LLM inference is the process of running a trained language model to answer prompts: the model receives text and returns a result. A Russian company can set it up in six ways: pay per token for a proprietary model, pay per token for an open model in the cloud, rent a dedicated endpoint, run its own or fine-tuned weights on a provider’s infrastructure, rent GPUs and assemble everything itself, or install a server inside its own environment. For getting started and variable workloads, pay-per-token is usually more cost-effective; for a steady high volume and strict data requirements, dedicated inference or an on-prem environment is better. The exact threshold is determined by calculations based on your requests.
Prepared by AIrassvet using open sources, verified on October 9, 2026. It is intended for business owners, IT directors, and digital project leaders who are choosing where and how to run language models. Provider pricing and model catalogs change often, so specific rates are not listed in the article. AIrassvet is not affiliated with the providers mentioned, and the final section describes the publisher’s services.
Contents
- Why this became a hot topic in October 2026
- Quick overview: six inference formats
- Pay per token: proprietary and open models
- Dedicated inference and your own weights
- GPU rental and your own server
- When to move from token-based pricing to dedicated inference
- What to check with the provider and in the contract
- How to choose a format: step-by-step process
- Common mistakes when choosing
- Frequently asked questions
- Primary sources and verification method
Why this became a hot topic in October 2026
October 8, 2026 CNews reportedthat Selectel updated its AI platform and added two services:
- Token Factory — access to the open-source models GLM-5.3-Flash, DeepSeek-V4-Flash, Qwen3.8-Flash-Next, and Qwen3.8-27b deployed on the provider’s infrastructure through a single API with pricing in rubles for actual token usage. According to the company, data is stored and processed in Russia on infrastructure compliant with Federal Law 152-FZ.
- Custom Inference — running models from the catalog or your own models, including those fine-tuned on corporate data, on dedicated provider infrastructure.
According to the report, the same platform also includes a previously launched AI router with access to more than 400 international models through a single API, dedicated inference with an isolated endpoint compatible with the OpenAI API, and a marketplace of ready-made solutions.
On the same day, CNews published a study by Selectel and the analytics firm Apple Hills Digital on the cloud AI services market:
- total consumption of cloud AI services in Russia in 2025 — RUB 65.5 billion, of which RUB 40.8 billion (62%) was internal consumption within provider ecosystems and RUB 24.7 billion was market consumption;
- GPU infrastructure in the public cloud — RUB 8.2 billion, up 68.7% from 2024; AI/ML platforms — RUB 9.4 billion, up 97%;
- according to a Selectel representative, over the first nine months of 2026, the company’s GPU server consumption grew 2.6 times compared with the same period last year, with most demand coming from large enterprises;
- an Apple Hills Digital analyst noted that the cloud AI services market in Russia is currently constrained more by supply than by demand.
The study was prepared with the participation of the provider that sells these services itself; the figures should be read with that in mind.
Quick overview: six inference formats
| Format | How it is billed | Control over the model and data | Best for |
|---|---|---|---|
| Proprietary model API | Per token | The developer sets the model and terms; data is sent to the developer | Fast launch, use cases where a strong model matters |
| Open model via API (tokens) | Per token | Choice from the provider’s catalog; data stays with the cloud provider | Pilots, variable workload, requirement for processing in Russia |
| Dedicated endpoint for a catalog model | Per runtime of dedicated resources | Isolated endpoint, predictable speed | Steady request volume, latency requirements |
| Your own or fine-tuned weights on provider infrastructure | For resources | Your own model without your own infrastructure | Companies with a fine-tuned model |
| GPU rental (virtual machine or dedicated server) | Per rental period | Full control over the stack; all operations are on you | Teams with ML engineers, nonstandard models |
| Own server inside the company environment | Capital expenses and operations | Maximum control, data never leaves the environment | Strict security requirements, large stable volume |
The lower the row, the more control you get and the more work falls on the company. For a detailed comparison of cloud and on-prem environments by total cost of ownership, see the article “Cloud or On-Prem LLM: How to Compare TCO”.
Pay per token: proprietary and open models
Pay per token is the fastest way to get started: there is no need to rent servers, and you only pay for the text processed. The difference between a proprietary model and an open model in this format is mainly about control.
- A proprietary model gives access to strong commercial models, but its behavior, pricing, and availability are determined by the developer. If it is a foreign model, requests involving personal data raise the issue of cross-border data transfer — covered in the article “Federal Law 152-FZ and Foreign Neural Networks”.
- An open model from a Russian cloud provider runs on its infrastructure. If you later need to move to dedicated inference or your own server, the same open-weight model can run there too — prompts and tests migrate with fewer losses.
The weak points of this format are that a sudden spike in traffic can lead to request limits, response speed depends on overall service load, and costs rise linearly with volume. For Russian text, it also matters how the model’s tokenizer breaks up words — more on that in the article “Why Russian Text Is More Expensive in LLMs”.
Dedicated Inference and Your Own Weights
Dedicated inference is a model running on resources reserved exclusively for your company, with a separate endpoint. You pay for the runtime of the resources, not for tokens, so costs do not increase with every request, but they also do not drop when there are no requests.
When this format makes sense:
- the request flow is stable and high — for example, an agent processes incoming documents throughout the workday;
- you need predictable response latency that is not affected by other provider customers;
- you need isolation: a separate endpoint is easier to describe in a threat model and approve with security;
- the company has a fine-tuned model that is not available in the public catalog.
The option of “your own weights on the provider’s infrastructure” removes the need to build and maintain a cluster. But fine-tuning itself is a separate project involving data and quality validation; for when you need it and when knowledge base search is enough, see the article “RAG, Fine-Tuning, or Prompt Engineering: Which Should You Choose?”.
GPU Rental and Your Own Server
GPU rental gives you full control: you choose the model, inference engine, quantization, and settings yourself. The price of that control is operations: updates, monitoring, scaling, on-call coverage. According to a K2Tech study published by CNews on October 5, 2026, among large industrial companies, 27% named insufficient computing capacity as the main constraint, 24% rent GPUs from third-party providers, 21% combine their own capacity with the cloud for peak loads, and another 21% have their own GPU clusters with spare capacity.
An on-premises server inside the company’s environment makes sense when data cannot be sent even to a Russian cloud provider, or when the request volume is large and stable for years ahead. For guidance on choosing models and hardware for this option, see the review “Local LLMs 2026: Models, Hardware, and Cost”.
When to Move from Tokens to Dedicated Inference
The decision should be based on calculation, not intuition. A simple monthly comparison formula:
| Metric | How to Calculate |
|---|---|
| Token-Based Pricing Cost | (input tokens × input token price + output tokens × output token price) per month |
| Dedicated Inference Cost | hourly price of dedicated resources × operating hours per month + team setup costs |
| Break-Even Point | the token volume at which the two totals are equal |
| Load Headroom | peak number of simultaneous requests / throughput of the dedicated endpoint |
Use actual request logs from the past 2–4 weeks as input for the calculation, not estimates. Keep in mind that dedicated resources can be turned off at night only if the process allows it, and that if load is estimated incorrectly, the dedicated endpoint will either sit idle or fail to handle peaks.
At the X5 Group session on October 9, 2026, according to CNews’ retelling, representatives of large companies said much the same thing: the head of Skala^r said the company introduced metrics for token cost and pipeline execution, while T-Bank’s CTO said the bank is looking for the right balance between buying tokens and deploying models. For a small or mid-sized business, the takeaway is the same on a smaller scale: measure the cost of one completed request or task, not the token price.
What to Check with the Provider and in the Contract
- Where data is processed. A site in Russia, compliance with Federal Law No. 152-FZ, and a data processing agreement if personal data is transferred.
- Request storage. Whether the provider stores requests and responses, for how long, and for what purpose; whether they are used for training.
- Load limits. Request and token limits per minute, and how to increase them.
- API compatibility. Compatibility with the OpenAI API makes it easier to switch providers, but model behavior still needs to be checked on your data.
- Fault tolerance. How many availability zones the service runs in, what is promised if a site goes down, and how quickly a dedicated endpoint can be moved.
- Catalog stability. How the provider notifies customers about removing a model from the catalog and how much time they give for migration.
- Transition provisions under 243-FZ. Under Federal Law No. 243-FZ starting March 1, 2027, the government will determine cases in which only sovereign or national models are permitted. For systems operating as of that date, a transition period applies until September 1, 2032, provided that data is processed and stored in Russia. If your process could fall under such a list, it is worth documenting the data processing location now.
How to Choose a Format: Step-by-Step Process
- Describe the use case. What task, how many requests per day, what data comes in, and what class it belongs to.
- Build a test set. 50–100 typical requests with reference answers for quality checks.
- Check quality on two or three token-based models. Start with token-based pricing: it is the cheapest way to test the hypothesis.
- Measure actual volume. Two to four weeks of request logs will provide the numbers for the formula above.
- Compare formats using the formula and data requirements. If processing must happen only within the company’s environment, some formats are ruled out immediately.
- Build in a backup option. A second model or second provider for critical scenarios, with a routing layer between them. More on that in the article “An AI Model Router for Business”.
Common Mistakes When Choosing
- Comparing token price instead of task cost. Different models use different numbers of tokens for the same Russian text and produce answers of different lengths.
- Buying dedicated resources right away. Without measuring real load, it is easy to overpay for an idle endpoint.
- Choosing a model based on general rankings. Quality on your documents and questions may differ from public benchmarks.
- Ignoring operations. GPU rental looks cheaper on paper until engineer hours are counted.
- Keep everything with one provider. If the platform goes down or the model is removed from the catalog, the entire process stops.
Frequently Asked Questions
What is LLM inference?
It’s the process of running a trained language model to answer requests: the model receives text and returns a result. Inference differs from training in that the model weights do not change, and the main costs are tied to computation for each request.
How is an open model via API different from a proprietary one?
An open model with publicly available weights can be run with different providers or on your own server, so switching between formats is easier. A proprietary model is available only from its developer or partners, and the terms are set by the developer.
When should you move from per-token pricing to dedicated inference?
When request traffic is steady and high, and you need predictable latency or isolation. The threshold is calculated from a log of real requests: monthly token costs are compared with the cost of dedicated resource hours and team operations.
Can you run an LLM in the cloud with personal data?
Yes, if the provider processes the data in Russia, complies with Federal Law 152, and you have a data processing agreement in place. For especially sensitive data, companies often choose a dedicated endpoint or their own server.
Do you need your own ML engineering team for cloud inference?
For per-token pricing and a dedicated catalog endpoint, usually no—a developer for integration is enough. For GPU rental and your own server, you need specialists who can configure the inference engine, monitoring, and updates.
How does Federal Law 243 affect the choice of inference format?
Starting March 1, 2027, the government will define cases where only sovereign or national models are allowed. For systems operating on that date, the transition period through September 1, 2032 applies provided data is processed and stored in Russia, so it makes sense to document the processing location now.
Primary sources and verification method
Checked on October 9, 2026:
- CNews, October 8, 2026 — Selectel’s updated AI platform: Token Factory, Custom Inference, AI router, dedicated inference, marketplace.
- CNews, October 8, 2026 — a Selectel and Apple Hills Digital study of the cloud AI services market in 2025 and scenarios through 2030.
- CNews, October 9, 2026 — an X Holding session on enterprise AI infrastructure and token economics.
- CNews, October 5, 2026 — a K2Tech study on compute constraints facing industrial companies.
- ConsultantPlus, Federal Law 243 — rules on sovereign and national models and the transition period.
Service details are based on the provider’s press release as retold by CNews and were not verified through testing. The format table, comparison formula, selection process, and list of issues are an editorial analysis by AIrassvet; the article does not include its own measurements or project results.
How AIrassvet helps you choose and deploy a language model
If you are choosing where and how to run a language model for your workflows, the task can be broken down step by step. On the AIrassvet website, which also uses the name “AI Studio by Vladimir Lomtev,” the following relevant directions are described:
- Use-case audit and data preparation: request volume, data categories, and a benchmark set for evaluating model quality.
- Integrating models, including domestic ones, through a unified layer so you can compare them on your tasks and switch without rewriting the application.
- Deploying a language model on the company’s servers or in a secure cloud when data cannot be sent to external services.
- Fine-tuning an open model on company data when knowledge base search is not enough.
The first step is to choose one process, define its current metrics, data sources, constraints, and the criterion by which you will consider the selected format a fit.
Bottom line: choose the inference format based on the numbers
The Russian cloud AI services market is growing fast, and providers are offering more and more formats—from per-token pricing for open models to dedicated endpoints and running your own weights. That expands your options, but it doesn’t replace the need to do the math.
Start with per-token pricing on open models with data processed in Russia, build a benchmark set and a log of real requests, and then compare the cost per task across formats. Moving to dedicated inference or your own server makes sense when the numbers, data requirements, and your team’s readiness to operate it all point in that direction.