US CLOUD Act a US provider must hand over data when a US court or authority orders it, wherever the server stands
  • on your own server these questions disappear: the data stays in your network

What hardware do I need for how many people?

Model size and daily text volume decide, not headcount.

  • a model takes about one byte of memory per parameter at 8-bit precision and half that at 4-bit, plus room for context: gpt-oss-120b fits a single 80 GB GPU, gpt-oss-20b fits in 16 GB
  • up to 20 people and pilots: compact AI computer with 128 GB unified memory (NVIDIA DGX Spark, Mac Studio), from €6,000
  • up to about 150 people: server with one NVIDIA RTX PRO 6000 Blackwell (96 GB), around €24,000
  • up to about 300 people: two such GPUs, around €39,000
  • hundreds of users, round-the-clock agents or models above 200B parameters: 4 to 8 GPUs, from €69,000
  • our estimates for October 2026, excl. VAT, sized conservatively at 20 requests per person a day with a medium model; the pilot load test measures real capacity

Which models do you deploy?

Open models whose licence allows business use on your own hardware.

  • for example gpt-oss from OpenAI, or Mistral Small 4 and Mistral Large 3 from the French company Mistral AI, all under Apache 2.0
  • we choose by your documents and languages
  • the 50-document test compares two or three candidates
  • if an off-the-shelf model is not enough, we fine-tune it on your data with a LoRA adapter
  • the adapter and all its updates are yours

How does the cost compare with the cloud?

For 100 people at 20 requests a day with a medium model, your own server pays back against ChatGPT Business in month 25 and, for plain chat, never against the cloud API.

  • cloud API (gpt-6.1-sol with EU processing): about €450 a month
  • ChatGPT Business: about €1,790 a month
  • your own server: about €27,200 one-off including the pilot, plus €670 a month for operation and power
  • it makes sense when the data must not reach a provider, when it replaces per-user licences, or when agents process documents around the clock
  • run your own numbers in our AI cost calculator

OFFER

From a free test to your own AI server

Step 1 · Start here

50-document test in 5 working days

You see how the AI answers from your documents and what hardware you would need.

Free

What you get

  • 30 minutes with an engineer to pick documents and 20 test questions
  • you send us the documents; isolated EU test server under NDA, no access to your systems
  • answers with links to the source document
  • two or three open models compared
  • initial hardware recommendation and a fixed pilot price
  • documents deleted after the test
Get the free 50-document test

Step 2

14-day pilot

The AI runs on your documents and your team uses it in daily work.

from €3,200excl. VAT

What you get

  • review of your infrastructure and document sources, installation on your server or a dedicated EU server
  • file shares or SharePoint connected
  • web interface for up to 20 users
  • answers evaluated against your question set
  • load test and final hardware sizing
Request a quote

Step 3

Your own AI server in operation

The AI runs permanently on your premises, current and monitored.

from €490 per monthexcl. VAT + hardware

What you get

  • server delivered or installed on your hardware
  • model updates and security patches
  • monitoring, backups and new documents indexed
  • quarterly answer check against your question set
  • support on working days
Request a quote

AT NO EXTRA COST

Included with the pilot

  • material for your data protection officer: data flow description and a basis for the impact assessment (DPIA)
  • an employee policy on what may go into which AI
  • one-hour user training as evidence of AI literacy measures (Art. 4 AI Act)

GUARANTEE

No payment for the pilot if it misses the agreed criteria

Before the pilot we agree in writing on test questions and what a correct answer is. If the AI does not meet them at the end, you do not pay for the pilot.

Get the free 50-document test

Under NDA · results within 5 working days

Leave your number, I'll call you

Krystof Olik · founder and CEO

Same working day (CET), otherwise the next working morning.

Or email me directly at info@ollsoft.ai

We use your details only for this request. Privacy

PROCESS

How it works

  1. 01

    We call you today

    A short call: which documents, how many users, where data may live. No access to your systems.

  2. 02

    50-document test

    You send the documents. Answers to your questions and a hardware proposal within 5 working days.

  3. 03

    14-day pilot

    Infrastructure review, installation, your team uses it. We measure quality and load.

  4. 04

    Operation

    Server on your premises, updates and monitoring by us.

COMPARISON

Chatbot subscription, cloud API or your own server?

Chatbot subscriptionCloud APIOwn server
Where data is processedChatbot subscription: at the provider; EU data residency only on EnterpriseCloud API: at the provider; EU region at +10%, system data may still be processed outsideOwn server: inside your network, even offline
PriceChatbot subscription: per user, e.g. $20 a monthCloud API: per token, e.g. $2 per 1M input and $10 per 1M output tokensOwn server: hardware one-off + fixed monthly fee
Model choiceChatbot subscription: provider modelsCloud API: provider modelsOwn server: open models, tunable on your data

Category comparison. Price examples: ChatGPT Business and gpt-6.1-sol from OpenAI price pages as of 23 Oct 2026.

REFERENCES

Where our AI already runs without public cloud

Olingo Medical

More than 10 hospitals and clinics use its AI for medical documentation (speech recognition and OCR). In two university hospitals it runs entirely on their own GPU servers with no cloud services, and patient data never leaves the hospital network.

Case study

Rowan Legal

AI agents on a law firm’s own servers: search in case files and case law, no external API calls.

Case study

CONTACT

See AI on your own documents within 5 working days

Leave your number. We call you today, agree on 50 documents and test questions, and show you the answers within 5 working days. Free, no commitment, no access to your systems.

Leave your number, I'll call you

Krystof Olik · founder and CEO

Same working day (CET), otherwise the next working morning.

Or email me directly at info@ollsoft.ai

We use your details only for this request. Privacy

FAQ

What clients ask us

On general questions the best cloud models usually lead. For answers from your documents, retrieval matters most. The free 50-document test shows whether an open model is enough before you pay anything.

No. A compact AI computer fits on a desk; a GPU server belongs in a rack or an air-conditioned room. If you do not want a server on site, we run it on a dedicated EU server that you do not share with other customers.

Your users, and we for server administration under a data processing agreement (Art. 28 GDPR). Run air-gapped, the server has no outside connection and we install updates on site.

The test is free. The pilot starts at €3,200 excl. VAT and includes the review of your infrastructure and the installation. Operation from €490 a month excl. VAT; hardware depends on volume, from about €6,000. A fixed amount, no tokens.

Your own server removes transfers to providers and third countries. Legal basis, notices and records of processing stay with you; we provide material for your DPO. We do not give legal advice.

Add GPUs or swap the model. Documents, search index and interface stay; your users have nothing new to learn.

MORE SERVICES

Often needed alongside

AI cost calculatorWhat does company AI cost via cloud API, per-user subscription or your own server? OpenAI prices as of 23 Oct 2026, payback in months, open formula.MoreDocument anonymizationAnonymize contracts and case files in Word, Excel, PDF and scans: true PDF redaction or labels like [PERSON_1], key kept by you. 14 days free, then on-premise.MoreAI meeting minutesAI meeting minutes from a recording: transcript with speakers, decisions, action items with deadlines. Processed on your server or in the EU. 14 days free.More

Sources

Updated: