Local AI in your company: your data, your server, lower costs
We deploy language models (LLM and SLM) on hardware in your company. Employees get an AI assistant, while contracts, documents and client data never leave the company.
Data stays with you
Queries, documents and answers do not leave your network. No foreign data processor.
No per-user licenses
You pay for hardware once. A new employee does not increase the monthly bill.
Payback in 1 to 2 years
In a typical company with 100+ users, the server pays for itself faster than the warranty lasts.
Works without internet
AI is available even when the provider has an outage or changes terms and prices.
Cloud AI tools are great until you hit two problems: data goes to a foreign provider, and the bill grows with each new user. Local AI solves both. Open models such as gpt-oss, Gemma 4, Qwen, Mistral and DeepSeek now reliably handle most everyday tasks: writing and summarizing documents, answering over internal documentation, extracting data from invoices and contracts, translations and developer assistance. We choose the right model and hardware for your tasks, set up the server in your network and connect it to your systems. Employees get an internal AI chat on your domain, with login through existing accounts, and you pay for hardware once instead of monthly licenses. For the most demanding tasks we recommend a hybrid model: most work locally, and a smaller number of cloud licenses only where they truly make a difference.
What we deploy
Model and hardware selection
We test models on your real tasks and choose the smallest hardware that runs them reliably.
Internal AI chat
A ChatGPT-like application on your domain, with login via Microsoft 365 or Active Directory accounts.
AI over documents, locally
Search and answers over contracts, procedures and archives, with a source for every answer (RAG).
Small specialized models (SLM)
Fast and inexpensive models for a single task: email classification, data extraction, OCR and quality control.
Integrations with your systems
OpenAI-compatible API, so ERP, CRM and existing applications use local AI without modifications.
Hybrid model
Local models for most work and cloud models only for the hardest tasks, with clear rules on what may leave the company.
Security and access control
Team-based access rights, internal logs, encryption and server isolation in the network.
Cost and performance measurement
Reports on usage, speed and savings compared to cloud licenses.
Growth without rewrites
We design the system so that another GPU or server can be added as the number of users grows.
Indicative hardware prices (November 2026)
Hardware purchase only. Our implementation, model customization and maintenance are contracted separately.
Mini AI server
€5,500 to €8,000
- Configuration
- NVIDIA GB10 (DGX Spark or ASUS Ascent GX10), 128 GB memory, or Mac Studio with 128 GB and more
- For whom
- Pilot and smaller teams, up to about 15 users
- Models
- gpt-oss-20b, Gemma 4, small specialized models, gpt-oss-120b at lower speed
AI server with 1 to 2 GPUs
€18,000 to €40,000
- Configuration
- 1 or 2 NVIDIA RTX PRO 6000 cards (96 GB each), server platform, 256 GB RAM, NVMe disks
- For whom
- Companies with 20 to 150 users and AI over documents
- Models
- gpt-oss-120b, Mistral Small 4, Gemma 4 and multiple models in parallel
AI server with 4 GPUs
€65,000 to €85,000
- Configuration
- 4 NVIDIA RTX PRO 6000 cards (384 GB total), rack server with redundant power supply
- For whom
- 150 to 400 users, AI agents and processing of large archives
- Models
- Models in the 300 billion parameter class or multiple models simultaneously
Data center class
from about €280,000
- Configuration
- NVIDIA HGX with 8 H200 GPUs (1,128 GB total)
- For whom
- Banks, telecoms, public sector and large enterprises
- Models
- The largest open models, for example DeepSeek class
Prices are indicative, excluding VAT and transport, based on public prices from November 2026. Graphics cards and memory became significantly more expensive during 2026 due to a global memory shortage, so prices change from week to week. Before purchase we check current supplier offers. If needed, add UPS, rack cabinet and cooling, about €1,000 to €5,000.
Example: a company with 120 employees
A consulting and accounting firm works with contracts, financial reports and personal data of clients. It wants AI for all employees, but must not send client data to a foreign provider. We compare two paths over three years.
| Item | Cloud AI licenses | Local AI server |
|---|---|---|
| Initial investment | €0 | about €35,000 (server with 2× RTX PRO 6000) |
| Licenses for 120 users | €25,000 to €37,000 per year ($20 to $30 per user per month) | €0 |
| Electricity | included in price | about €1,000 to €1,500 per year |
| Cloud for the most demanding tasks | included | 10 licenses, about €3,000 per year |
| Total for 3 years | €75,000 to €112,000 | about €48,000 |
| Where the data is | at the provider, outside the company | in the company's server room |
The investment pays back in 13 to 20 months, and the three-year savings are €27,000 to €64,000. The server continues to run after that, and each new employee does not increase the cost.
- Contracts, reports and client data do not leave the company, so there is no new data processor and GDPR and local law compliance is simpler.
- Access is through existing employee accounts, by teams, and all logs remain internal.
- The local model handles summarization, writing, document search and data extraction. Ten users with the most demanding tasks keep a cloud license.
- As the company grows, a GPU is added instead of hundreds of new licenses.
The example is illustrative, with prices from November 2026. Our implementation, training and maintenance are not included.
Models and technologies
How we introduce local AI
- 1
Assessment
Which tasks AI should perform, how many users, which data is confidential and how much you currently pay for cloud tools.
- 2
Pilot before purchase
We test models on your real examples, so you buy hardware only when you know the solution works.
- 3
Hardware and installation
We propose a configuration, assist with procurement and set up the server, models and internal AI chat.
- 4
Go live
Integration with accounts and systems, brief employee training and measurement of usage and savings.
Why us
We developed Intrascope, a platform through which companies use more than 120 AI models, and we daily compare both open and commercial models on real business tasks. That is why we know where a local model is good enough and where cloud still pays off, and we propose only the hardware you truly need. We also maintain IT infrastructure for companies with more than 100 workstations, so the server does not remain unsupervised after installation.
100%
of data stays in your network
€0
per-user licenses per month
13 to 20
months to payback in the example
Frequently asked questions
What is the difference between LLM and SLM?
LLM (large language model) is a general assistant for writing, analysis and conversation. SLM (small language model) is a smaller model for one specific task, for example email classification or data extraction from invoices. SLM runs faster and on more modest hardware, so we often combine them.
Is a local model as good as ChatGPT?
For most everyday business tasks, such as summarization, writing, document search and data extraction, today's open models are good enough. For the most demanding analyses we recommend a hybrid model with a smaller number of cloud licenses.
How much does hardware for local AI cost?
According to November 2026 prices, a mini AI server for a pilot costs about €5,500 to €8,000, a server with 1 to 2 GPUs for most companies €18,000 to €40,000, and a server with 4 GPUs €65,000 to €85,000, excluding VAT. Hardware prices change quickly in 2026, so before purchase we check current offers.
How long does the investment payback take?
It depends on the number of users and how much you currently pay for cloud tools. For companies with 100 or more users, payback is most often between one and two years.
Can we start without buying a server?
Yes. We run the pilot on smaller hardware or in a rented test environment, with your examples. You buy hardware only when you see that the solution works.
How much electricity does an AI server consume?
A server with two professional graphics cards under load consumes about 1.2 to 1.5 kW, and on average significantly less. For a typical company that is about €1,000 to €1,500 per year.
Do we need internet?
Not for operation. The model, documents and application are in your network. Internet is needed only for updates and, if you use it, hybrid access to cloud models.
Who maintains the system after installation?
We can, through a monthly maintenance package: model updates, monitoring, backup and user support. Maintenance is contracted separately and is not included in the hardware price.
Free fixes from our knowledge base
Step-by-step guides for the problems we solve most often as part of this service.
- ChatGPT reports an error or limit: how to continue working
- AI gives wrong answers: how to recognize hallucinations
- Which laptop to buy for office and students in 2026
- Laptop for programmers: which configuration is enough
- Laptop overheating and noisy fan: how to cool it down
- PC Won't Turn On: Check Power Supply and Power Button
Want AI without sending data outside the company?
Tell us how many employees you have and what you would use AI for. We will get back to you within one business day with a pilot proposal and an indicative configuration.