Local AI for companies in Romania: what it is, what it costs and how to choose (2026)
Local AI means a model running on a server dedicated to your company, not in someone else's cloud. The options a Romanian or EU company has, what they cost, and what to check before you choose.
Local AI for a company is an AI model that runs on a server dedicated to that company only, at its site or in a managed data centre. Documents, email and ERP data never go to a public AI service, and the model does not train on them. The difference from ChatGPT or Copilot is not the interface; it is where the model runs and who else has access to it.
This guide covers the options a company in Romania or elsewhere in the EU has, what each costs, and what to check before choosing.
What "local AI" actually means
An AI is local to your company only if three conditions hold:
- The model runs on a server dedicated to your company. Not on infrastructure shared with other customers.
- Data never reaches a public AI. Not the questions, not the documents, not the answers.
- Your data trains no model. Not the vendor's, not a third party's.
"Hosted in Romania" does not automatically mean local. A cloud service with servers in the country keeps data in Romania, but it still sits on the provider's infrastructure, next to other customers' data.
The four options a company has
| Option | Data leaves the company | Works without internet |
|---|---|---|
| Local AI, on a dedicated server | No | Yes, optionally |
| Cloud AI hosted in Romania | Yes, to the provider | No |
| Copilot / ChatGPT Enterprise | Yes, to the provider | No |
| Built in-house, open-source | No | Yes |
- Local AI, on a dedicated server. The model runs on the company's server, on site or in a managed data centre, and no one else uses it. You pay a fixed software subscription and buy the server once. The vendor maintains it.
- Cloud AI hosted in Romania. The model runs on the provider's servers in the country, usually alongside other customers. You usually pay a subscription or per user.
- Copilot / ChatGPT Enterprise. The model runs in Microsoft's or OpenAI's cloud, on shared, logically isolated infrastructure. You pay per user, monthly.
- Built in-house, open-source. The model runs on the company's servers, but the hardware and the people who maintain it are yours.
The last option, built in-house with tools such as Ollama, vLLM or Open WebUI, is within reach for a simple chat. The hard part comes after: connecting email, SAP and the scanned archive, finding the right document and showing the source of every answer. That is where most of the time goes.
When local AI makes sense
- When you want to use AI on contracts, salaries, offers and unpublished figures you would not put into a public AI.
- When you work in a regulated industry, or your customers need to know exactly where their data sits.
- When many people ask many questions: on a fixed subscription, one more question does not cost one more licence.
- When the AI has to work without an internet connection.
For generic writing with no internal data, a public cloud AI is usually the simpler choice.
What local AI costs
There are two parts: the server, bought once, and the software, paid as a subscription. DOYNA's prices, excluding VAT:
| Plan | People | Software, per year | Server, one-off |
|---|---|---|---|
| Starter | Up to 10 | €24,000 | ~€26,000 |
| Growth | Up to 30 | €48,000 | ~€61,000 |
| Scale | Up to 60 | €84,000 | ~€99,000 |
| Enterprise | 60+ | On quote | On request |
The servers: Starter is 1× RTX PRO 6000, Growth 1× H200 NVL, Scale 2× H200 NVL. The subscription starts at €2,000 a month and is not charged per user. On your own server, one more question costs only the electricity it uses: less than a few cents. Server prices are estimates from supplier quotes in October 2026 and are re-quoted at order, because GPU prices move fast. Details: pricing and hardware.
What to check before choosing
- Where does the model run? On a server that is only yours, or on shared infrastructure?
- Does it train on your data? The answer should be "no", in the contract.
- Does it show the source? Every figure should lead to the document, page and date it came from.
- Which sources does it read? Email, SAP or another ERP, SharePoint, scanned contracts.
- Does it handle your language natively? For Romanian companies, contracts, email and meetings are in Romanian.
- Does it work offline, if you need it to?
- Can you test it on your own data? An AI that looks good on someone else's data tells you nothing about yours.
What DOYNA looks like
DOYNA is private AI for companies, built in Cluj-Napoca. It reads the company's email, SAP, contracts and files and returns the finished work: the deck (.pptx), the PDF or the direct answer, with the source behind every figure. It runs on a dedicated server that serves only your company, on site or in a managed data centre, and can run fully offline. Romanian and English are both native.
It runs in production at a Romanian company of more than 890 people, on more than 600,000 documents. DOYNA has been covered by Ziarul Pozitiv and Business Catalog (both in Romanian).
How to start
It starts with a pilot on your own data. We connect the first source together, on a call: Gmail, Outlook or a set of documents. The pilot runs on a machine we own, at our office in Cluj-Napoca, and what we hold is deleted within 60 days of the pilot ending.
Write to us or at [email protected]. We answer within one business day.
Related: Private AI guide for enterprises · AI assistant for business — how to choose · Doyna vs Glean vs Copilot