What it includes
A server or workstation with a graphics card, a serving engine, model files, a vector database of documents, the agent layer, monitoring and a log of requests. All updates, including new model versions, come in as a portable image through a gateway under the security procedure. The airport from this section’s cases built an AI code-check plugin into its development environment this way, and the operator of an insurance system ran requests, support and code review inside the circuit on one model.
What the circuit does not have
Cloud APIs, and so no flagship models: only models deployed inside are available, up to the size that fits in memory. Automatic updates: each update is an event with a repeat test. Internet search for agents: sources are internal only. Outside support by remote access: the engineer is inside.
Which models
Open models that handle Russian: a 32-billion and an 8-billion model from Russian developers, global open model families, and if needed an on-site delivery from a Russian provider. Size by memory in 4 bits, by the calculator: 32 billion on a card with 32–96 GB, 70 billion on 96 GB or two cards, 235 billion on two 96 GB cards in 4 bits. A telecom operator required the option to choose the model in its specification: plan for this from the start.
What it costs as of 29 September 2026
By our calculation, a workstation with an RTX 5090 costs about $13,000, for a pilot with 1–3 users. A server with two RTX PRO 6000 cards costs about $30,000 for 10–20 users, or about $2,100 a month with hosting and an engineer. A server with four H100 cards starts at $120,000 for hundreds of users, about $6,100 a month. Rollout: a pilot in 14 days from $1,499, a department in 4–6 weeks from $4,999.
How to update without the internet
The model is downloaded on an outside machine, checked against checksums, moved through the gateway and installed next to the old version. It passes a test on 50 requests, and only then becomes the main one. The old version stays for rollback. By our estimate, a quarterly procedure is enough: there is no point in chasing every new model version if the task is already solved.