MMOOZZGG.com

Architecture

On premise AI: three levels and how to choose

3 min read15 sources

On premise AI, or local deployment, means the model runs on your own hardware, while the network can stay open: updates and outside sources remain available. Below are three levels by number of users, with prices, and who works at which level.

On this page · 5 sections

How it differs

Versus an air-gapped setup: there is internet, so updates are easier, and you can go hybrid and use the cloud for anonymized tasks; security demands are softer. Versus GPU rental: the hardware is yours, and against rental it pays back only at high, steady load; the payback calculator shows the term, and you need an engineer and space. Versus a Russian cloud: the environment is yours, there are no third parties, and no data processing agreement is needed.

Level 1: one computer, 1–3 users

A Mac Studio M4 Max with 128 GB for $11,000, or a workstation with an RTX 5090 for about $13,000. Models up to 70 billion parameters on the Mac and up to 32 billion on the RTX 5090, with LM Studio or Ollama, at about $400–$800 a month in full cost by our calculation. The IT director of a bank runs a personal agent on a Mac at about 70 tokens per second.

Level 2: a team server, 10–20 users

A server with two RTX PRO 6000 cards of 96 GB each costs about $30,000, runs vLLM, and costs about $2,100 a month with hosting and an engineer by our calculation. It runs 70-billion-parameter models in 16 bits or 235-billion models in 4 bits, with a queue of agents. The legal department of an IT holding cut contract approval from two weeks to three days with its own model.

Level 3: server cards, hundreds of users

A server with four H100 cards from $120,000, about $6,100 a month by our calculation, or several such servers. A bank running its own models resolves about 45% of requests with no operator and estimates that support costs fall by at least 25%. According to an industry publication, a server for 1,000 simultaneous users costs about $660,000 and pays back at 80–90% load.

Platform or custom build

Integrators sell on premise AI as a platform with an interface, pipelines, and access rights; this is how the LANIT group presented a platform for deploying models on the client’s infrastructure. A build from open components costs less in licenses and needs an engineer. We build from open components and hand the result to your engineer with documentation.

Questions and answers

Do you need your own data center?

For levels 1 and 2, no: the Mac sits on a desk, and a team server can be placed in a commercial data center from $90 a month for a server of up to 300 W. A level 3 server at 3.8 kW costs about $900 a month at the same rate by our calculation; an air-gapped setup needs a separate network with no internet.

How do you calculate payback?

Use a break-even calculator: the hardware price, the monthly costs, and the cloud bill. At level 1, payback exists only against expensive models; for a large server, according to an industry publication, it comes at 80–90% load.

Sources and fact-checking

Checked by the editors on 29 September 2026.

  1. store-apple.msk.ru, Mac Studio M4 Max 128 GB, 29 September 2026 (store-apple.msk.ru)
  2. Regard, GeForce RTX 5090, prices as of 29 September 2026 (regard.ru)
  3. iXBT, RTX PRO 6000 Blackwell 96 GB for $13,000 at DNS, 2 February 2026 (ixbt.com)
  4. Servermall, NVIDIA H100 SXM5 80 GB, price as of 29 September 2026 (servermall.ru)
  5. HOSTKEY, colocation in Moscow, 2026 (hostkey.ru)
  6. ComNews, local AI and the cost of automation, 19 February 2026 (comnews.ru)
  7. Habr (Slurm), DevOps engineer salaries, 3 September 2026 (habr.com)
  8. Digital Platforms, case no. 2447, October 2025: T-Bank (platforms.su)
  9. Digital Platforms, case no. 3855, 2 September 2026: T-Technologies (platforms.su)
  10. Digital Platforms, case no. 4067, 2 September 2026: Alfa-Bank, IT director (platforms.su)
  11. Digital Platforms, case no. 2929, 2025: T1 IT holding (platforms.su)
  12. Digital Platforms, case no. 3475, 2 September 2026: NSIS, operator of an insurance information system (platforms.su)
  13. Digital Platforms, case no. 3661, 2 September 2026: Digital Sector and a client from IT infrastructure (platforms.su)
  14. Digital Platforms, case no. 4084, 2 September 2026: VK Tech (platforms.su)
  15. Digital Platforms, case no. 2874, 2025: LANIT (platforms.su)

Found a mistake? Tell us

Publications

Cases on this topic: 8

  • Market case · Russia

    A large Russian bank in-house LLMs solve 45% of requests; support costs estimated down at least 25% a year

    The bank put its own family of language models into its chatbot for LLM customer support: about 45% of customer requests are solved without an operator. In October 2025 it launched AI copilots for first-line operators, and more than 90% of support staff use them every day. The models run inside the bank’s network perimeter, so data does not leave the bank.

    Banks and finance

    source: platforms.su

    Read the case
  • Market case · Russia

    A large Russian financial group AI financial assistant on in-house language models, beta since 20 July 2026

    The group built an AI financial assistant for tasks linked to money and shopping, on its own family of large language models. To avoid made-up answers, the system relies on verified data sources, not only on what the model has learned. The beta was presented on 20 July 2026, and the full launch is planned for the autumn.

    Banks and finance

    2 September 2026 · source: platforms.su

    Read the case
  • Market case · Russia

    Alfa-Bank, IT head local agent on the Qwen 3.6–35B-A3B model: about 70 tokens per second on a Mac

    A local AI agent on Mac does autonomous work on a model with 35 billion parameters, about 3 billion of them active per token. It runs through LM Studio at about 70 tokens per second. The agent first ran in Docker on the Mac, then moved to a virtual machine, and the author calls that move a turning point for ease of use.

    Banks and finance

    2 September 2026 · source: platforms.su

    Read the case
  • Market case · Russia

    A large Russian IT holding contract approval in 3 days instead of two weeks, document comparison costs 40% lower

    The legal department of the holding, more than 70 lawyers, trained its own language model on the company’s legal data for over a year. Contract approval fell from two weeks to three days. AI contract comparison of two versions of a document is five times faster than by hand.

    IT companies

    source: platforms.su

    Read the case
  • Market case · Russia

    NSIS, insurance information system operator requests handled 20 times faster on a local LLM, 1.5 times more requests with no new hires

    The operator of the national insurance information system deployed a local LLM, a large language model on its own servers, and built three uses on it. The first is automatic processing of requests with many fields. The second is a RAG service for technical support that answers from company documents. The third is AI agents in software development, including automatic code review.

    IT companies

    2 September 2026 · source: platforms.su

    Read the case
  • Market case · Russia

    Digital Sector and an IT infrastructure client AI contract analysis agent in the client’s closed environment; in the integrator’s own testing, agents freed 20% of team time

    The integrator built an AI agent for first-pass contract analysis inside the closed network of a client that designs and supplies IT infrastructure. The agent extracts and matches data, finds discrepancies and prepares the result for a specialist. The integrator also put agents into its own testing process, which freed 20% of the team’s working time per sprint.

    IT companies

    2 September 2026 · source: platforms.su

    Read the case
  • Market case · Russia

    VK Tech two models fit into 22.4 GB of 24 GB of video memory with a 65,536-token context

    The team built a local AI agent in a closed environment with Ollama and an agent layer, on a rented GPU server with one NVIDIA A30 card of 24 GB. The models were gpt-oss 20b (13 GB) for text and tools and qwen3-vl 8b (6.1 GB) for images; with the KV cache for a 65,536-token context, they used 22.4 GB.

    IT companies

    2 September 2026 · source: platforms.su

    Read the case
  • Market case · Russia

    LANIT requirements gathering 4 times faster, document search time 66% lower on a local LLM platform

    The group built a platform that runs large language models on the customer’s own infrastructure: a chat and complex pipelines for documents, audio and images. A subsidiary built its own model, trained further on Russian-language data. It pilots the model in CRM, SRM and HRM projects and in corporate chatbots.

    IT companies

    source: platforms.su

    Read the case
Contact us