AI for confidential documents, installed at your office
A dedicated system with open-weight models preinstalled, running without an internet connection.
For case files, investigations and archives
Insolvency and restructuring
“Tell us what happened to this company.”
Construction and infrastructure disputes
“Reconstruct everything related to Variation Order 73, with sources.”
Digitisation backlogs
“Classify the 1,700 documents the OCR pipeline could not.”
Fraud investigations
“Find every payment that does not match a contract.”
In a 2026 survey by the Romanian National Bar (UNBR), 30% of lawyers said they avoid entering sensitive data into AI systems. Source: juridice.ro
How a project runs
- 1
Test on a document sample
3 to 5 days · €1,500
You receive a written report and a fixed quote before you confirm the project.
- 2
Delivery and installation
1 day
The system arrives sealed. We inventory and photograph the equipment with you.
- 3
The project
From 5 days
Your team asks the questions. A Piatra engineer is on call.
- 4
Collection and wipe
Last day
Storage is cryptographically erased in your presence, with a report signed by both sides.
6 configurations
Pod 0.5T2 × 256 GB512 GB · 0.96 kW€3,000per 30 days Pod 1T-44 × 256 GB1 TB · 1.92 kW€5,400per 30 dayshigher throughput Pod 1T-22 × 512 GB1 TB · 0.96 kW€4,200per 30 dayslarger models per node Pod 2T4 × 512 GB2 TB · 1.92 kW€7,600per 30 days Pod 3T6 × 512 GB3 TB · 2.88 kW€10,500per 30 daysflagship Pod B3008 × 288 GB2.3 TB · 11 kW€55,000per 30 days20 kW feed
Each configuration arrives sealed: Apple Mac Studio M5 Ultra nodes linked directly over Thunderbolt 5, their own network, encrypted storage, UPS, cooling and the models loaded: Kimi, GLM, DeepSeek, Qwen or your own model.
Pod B300: 8 NVIDIA B300 GPUs in one cooled rack, for workloads a Mac cluster cannot carry. It needs a dedicated 20 kW feed.
Services and options
- Data Vault, encrypted storage server, 32 to 40 TBper 30 days€900
- Test on a document sampleincluded from 30 days€1,500
- On-premises installation, anywhere in Romaniaone-timefrom €1,500
- Controlled room: deployment protocol, numbered seals, named accountsone-time€4,000
- Engineer, 2 days: setup and handoverin every projectincluded
- Additional engineer dayper day€1,000
Indicative prices, in EUR, excluding VAT, per 30 days of use, as of 11 September 2026. A fixed quote follows the sample test.
Models and capacity
Pick a model and see which configurations it fits, how fast it answers one user, and how many users or agents it can serve at the same time.
A comfortable pace for interactive work: about 630 words per minute.
| Configuration | Fits | tok/sOne user, tok/s | usersConcurrent users |
|---|---|---|---|
| Pod 0.5T | no | – | – |
| Pod 1T-4 | no | – | – |
| Pod 1T-2 | no | – | – |
| Pod 2T | yes | 18 | 1 |
| Pod 3T | yes | 20 | 1 |
| Pod B300 | yes | 100 | 279 |
Estimates from memory bandwidth (1.2 TB/s per Apple M5 Ultra node, 8 TB/s per NVIDIA B300 GPU) and each model’s active parameters, calibrated on published measurements. For Mac: Qwen3.5-122B-A10B on an Apple M5 Max (Silicon Score) and DeepSeek V3.1 on 1, 2 and 4 Mac Studio M3 Ultra (Jeff Geerling). For NVIDIA: Kimi K3 on 8 B300 GPUs, at 100 tok/s for one user and 38 tok/s each for 128 concurrent users (BIZON). The real figures are established in the sample test, on your documents.
Estimate a project
Pick a configuration, a duration and where the system runs. Add the estimate to your message and we come back with a fixed quote after the sample test.
What stays with the AI provider
In September 2026, mathematicians who had done their work inside OpenAI's products asked publicly where that work had gone.
The Navier-Stokes proof
Tristan Buckmaster and Levent Alpöge asked whether the OpenAI model had been trained on the Codex sessions holding their drafts, after OpenAI published a proof with the same approach. OpenAI denied any access to their data, but wrote that it cannot rule out that de-identified data from their use of its products helped improve its models.
Non-sofic groups
Andreas Thom, a professor at TU Dresden, had spent months in ChatGPT on questions around the problem OpenAI later announced as solved, and asked whether those conversations had entered the training data. He received a one-line answer, which he calls misleading.
On a system installed at your office, without internet access, the data passes through no provider. What you enter stays on machines you control and is erased at the end of the project, in your presence.
Your data stays with you
- The machines stand in a room you control. Your security team can inspect them at any time.
- Piatra processes the data directly, with no cloud and no subprocessors, under a data processing agreement and an NDA signed before the test.
- Every request to the model is logged in the system. The log is yours and you can read it at any time.
- At collection, storage is cryptographically erased in your presence. You receive the report signed by both sides.
If the system cannot be installed at your office, the same system runs on dedicated equipment in Piatra Neamț, reached over a private link.
What we do not promise
- A system serves one client and one project for the agreed period.
- Availability is agreed per project. A spare node is priced separately.
- The system arrives, runs the project and is collected, unless you choose to buy the equipment.
- GDPR compliance rests on the data processing agreement and the retention and deletion terms.
Questions
Is this GDPR compliant?
A data processing agreement, retention and deletion terms, no subprocessors, and the query log, available at any time. Your DPO is welcome in the scoping call.
We already have a NAS. Do we need the Data Vault?
No. The system connects to your storage over the isolated network and reads from there.
How long can a project be?
Any number of days from 5. After the first 30 days, the daily rate drops by 15%.
What does the controlled-room option include?
A deployment protocol agreed with your security officer, numbered seals on the enclosure and unused ports, checked at every visit, and accounts only for the people you name, with every session logged. Access to the room is decided by you.
Can we keep the equipment?
Yes. We quote the equipment separately, and half the rent you have paid comes off the price.
Do you sign an NDA before the test?
Yes, before receiving any document.
How is this different from an AI service in the cloud?
In the cloud, your documents travel to someone else’s servers. Here the models run on a machine in your room, with no third party and no internet access.
Does it need special power or cooling?
Every configuration arrives with its own cooling. The Mac configurations run from ordinary office circuits; Pod B300 needs a dedicated 20 kW feed, which we confirm with your building’s electrician.
Tell us about the project
We come back with a concrete proposal and a date for the sample test.