Tatneft brings AI coding into an air-gapped environment: one second for a line suggestion

The company is part of critical infrastructure, so its environment is completely isolated from the internet, and only software from the Russian registry may be selected. Its subsidiary TatITneft filed the application for an AI Development Platform on B2B-Center.

The servers are sized for 50–200 users and 80 requests per minute; the platform is provisioned with Nvidia Tesla H100 NVL accelerators and Intel Xeon Gold 6448Y processors. The assistant must understand code from C# and Java through PL/SQL and Bash, connect to self-hosted GitLab, Active Directory, and Test IT, and speak Russian with engineers.

Previously, a corporate coding assistant was bought as a per-user subscription: SourceCraft by Yandex has sold for $5,74 per seat per month since 6 April 2026. Now the oil company is buying an installation on its own servers, and vendors have already adapted. In June 2026, Sber released GigaCode CLI with local deployment and explicitly named its target customers: banks, industry, telecoms, and the public sector.

**What gets deployed in such an environment.** Kodify from MWS AI requires 8–32 CPU, at least 32 GB RAM, one 80 GB Nvidia A100, and an SSD from 500 GB, and runs on RockyLinux or AstraLinux. The 7-billion-parameter model holds 32 768 tokens of context and knows 90 programming languages. Licenses cost from 20 000 ₽ per user excluding VAT; plugins run in VS Code and JetBrains.

The price differs threefold. Akhmetzhan Makhmutov of the St Petersburg Oil Terminal estimates such an implementation at 400–600 mln ₽ in the first year and 150–250 mln ₽ annually after that. Nikita Vekesser of Nova AI names 100–200 mln together with the equipment. The tender allocates 7 working days to patch a discovered vulnerability.

Pilot operation will start no later than 1 December 2026, with the final acceptance certificate scheduled for 28 February 2027.