Skip to Content
← All articles

LLM in a company: cloud, your own account or an AI server?

Four ways to run it, prices from official price lists and the question of where your data may go.

Team Lumenax Nov 17, 2026 · 10 min read
LLM in a company: cloud, your own account or an AI server?
November 17, 2026 by
LLM in a company: cloud, your own account or an AI server?

You can run an LLM (large language model) in a firm in four ways: through a business plan of a cloud service, through your own account and the provider's API, through a model on servers in the EU, or on an AI server in your own network. Start with the question of where your data may go. Price and quality come after that. Claude Team costs 20 USD per user per month on annual billing, the API is paid per token, and your own server costs from 5,999 EUR one-off. We verified the prices on the providers' pages on 25 October 2026.

Key takeaways
  • Start with the question of where the data may go, not with the price of a model.
  • Claude Team costs 20 USD per user per month on annual billing, 600 USD for 30 people.
  • Anthropic for commercial products and OpenAI for its API don't train models on your data by default; personal plans differ.
  • With a provider you need a written processing contract; at Anthropic it is part of the terms.

An LLM is a program that produces further text from text. ChatGPT and Claude are programs built on such models. You can also call the same models from your own systems or run them yourself.

Four ways to run it and what differs between them

WayWhere the data sitsWho trains on the dataPricePreparation
Business plan of a cloud (e.g. Claude Team)With the providerNobody by default, see belowClaude Team from 20 USD per user per month, 25 USD on monthly billingHours: create accounts and write the rules
Your own account and APIWith the provider; with OpenAI you can choose a European endpointNobody by defaultPer token; Claude Sonnet 5.5 is 2 USD input and 10 USD output per million tokensDays: connect programs and watch the budget
A model on infrastructure in the EU (at Lumenax)On Lumenax's servers in the EUAccording to Lumenax, nobodyFrom 4,900 CZK a monthDays: pick a model and connect it
An AI server in your networkIn the firm's networkNobodyFrom 5,999 EUR one-off plus installation, plus running costsLongest: get the hardware, install, maintain

The Lumenax prices come from its xOffice page, where the firm marks them as starting prices. It confirms the final amount after an analysis. The Preparation column is our estimate, not a figure from the sources. The cloud prices are in dollars and the providers' pages show no date, so we state the date we read them.

Who trains on your data and where the data sits

On its page dated 18 August 2026, Anthropic writes that by default it doesn't use inputs and outputs from commercial products (Claude for business, the API) to train its models. Two exceptions: feedback that someone sends with the thumbs up and down buttons, where the whole conversation is stored for up to 5 years and may be used for training, and explicit consent. Personal plans (Free, Pro, Max) follow different rules. Employees with a personal account are therefore not covered by the business protection.

For its API, OpenAI writes that data sent to the API isn't used to train or improve models unless you explicitly opt in. Abuse-monitoring logs are kept for up to 30 days unless the law requires longer. For Europe (EEA and Switzerland) there is the endpoint eu.api.openai.com with storage and processing in the region, but it requires one of the options for how data is retained, such as Zero Data Retention, which OpenAI approves in advance. Processing in the region costs 10% more for models released since 5 March 2026.

What we haven't verified: the terms of the ChatGPT business plans. OpenAI's pages about them return a 403 error, so we write nothing about them that couldn't be looked up. Check them directly.

Not verified. We could not get the terms of the ChatGPT business plans from OpenAI's public pages.
20 USDClaude Team per user per month, annual billing
600 USDClaude Team for 30 users a month, annual billing
from 4,900 CZKa model on infrastructure in the EU per month (Lumenax)
from 5,999 EURan AI server in your own network, one-off plus installation

What an LLM costs in a firm

With 30 users, Claude Team comes to 600 USD a month on annual billing and 750 USD on monthly billing. The page says Team is for 2 to 150 people. We have no source to state the price of ChatGPT Business.

With the API you pay per token. A query with 2,000 input tokens and 500 output tokens costs 0.009 USD on Claude Sonnet 5.5 at 2 and 10 USD per million tokens, so 10,000 such queries cost 90 USD. In OpenAI's price list, text models range from 0.10 USD input and 0.50 USD output per million tokens (gpt-6-luna) to 30 and 180 USD (gpt-5.5-pro and gpt-5.4-pro) at short context. Use the dearer model only where it pays off.

Your own server has the opposite cost structure: a big first payment and then running without per-token fees. How much memory each model needs is described in the article on Ollama.

The contract with the processor

If personal data goes into the model, the provider is a processor. The European Data Protection Board (EDPB) writes in its Guidelines 07/2020 of 7 July 2021 that the work of a processor must be governed by a contract between the controller and the processor under Article 28(3) GDPR, and that it must be in writing, including electronically.

At Anthropic, the data processing addendum (DPA) including the standard contractual clauses is automatically part of the commercial terms, so you accept it along with them. At OpenAI we didn't find a DPA on the pages we opened. That doesn't mean it doesn't exist. Ask for it and check the list of sub-processors.

In Opinion 28/2024 of 17 December 2024, the EDPB stated that models trained on personal data can't always be considered anonymous. Whether a model in a European region meets your rules will be shown by the contract, not by the provider's marketing. We go through the GDPR links in the article on the technical side of GDPR.

Where may your data go?Simplified from the list that follows.
  1. To the provider's cloudYou want a quick start
    A business plan, business accounts instead of personal ones
  2. Only to a server in the EUYou don't want to maintain hardware
    A model on infrastructure in the EU, from 4,900 CZK a month
  3. Nowhere outside your networkContracts, payroll, health data, production
    An AI server, from 5,999 EUR one-off
  4. You don't know what you needTry 20 real tasks in each variant

How to decide

  1. The data may go to the cloud and you want a quick start: a business plan. You get working with a model within hours, and business accounts instead of personal ones.
  2. You want AI inside your own programs: your own account and the API. You pay for actual use and set a monthly limit in the system.
  3. You want data in the EU and don't want to maintain hardware: a model on infrastructure in the EU, with a written contract.
  4. The data must not leave the network (contracts, payroll, health data, production): an AI server in your own network.

Many firms end up using both. Routine queries go to the cloud, sensitive ones to a local model. Then you need a rule on which data may go where, and enforce it technically, not just with a handbook.

Steel cabinets housing servers, network devices and cables
Your own server means a big first payment and then running without per-token fees. Photo: Brett Sayles

Which model is good enough

We found no comparison of the quality of cloud and local models for Czech in open sources. So we recommend a test: take 20 real tasks from the firm and run them in each variant. You learn more from your own results than from a leaderboard. Repeat the test in six months, because models and prices change.

What the AI Act says

On a page updated on 19 October 2026, the European Commission states that the bans and the obligations around AI literacy have applied since 2 February 2025, the rules for general-purpose AI models since 2 August 2025, and that the AI Act began to apply on 2 August 2026 with some exceptions. The rules for high-risk systems under Annex III were pushed back to 2 December 2027 after an amendment (the AI Omnibus, which has been in force since 27 July 2026). The rest is summarised on the AI Act page.

What the AI assistant in xOffice does

In its xOffice ERP, Lumenax offers an optional AI assistant, and the page lists exactly the routes from this article: none, your own account with a provider, a model on Lumenax's infrastructure in the EU, or an AI server in your network. According to the page, the assistant sees only what the asking user may see, every change to the data has to be confirmed by a person, and a monthly budget applies per user and for the firm.

Common questions

Can employees use ChatGPT with company data?

Only if the firm sets rules for it. Anthropic states that it doesn't use data to train models for commercial products and the API, not for personal plans. The simplest thing is to give people a business account and ban personal ones.

Do I need a data processing agreement with the provider?

Yes, if personal data goes into the model. The EDPB writes that such a contract must be in writing. At Anthropic the DPA is part of the commercial terms. At OpenAI, ask for it.

How much does your own AI server cost?

For xOffice from 5,999 EUR one-off plus installation, plus running costs. How much memory each model needs is in the article on Ollama.

Is a model in the EU the same as a local server?

No. The data leaves your network for the provider's server, even if it stays in the EU. The contract, the ban on training models on your data and the provider's security rules decide.

Which model is best for a firm?

None universally. Decide by the sensitivity of the data, the volume of queries and a test on your own tasks.

Sources

  • Claude, price list, viewed 25 Oct 2026: https://claude.com/pricing
  • OpenAI, API price list, viewed 25 Oct 2026: https://developers.openai.com/api/docs/pricing
  • OpenAI, Your data, viewed 25 Oct 2026: https://developers.openai.com/api/docs/guides/your-data
  • Anthropic, training models on commercial product data, page of 18 Aug 2026: https://privacy.claude.com/en/articles/7996868
  • Anthropic, DPA, page of 16 Mar 2026: https://privacy.claude.com/en/articles/7996862
  • EDPB, Guidelines 07/2020 on the concepts of controller and processor, version 2.1, 7 Jul 2021: https://www.edpb.europa.eu/system/files/2023-10/edpb_guidelines_202007_controllerprocessor_final_en.pdf
  • EDPB, Opinion 28/2024, 17 Dec 2024: https://www.edpb.europa.eu/system/files/documents/2024-12/edpb_opinion_202428_ai-models_en.pdf
  • European Commission, Regulatory framework on AI, updated 19 Oct 2026: https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
  • Lumenax, xOffice ERP system, verified on the page 12 Oct 2026, viewed 25 Oct 2026: https://www.lumenax.cz/erp-system
Team Lumenax

We write about security, AI and software for Czech companies with 10 to 100 people.

AI implementation
ISMS: what it is and why not to build it only in Excel
An ISMS is a cycle, not a document. The higher regime of Decree 409/2025 Coll. asks for more than a workbook can hold, the lower regime gets by with an overview.