Private AIwhere datacannot leave.
Banks, hospitals, government suppliers and plants hold data that cannot go to someone else's API. Open-weight models run inside your network, on GPUs you control, behind an OpenAI-compatible endpoint.
Four sectors, one rule: the data stays inside.
The stack is the same everywhere: an open-weight model, GPUs you control, an OpenAI-compatible endpoint. What changes is the data constraint, and with it the deployment shape.
Finance
Customer records and trading context stay in your data centre.
Account data, transaction history and internal research are covered by confidentiality duties, and a regulator or auditor can ask where every prompt went. A third-party API makes that answer harder.
- Deployment shape
Own GPUs. GPUs in your data centre or colocation rack, with the endpoint reachable only from your network. Audit logging and role-based access are part of a production deployment.
- Typical models
- gpt-oss-120b~117B MoE, Apache 2.0. 1× H100 80GB at 4-bit, 1× H200 141GB at FP8.
- Qwen38B / 32B / 235B-A22B, Apache 2.0. CPU only at 4-bit, 1× T4 16GB at FP8 (8B).
- Mistral Small 3.124B dense, Apache 2.0. 1× T4 16GB at 4-bit, 1× RTX 5090 32GB at FP8.
All three are Apache 2.0, which keeps legal review short. Hardware fit is listed per model.
- What we claim
- Keeping inference inside your network supports your GDPR and data-residency obligations. Any certification, such as SOC 2 or PCI DSS, remains yours to hold.
Healthcare
Patient data never has to leave the hospital network.
Clinical notes, imaging reports and patient messages are special-category data. Sending them to an outside model needs a processor agreement and a transfer basis, and sometimes cannot be justified at all.
- Deployment shape
Own GPUs, or confidential compute. GPUs on your premises for the strictest sites. Where you cannot host, confidential compute keeps prompts and weights encrypted in GPU memory on hardware you do not own.
- Typical models
- Gemma 3270M–27B dense, Gemma Terms of Use. 1× RTX 4090 24GB at 4-bit, 1× A100 40GB at FP8 (27B).
- Qwen38B / 32B / 235B-A22B, Apache 2.0. CPU only at 4-bit, 1× T4 16GB at FP8 (8B).
- gpt-oss-20b~21B MoE, Apache 2.0. 1× T4 16GB at 4-bit, 1× RTX 5090 32GB at FP8.
Small and mid-size models cover documentation and summarisation on one GPU. gpt-oss-20b needs about 16 GB.
- What we claim
- Keeping data inside your network supports your HIPAA and GDPR obligations. We do not claim HIPAA certification, and we do not claim FDA validation. Compliance remains yours to establish.
Government and defence contractors
A model that runs where the network has no way out.
Classified-adjacent and controlled data often sits on networks with no outbound route, so a hosted API is not an option. The model has to live in the enclave with the data.
- Deployment shape
Air-gapped. The endpoint has no outbound route. Weights and updates arrive on media your security team approves. Scope and sequencing are agreed in the pilot call.
- Typical models
- gpt-oss-120b~117B MoE, Apache 2.0. 1× H100 80GB at 4-bit, 1× H200 141GB at FP8.
- Qwen38B / 32B / 235B-A22B, Apache 2.0. CPU only at 4-bit, 1× T4 16GB at FP8 (8B).
- Mistral Large 2123B dense, Mistral Research License (MRL). 1× H100 80GB at 4-bit, 2× H100 (160GB) at FP8.
Open weights can be inspected, hashed and pinned. Licence and model origin are policy questions for your accreditor, so we list both.
- What we claim
- We do not hold FedRAMP or similar authorisations and do not claim them. We build to your accreditation boundary, and you or your sponsor accredit the system.
Manufacturing
Process know-how and designs stay on the plant floor.
CAD files, process parameters and supplier terms are trade secrets. An engineer pasting them into a public chatbot is the leak you are trying to prevent.
- Deployment shape
Own GPUs at the plant. A small GPU server at the site, on the factory network. If you would rather not run hardware, confidential compute offers hardware-verified attestation for the host.
- Typical models
- gpt-oss-20b~21B MoE, Apache 2.0. 1× T4 16GB at 4-bit, 1× RTX 5090 32GB at FP8.
- Gemma 3270M–27B dense, Gemma Terms of Use. 1× RTX 4090 24GB at 4-bit, 1× A100 40GB at FP8 (27B).
- Qwen38B / 32B / 235B-A22B, Apache 2.0. CPU only at 4-bit, 1× T4 16GB at FP8 (8B).
Small models handle manuals, work instructions and quality reports on modest hardware.
- What we claim
- Keeping documents inside your network supports your confidentiality and export-control obligations. We do not certify your controls. They remain yours to hold.
Talk to an engineer.
Tell us the model, the data and the GPUs you have. An engineer from Limassol or Auckland scopes the 72-hour pilot with you.