AI server for business, work and home: your own AI without the cloud

Your own AI server means the language model runs on your hardware and company data never leaves the building. You don't pay for tokens or per-user subscriptions, you can try new open models the day they are released, and you can expand the server step by step with more graphics cards. For a few people writing everyday e-mails, a cloud subscription is cheaper. A server pays off where data must not leave the company, where AI processes large volumes of documents, or where you want to integrate it into your own systems.

Summary

• A local model sends nothing to the cloud – contracts, customer data and source code stay in the company.
• For 20 people the cloud costs roughly 8,700–13,000 Kč a month for basic licences alone. Your own server is a one-off investment plus electricity.
• Which model will run is decided mainly by graphics memory (VRAM): 32 GB is enough for models of around 30 billion parameters, 96–128 GB for models of around 120 billion.
• You can start with one card and expand the server later. We build systems from roughly 110 thousand Kč up to servers costing millions.

Why companies want AI without the cloud

Employees paste contracts, quotes, spreadsheets with customer data and source code into chatbots. According to a Check Point analysis from autumn 2025, one in every 54 prompts to public AI tools carried a high risk of leaking sensitive data, and this affected 91 % of the organisations monitored.

Under the GDPR, the company using an AI tool is responsible for the processing of personal data in it. It must know where the data goes and, when transferring it outside the EU, have a legal basis. The framework for data transfers to the USA (EU–US Data Privacy Framework) did survive before the EU General Court, but the decision has been appealed and legal uncertainty remains.

Business versions of cloud services have safeguards: according to their terms, ChatGPT Business, Claude Team and Microsoft 365 Copilot do not train on company data. But the data is still processed by another company on its servers. A local model is the only option where the text never leaves your network at all.

One thing to watch: in 2025 the Czech government banned DeepSeek products in the state administration and the Czech Army warned against Chinese cloud chatbots because they store data in China. This applies to cloud services. A downloaded open model runs on your premises without connecting to their servers – even so, choose models with a clear licence and verified origin.

How much AI in the cloud costs

Prices excluding VAT according to the providers' price lists as of 25 October 2026, converted at the ČNB rate of 21.74 Kč/USD. If you integrate AI into your own applications via an API, you pay for every million tokens – for powerful models 2–10 USD for input and 10–50 USD for output. With automatic processing of thousands of documents a month, the bill adds up quickly.

What your own AI server can do

• Company chat in the style of ChatGPT for the whole team – for example Open WebUI with user accounts, running only on your network.
• Chat over company documents (RAG) – the model answers from your policies, contracts, manuals or e-mails and cites the source.
• Assistant for programmers – code completion in VS Code or JetBrains (Continue) without sending code outside.
• Transcription of meetings and calls – the Whisper model handles Czech too, so meetings don't have to be uploaded to someone else's server.
• Images, video and rendering – image generation (ComfyUI), Blender, DaVinci Resolve.
• Automation – your own applications call the model via an API just like the cloud, only without a per-token fee.
• Experiments – download a new open model (Qwen, Gemma, Mistral, gpt-oss, DeepSeek…) and try it out the same day.

How much VRAM you need: which server for which model

The whole model (including context) must fit in graphics memory. Rule of thumb: a model compressed to 4 bits needs around 0.6 GB for every billion parameters, plus memory for the context.

Speeds based on public tests from 2026; they change with the model and software version. For reference: 30 tokens per second is faster than you can read.

An AI server for the company, for work, or for home

For the company

A server in your server room or in our data centre, chat for the whole team, connected to documents and internal systems. For 10–50 people, 2–4 cards and a model of around 30–120 billion parameters are usually enough. For more concurrent users, vLLM is deployed, which queues requests efficiently.

For work – for developers, analysts, creatives

A workstation with one or two cards under the desk. A coding assistant, transcriptions, image generation and testing new models. A single RTX 5090 today handles models that a year ago were reserved for servers.

For home

For enthusiasts and small business owners: a local chat that doesn't send your queries anywhere, image generation, games and performance for other projects on top. Expect noise and heat – a card drawing 575 W heats like a small electric heater.

When your own server pays off and when it doesn't

Let's be honest: if your team only uses AI occasionally for writing texts, a subscription at 20 USD per person works out cheaper than a server. Your own server makes sense when at least one of the following applies:

• Data must not leave the company – healthcare, law, accounting, development, trade secrets, clients' personal data.
• You process large volumes – thousands of documents, transcripts of hours of recordings, automatic replies. An academic study from 2025 found a payback period of months for a small deployment with one RTX 5090 compared with a paid API.
• You want to experiment – try models, fine-tune them on your own data, build your own applications.
• You will use the server in several ways – AI during the day, rendering or training at night.

Also allow for additional costs. A server with 4 cards draws around 2.7 kW at full load – at 8 hours a day and 22 working days that is about 475 kWh and 2,600 Kč a month excluding VAT. It also needs administration: according to international surveys, updating models and drivers takes 10–20 hours a month unless the supplier handles it.

And quality: the best open models lag behind the top cloud models by roughly four months of development on average. The strongest ones, however, have hundreds of billions of parameters and need a big server. A single card will run a smaller model – which is usually enough for summaries, answers from documents or writing code.

How to start cheaply and expand

1. Start with one card and a model of around 30 bn parameters. You'll see how AI fits into your company.
2. Choose a base with headroom – a motherboard and processor with enough PCIe slots, a power supply with spare capacity and a case or frame for additional cards.
3. Add cards as needed – each additional RTX 5090 adds 32 GB of memory. Splitting a model across several cards reduces speed slightly but allows larger models.
4. As you grow, consider professional cards with 96 GB (RTX PRO 6000) or server cards – fewer cards, more memory on each.
5. Old cards don't lose their value – with the memory shortage in 2026, graphics cards are selling for more than they cost new. An upgrade therefore returns part of the investment.

How it works with us

We build AI servers to order in Prague – from a workstation for roughly 110 thousand Kč to servers costing tens of millions. We first agree on what will run on the server (model, software, number of users) and design the configuration and budget accordingly.

• We pre-install the system, drivers and software – switch it on and it works.
• We test every server for 72 hours under full load before handover.
• We deliver the server to your office or place it in our data centre – no noise and no cooling worries on your side.
• We keep spare parts in Prague and provide service within 24/48 hours.
• When the server isn't running your workloads, it can mine cryptocurrencies on its graphics cards.

What next

• Have a look at sample configurations and send an enquiry: AI servers with custom pre-installation.
• Specific configurations with prices: AI and GPU servers.
• What AI servers look like, how they are cooled and what determines their price: AI servers: cooling and price.
• How much a graphics card earns when the server isn't running AI: graphics card profitability.
• Which graphics card to choose for local AI: GPUs for local AI: RTX 5090 or RTX PRO 6000 Blackwell?

Sources

• Seznam Zprávy: every 54th prompt reveals a secret (27 Oct 2025, in Czech)
• epravo.cz: use of artificial intelligence tools and the GDPR (18 Dec 2025, in Czech)
• WilmerHale: appeal against the EU–US Data Privacy Framework (1 Dec 2025)
• BusinessInfo.cz: company AI without the cloud (14 May 2026, in Czech)
• ČT24: government bans DeepSeek in the state administration (9 Jul 2025, in Czech)
• OpenAI: ChatGPT Business pricing · Claude: pricing · Microsoft 365 Copilot
• Hardware Corner: RTX 5090 and language models
• DatabaseMart: RTX PRO 6000 test in Ollama (10/2026)
• arXiv 2509.18101: costs of running LLMs locally compared with an API
• Epoch AI: the gap between open and closed models (5/2026)
• ČNB: foreign exchange rates