International news agency Rossiya Segodnya has deployed a local AI compute cluster built on YADRO G4208P G3 servers, moving inference workloads inside its own IT perimeter. The system, operational since December 2025, runs transcription, multilingual audio/video processing, subtitle generation, and a photo bank search over 3.5 million images—all without shipping data to external APIs.

A custom language model trained on the agency’s own archive of RIA Novosti materials generates draft texts autonomously. The architecture uses eight GPU accelerators per node, chosen to handle sustained inference loads without hitting memory or latency ceilings. Engineering teams scoped priority editorial scenarios in March 2025, then selected hardware configs and deployment patterns before cutting over at year-end.

Executive director Dmitry Ushakov noted the shift from experimental AI to production tooling: as information volume grows, the team needs to accelerate data pipelines while maintaining editorial quality. The on-prem approach gives them full control over data governance—a non-negotiable for a media organization handling sensitive sources and proprietary archives.

YADRO’s commercial director Alexander Bakulin framed the deployment as a maturity signal. “This isn’t a sandbox. It’s a working system automating routine ops, freeing editors for tasks requiring judgment and creativity.” The setup is designed for horizontal scaling, with room to add more models without re-architecting. That pattern—local inference clusters behind a corporate firewall—is directly applicable to finance, healthcare, and government, where data residency and latency matter.

The project will be detailed at CIPR Forum 2026 in Nizhny Novgorod this May.

Source: RIA Novosti