Bastion
In-house AI
On-premise AI for companies that cannot send data to the cloud.
Your own AI. Inside your walls. Under your control.
Bastion is a private AI appliance delivered as one sealed unit: the hardware, the open model and the hardened software, ready to plug in. Your teams get a secure ChatGPT-style assistant that never sends a single word outside the building, at a flat monthly fee.
- Hardware included
- Data never leaves
- Works offline
The problem with cloud AI
Cloud AI costs you five things the invoice never shows.
Cloud assistants are convenient. For a bank, a hospital, a law firm or a public body, they carry five risks that on-premise AI removes.
- With cloud AI
Your data leaves the building
Every prompt travels to a foreign server. Contracts, patient files and source code included. Vendor terms decide what happens next, and those terms change.
- With cloud AI
AI costs you cannot predict
Cloud AI bills per token. Usage grows, prices move, enterprise tiers start at large annual minimums. Finance cannot budget a line that doubles when a team gets productive.
- With cloud AI
You depend on someone else's switch
Rate limits, outages, model deprecations and policy changes happen on the vendor's schedule. When they pause, your teams pause with them.
- With cloud AI
GDPR and DORA compliance lands on you
You must document where processing happens and who can reach the data. No AWS, Azure, Google Cloud or Oracle region has been announced in Romania. Every cloud answer is a foreign one.
- With cloud AI
No internet, no intelligence
Cloud tools need the network up and the grid on. A comms outage or a power cut turns the smartest assistant into a blank screen exactly when you need it.
With Bastion“The safest data is the data that never left.”
Bastion starts from that sentence and builds everything else around it.
The solution: a private AI appliance
One sealed box.
Everything an in-house AI needs.
Bastion is not a parts list you integrate. It is a finished on-premise AI server you unbox. We bring the hardware, the model, the secure software and the support. You bring a power socket and a network cable.
One standard node serves
- concurrent users
- 35
- concurrent users
- tokens of context
- 8k
- tokens of context
- tokens/s each
- 30
- tokens/s each
Calculated for the fp8 KV cache the node ships with. Enough for the everyday AI use of a mid-sized office. Larger teams get more nodes.
Enterprise hardware
A professional GPU server, built for continuous use and sized for your team. Delivered, installed, and replaced as a whole unit if anything fails.
An open model, tuned and ready
A state-of-the-art open-weight LLM with a commercial licence. Chat, drafting, code, Romanian and English, through a ChatGPT-style interface.
Sealed and hardened
A locked operating system with signed software, no outbound internet, and a full audit log you own. Security is a property of the box, not a promise in a policy.
Plugs into your network
Managed like a network appliance. Your staff open a browser, your applications use a standard API. Existing tools switch over by changing one address.
Optional power backup
Add an external battery unit and a power cut costs you no work. The box finishes the answers in flight, shuts down cleanly, and starts again when the power returns.
Rented, not bought
One flat monthly fee covers hardware, model updates and support. No capital expense, no depreciation schedule, a fixed term you can end.
What your sector may run
The EU AI Act draws a line. We publish where it falls.
Bastion is offered for general assistance work only. Every purpose listed in Annex III of the EU AI Act is excluded from the offer, which keeps the box out of the high-risk regime. Here is what that means for your work, sector by sector.
Banking and finance
You may run
Regulatory and policy text Q&A, internal audit and complaint classification. Fraud detection as well, which point 5(b) excepts by name. Annex III does not list anti-money-laundering work at all.
Annex III does not allow
- 5(b)Judging whether a person is creditworthy, or setting their credit score.
Insurance
You may run
Policy wording, claims correspondence and internal research.
Annex III does not allow
- 5(c)Risk assessment and pricing for life and health insurance.
Healthcare
You may run
Discharge summaries, clinical letters, coding support and literature Q&A. Annex III lists none of them.
Annex III does not allow
- 5(d)Emergency patient triage, and the dispatch of emergency services.
- 5(a)Deciding, as a public authority, who is eligible for a healthcare benefit.
Annex III is not the only law here. Medical device rules can also reach clinical software.
Law firms
You may run
All of it, legal research and interpretation included. You act for your client, not for the court, so point 8(a) does not reach that work.
Annex III does not allow
- 8(a)Working on behalf of a court, or sitting as an arbitrator. Representing a client is neither.
This is the narrowest limit any buyer on this page carries.
Public sector and courts
You may run
Judgement anonymisation, registry and publication work, classification and scheduling.
Annex III does not allow
- 8(a)Helping a judicial authority to research or interpret the law.
- 5(a)Deciding who is eligible for an essential public benefit or service.
Annex III has eight areas and a public body can fall in most of them: biometrics, critical infrastructure, education, employment, law enforcement, migration, and point 8(b) on elections. Ask us which one covers your service.
One limit crosses every sector: no AI in recruitment, promotion, task allocation, termination, or monitoring the performance of your own staff. Annex III points 4(a) and 4(b). This is our reading of Regulation (EU) 2024/1689, not legal advice.
Why on-premise AI pays
One flat fee. Zero data leaks.
For clients
What you get on day one
- Nothing confidential ever leaves your premises
- One fixed monthly fee, no per-token surprises
- Works during internet outages. With the battery, a power cut costs no work
- GDPR and DORA answers you can hand to an auditor
- The same familiar chat experience your staff already know
- Swap-out support: a failure is our problem, not yours
AI cost control
A bill that stops climbing
A 50-person team using cloud AI daily reaches 100 million tokens a month within months. That bill moves every month and rises as adoption rises. Bastion is one flat line, priced per box and never per token.
Illustrative shape, not a quote. Your offer states the exact monthly fee for your team size.
For investors
No law forces private AI. What forces it is the map: no hyperscaler runs a region in Romania, so keeping data in the country means keeping it in the building. Bastion is that building.
Recurring revenue, physical moat
Each node is rented on a fixed term. Revenue compounds with every box in the field, and a sealed appliance is far harder to churn than a software subscription.
A market with no local cloud
Banks, hospitals, law firms and public bodies need AI and must document where it runs. No hyperscaler has announced a region in Romania, so every cloud answer is a foreign one. That is a gap in the map, and a gap has a date on it.
The bundle is the product
Anyone can buy a GPU. Nobody sells the hardware, model, hardened OS, compliance file and support as one finished, inspectable unit. That integration is the asset.
How it works
From first call to first answer in three steps.
01
Tell us about your team
Industry, number of people, whether you need power backup. That is enough for a sized offer within days.
02
We build and deliver the node
Hardware assembled, model loaded, software sealed and tested. It arrives as one unit with its compliance file.
03
Plug in. Start asking.
Connect power and network. Your staff open a browser and have a private assistant. Updates arrive as a signed file your own IT applies on your network. The box opens no connection to us.
Frequently asked questions
On-premise AI, answered plainly.
What is an on-premise AI appliance?
A complete AI system, hardware and software, that runs inside your own building instead of in a vendor's cloud. Bastion delivers it as one sealed unit: GPU server, open-weight language model, hardened operating system and support. Your staff use it like ChatGPT, but nothing leaves your network.
Is Bastion GDPR and DORA compliant?
Bastion keeps all processing inside your premises, so there is no transfer to a third-country processor to justify. You receive a contract annex written to DORA Article 30, the EU Declaration of Conformity and the audit log. Compliance still depends on how you use it, but the hardest questions are answered before the auditor asks.
Does it work without internet?
Yes. The appliance has no outbound internet connection by design. It answers on your local network during an internet outage. The optional battery bridges a short dip, then shuts the box down cleanly, so a power cut costs you no work. It is a ride-through, not a generator.
How many people can use one Bastion node?
One standard node serves 35 concurrent users, at 8k tokens of context, at 30 tokens per second each, with the fp8 KV cache it ships with. In practice that covers the everyday AI use of a mid-sized office. Larger organisations get additional nodes.
How much does on-premise AI cost compared to ChatGPT or Claude?
Cloud AI is billed per token and grows with adoption. Bastion is a fixed monthly rental that covers hardware, model updates and support, with no capital expense. Your offer states one number for your team size, and it does not move when usage does.
Which AI model runs on the box, and is it as good as GPT or Claude?
A state-of-the-art open-weight model with a commercial licence, served with a production inference engine. For drafting, summarising, document search, coding help and Romanian or English chat it is on par with mainstream cloud assistants. For the very hardest reasoning tasks, frontier cloud models still lead, and many clients keep both.
Request an offer
Tell us who you are. We size the box.
Seven short answers are enough for a concrete monthly offer. No sales calls until you ask for one.
- Your details are used only to prepare the offer.
- Prefer email? Write to contact@bastion-inc.com
- Based in Cluj, Romania. We answer in Romanian or English.