Private AI for telco

An LLMOps control plane for operators: pool GPUs wherever they live, host and version models, see exactly who calls what with which data at what cost — and keep every token inside the perimeter.

4× H200

on-prem GPU pilot configuration with CETIN SK

100%

of inference inside the security perimeter — nothing leaves

LLMaaS

ambition: operator-hosted AI as a service for B2B customers (financial-sector PoC)

The idea

Telcos sit on exactly what regulated AI workloads need — secure facilities, power, connectivity, operational discipline — and on data that can never visit a public API. Private AI turns that into a platform:

• GPU pooling — on-prem cards, rented capacity and cloud burst managed as one resource, scheduled by workload.

• Model lifecycle — hosting, versioning and scaling of open and licensed models behind one stable interface.

• Observability — per-call accounting of who invoked which model with what data at what cost. AI spend becomes a dashboard, not a surprise.

• Governance — access control through the same Keycloak/AD identity fabric as the rest of the platform, with audit on every interaction.

At a glance

Status: In Development

Deployed at: Pilot with CETIN SK — 4× H200 on-prem

Domain: AI infrastructure

Built with: GPU pooling · Model serving · Observability · Governance

Why it matters

Every assistant on this site — the chatbots, the TESU ops assistant, the agents behind MCP — needs inference that is private, governed and economical. This platform is that foundation. And it points outward too: an operator that runs private AI well for itself can sell it — LLM-as-a-Service from telco facilities to enterprises with the same sovereignty requirements, with a financial-sector proof of concept already shaping the offer.

Where it stands

The pilot runs on 4× H200 on-premises with CETIN SK, serving the production assistant and agent workloads. The build-out is the control plane: pooling, multi-tenant serving, metering and the B2B packaging.