On October 12, 2026, it became known that the government working group “Infrastructure and Data Centers for AI” had proposed temporarily allowing foreign graphics processors (GPUs) for AI tasks at critical information infrastructure (CII) facilities and in government systems—so long as there are no domestically produced equivalents with the required performance in Russian registries. The conditions: the Ministry of Industry and Trade confirms that no equivalent exists, usage is declared, and compensating security measures are coordinated with the FSB. This is only a proposal: according to CNews, the Ministry of Digital Development had not received it at that point. Until amendments are adopted, the current restrictions remain in force, and AI projects on CII must be planned accordingly.
Prepared by AIrassvet October 13, 2026 based on publications by CNews, RB.ru, and vc.ru (citing Forbes) and the text of Presidential Decree No. 166. It is intended for heads of IT and information security at CII entities—banks, energy, transportation, industrial, and telecom companies—as well as operators of government systems and integrators planning AI projects in a protected environment. This is an overview of the news and current rules, not legal advice: requirements for a specific facility depend on its significance category and regulatory decisions. AIrassvet is not affiliated with the agencies or equipment manufacturers mentioned; the section at the end of the article describes the publisher’s services.
Contents
- At a Glance: What Happened and What It Means
- What CII and Trusted Hardware and Software Systems Are
- What Rules Apply Now
- What the Working Group Is Proposing
- Arguments For and Against
- Which AI Tasks Really Need Top-Tier GPUs
- What a CII Entity Should Do Now
- Three Possible Scenarios
- Who This Does Not Affect
- Frequently Asked Questions
- Primary Sources and Verification Method
At a Glance: What Happened and What It Means
- Who proposed it: the working group “Infrastructure and Data Centers for AI” within the government subcommittee prepared a package of amendments; Forbes, CNews, and RB.ru reported on it on October 12, 2026.
- What is proposed: temporarily allow foreign GPUs and AI accelerators at significant CII facilities and in government systems if there is no equipment with the required specifications in Russian registries.
- Under what conditions: confirmation from the Ministry of Industry and Trade, declaration of use, and compensating information security measures agreed with the FSB.
- Which documents are affected: the law “On the Security of CII,” Presidential Decree No. 166, Government Resolution No. 719, and Government Resolution No. 1912.
- Status: this is a proposal from the working group. The Ministry of Digital Development said it needs to be coordinated with the relevant agencies, and as of October 12 it had not been submitted to the ministry.
- What to do now: plan projects under the current rules, inventory AI workloads based on power requirements, and prepare supporting justification in case the temporary procedure is introduced.
What CII and Trusted Hardware and Software Systems Are
Critical information infrastructure (CII) refers to information systems, networks, and automated control systems of organizations operating in healthcare, science, transportation, communications, energy, banking, and other financial market sectors, the fuel and energy complex, nuclear energy, the defense, rocket and space, mining, metallurgy, and chemical industries, as well as in real estate rights registration—and the communication networks between them. This is how Article 2 of Law No. 187-FZ “On the Security of Critical Information Infrastructure” defines it. A significant CII facility is one assigned one of the significance categories and included in the register of significant facilities; the strictest requirements apply to it.
A hardware and software system (HSS) is equipment and software that work together to solve one or more related tasks. A “trusted” HSS is one that meets established requirements, including being included in the unified register of Russian radioelectronic products. A graphics accelerator for training and running neural networks is also part of an HSS, and that is where the problem arises: according to the working group’s estimate cited by CNews, there are still too few import-substituting GPUs in the registry with performance sufficient for AI tasks on the market.
What Rules Apply Now
Three documents are most often mentioned in connection with the proposal:
| Document | What it establishes | Source |
|---|---|---|
| Presidential Decree of 03/30/2022 No. 166 | As of March 31, 2022, customers under Federal Law No. 223-FZ may not purchase foreign software, including as part of an HSS, for significant CII facilities without approval from the authorized body or the Bank of Russia. As of January 1, 2025, government bodies and customers are prohibited from using foreign software on their significant CII facilities unless otherwise provided by federal law. The Government is tasked with determining the timeline and procedure for transitioning to trusted HSS | kremlin.ru |
| Government Resolution of 11/14/2023 No. 1912 | The procedure for CII entities to transition to the preferential use of trusted HSS at significant facilities; according to CNews, it prohibits the use of untrusted HSS purchased as of September 1, 2024 | ConsultantPlus, CNews |
| Government Resolution No. 719 | Rules for confirming the production of industrial goods in Russia, which underpin the register of Russian radioelectronic products | CNews |
Decree No. 166 is cited according to the official text in the versions of the decrees of 11/22/2023 No. 887 and 04/07/2025 No. 214. For Resolution No. 1912, we verified the title and the version dated 12/26/2024; we present the content of the provisions on the acquisition date based on CNews’s description.
The practical takeaway for AI projects: at a significant CII facility, you cannot simply buy a server with foreign accelerators and run a model on it. Equipment and software must comply with trusted HSS requirements and import-substitution rules, and exceptions require approvals.
What the Working Group Is Proposing
According to CNews and RB.ru, the package includes five changes:
| Document | Proposed change |
|---|---|
| Law on CII Security | Introduce the term “computing resources for AI tasks” — equipment and software, including GPUs and specialized accelerators for machine learning, big data processing, and other AI tasks. When setting requirements for significant CII facilities, the Government should take into account the availability of sufficient capacity for AI |
| Decree No. 166 | An exception for computing capacity for AI—until Russian GPUs and accelerators suitable for large foundational models and no worse than the best foreign equivalents appear in the registries |
| Resolution No. 719 | If the registry does not contain equipment for sovereign large foundational models, non-registry products are permitted under three conditions: confirmation from the Ministry of Industry and Trade, declaration of use, and compensating security measures agreed with the FSB |
| Resolution No. 1912 | Changes due to compatibility issues: according to the group’s materials, foreign GPUs (in particular, Nvidia) require driver modifications when used with Russian operating systems in an OpenStack environment |
| Powers of the Ministry of Digital Development | The right to introduce a temporary procedure for using untrusted HSS for AI tasks, coordinated with the Ministry of Industry and Trade, for systems purchased after September 1, 2024, when no trusted equivalent exists |
As emphasized by vc.ru According to Forbes, the “loophole” is intended as a temporary measure: until Russia has comparable alternatives with similar capacity.
Status as of October 13, 2026. The Ministry of Digital Development told CNews that the proposals require a coordinated position from the relevant agencies and had not been submitted to the ministry as of October 12; when considering them, it is necessary to take into account the development of information security requirements for AI at critical information infrastructure (CII) facilities. The Analytical Center under the Government told Forbes that proposals from working group participants are first reviewed by the relevant agencies and only then, if approved, submitted to the subcommittee. In other words, this is not an adopted rule, but the start of the approval process.
Arguments For and Against
| Position | Who | Summary |
|---|---|---|
| In favor of temporary access | Working group | Without high-performance accelerators, it is impossible to train and run models in a secure environment, and there are not enough domestically produced GPUs of the required class in the registries |
| Against prioritizing AI over security | Ivan Pokrovsky, Executive Director of the Association of Russian Electronics Developers and Manufacturers (ARPE) | The proposal puts AI deployment ahead of CII security; the lack of trusted hardware and software systems should be a reason to fund development and push back deadlines, not to allow foreign platforms. Several companies produce Russian GPUs and NPUs; they should be compared not with Nvidia, but with the real needs of CII systems |
| In favor of combining components | Andrey Pankov, CEO of HighTech | Hardware and software systems for a trusted environment should already combine foreign and Russian components in 2026; there are no technical barriers to running foreign accelerators with Russian Linux-based operating systems; if access is granted, it is important to preserve priority for Russian developments |
| In favor of requirements based on specifications | Aleksey Grigoriev, Vice President of Engineering at Sitronics Group | Regulatory requirements are logically formulated through technical specifications rather than the manufacturer; foreign GPUs are needed primarily for the most compute-intensive tasks |
Expert positions are cited from CNews and vc.ru.
Pay attention to the discrepancy in compatibility assessments: the working group materials mention technical issues with drivers in OpenStack, while the HighTech representative says there are no technical barriers for Linux-based operating systems. For a specific project, this means one thing: compatibility must be tested on your own hardware, OS, and virtualization stack, not taken on faith.
What AI tasks actually need top-tier GPUs
The most practical point in the discussion is the distinction proposed by Aleksey Grigoriev of Sitronics Group: training large models in Russia already requires thousands of GPUs, while applied tasks need dozens or hundreds of accelerators. According to him, foreign GPUs are needed primarily for training and fine-tuning large foundation models, high-performance inference, and AI and HPC clusters, but not for all AI systems. He also emphasizes that the entire system matters: accelerators, memory, interconnects, networking, storage, cooling, and cluster stability.
We turned this into a practical classification for planning. This is an editorial framework, not a regulatory one:
| Task class | Examples at a CII facility | Capacity requirements | What to check first |
|---|---|---|---|
| Training and fine-tuning large models | An industry-specific foundation model | Maximum, cluster-scale | Usually the model developer’s task, not the CII operator’s |
| High-performance inference | A model for thousands of concurrent users, long context | High | Throughput, memory for KV cache, fault tolerance |
| Applied inference | Search through regulations, document analysis, dispatcher assistant | Medium | Whether certifiable equipment is sufficient for a compact model |
| Traditional ML and computer vision | Failure prediction, quality control, video analytics | Often moderate | Whether it can run on available trusted hardware and software systems |
For many applied scenarios at a CII facility, choosing a compact model that solves the task with the required quality reduces dependence on the most powerful and scarce hardware. We wrote about how to choose a model based on available resources in the article “Local LLMs in 2026: Models, Hardware, and Cost”.
What a CII operator should do now
Until the proposal is adopted, it makes sense to act so the project can be implemented regardless of the outcome:
- Define the system status. Which systems where AI is planned are classified as significant CII facilities, and what is their significance category? The requirements for non-significant facilities and systems outside CII are different.
- Inventory AI tasks by class from the table above: training, high-performance inference, applied inference, traditional ML. For each one, estimate the required capacity based on the pilot, not the largest possible model.
- Check trusted options. Are there hardware and software systems and Russian accelerators in the registries that cover your task class, and what do tests on your model and data show?
- Separate the environments. Tasks that do not process data from a significant facility can sometimes be moved outside it—for example, model training on anonymized data. This decision should be made together with the information security team.
- Prepare the rationale in advance. If a temporary procedure appears, the proposal will require confirmation that no equivalent exists, a declaration, and compensating security measures. Technical requirements, test results for trusted options, and the threat model will be useful in any case.
- Check stack compatibility. Test the accelerator, OS, drivers, and virtualization platform in a lab before procurement.
- Monitor the status. Rely on published regulations and the positions of the Ministry of Digital Development, the Ministry of Industry and Trade, and the FSTEC, not on retellings.
Three development scenarios
This is our forecast based on publications, not information from government agencies:
| Scenario | What will happen | What it means for projects |
|---|---|---|
| The proposal is adopted almost as written | A temporary procedure will appear with approval from the Ministry of Industry and Trade, a declaration, and protection measures coordinated with the FSB | High-performance AI workloads on critical information infrastructure will become feasible, but with a long approval cycle |
| Partially accepted | For example, only the concept of “computing resources for AI tasks” and accounting for capacity when setting requirements | The restrictions will remain, but a basis for future exceptions will appear |
| Rejected | A focus on funding Russian accelerators and adjusting AI rollout timelines, as proposed by ARPE | Projects on critical information infrastructure are built on trusted hardware-software systems and compact models |
In any scenario, companies that have already separated their AI tasks by power requirements and tested available trusted options will come out ahead.
Who this does not apply to
The proposal applies to significant critical information infrastructure facilities and government information systems. If your company is not a critical information infrastructure entity or the AI system is not tied to a significant facility, the rules for purchasing and using equipment for it are determined by other requirements — for example, personal data legislation. For such businesses, the news is indirectly important: it shows how scarce high-performance accelerators are, and that affects the availability and pricing of computing capacity. We wrote about the general rules for business after the AI law was passed in the article “AI Law 243-FZ: What Businesses Should Do”.
Frequently asked questions
Have foreign GPUs been allowed on critical information infrastructure sites?
No. As of October 13, 2026, this is a proposal from the working group “Infrastructure and Data Centers for AI.” According to CNews, the Ministry of Digital Development had not received it as of October 12, and it still needs to be coordinated with relevant agencies.
Under what conditions are foreign accelerators being proposed for admission?
If there is no equipment with the required characteristics in Russian registries, the Ministry of Industry and Trade confirms that fact, the use of foreign products is declared, and compensating information security measures are coordinated with the FSB. The exception is intended to be temporary — until domestic alternatives with comparable performance appear.
What does Decree No. 166 prohibit?
Since March 31, 2022, purchasers under Federal Law No. 223-FZ have not been allowed to buy foreign software, including software integrated into hardware-software systems, for significant critical information infrastructure facilities without approval. Since January 1, 2025, government bodies and purchasers have been prohibited from using foreign software on their significant critical information infrastructure facilities unless otherwise provided by federal law.
What is a trusted hardware-software system?
This is hardware and software working together to solve similar tasks and meeting established requirements, including being included in the unified register of Russian radio-electronic products. The procedure for transitioning to them for significant critical information infrastructure facilities is defined by Government Resolution No. 1912.
Do all AI systems on critical information infrastructure need top-tier foreign GPUs?
According to Sitronics Group, no: they are needed primarily for training and fine-tuning large foundation models, high-performance inference, and large clusters. Applied tasks need dozens or hundreds of accelerators, and some of them can be handled by compact models on available equipment.
Does this apply to companies that are not critical information infrastructure entities?
Directly, no: the proposal concerns significant critical information infrastructure facilities and government systems. Other companies are subject to different requirements, such as the personal data law, and the news matters as a signal of a shortage of high-performance accelerators.
Primary sources and verification method
We reviewed the primary sources on October 13, 2026. The content of the working group’s proposals, the positions of the Ministry of Digital Development, the Analytical Center, and experts are presented according to CNews, RB.ru, and vc.ru, which cite Forbes; the text of the amendments itself has not been published, and we have not seen it. The provisions of Decree No. 166 were checked against the official text on kremlin.ru; for Resolution No. 1912, the title and version were verified. The classification of AI tasks, action plan, and scenarios are AIrassvet’s editorial methodology.
- CNews, 10/12/2026: Foreign graphics processors will be allowed for use on critical infrastructure and government systems in Russia due to the lack of domestic ones — list of amendments, positions of the Ministry of Digital Development and experts.
- RB.ru, 10/12/2026: Russian developers may be allowed to use foreign chips for AI — conditions for approval.
- vc.ru, 10/12/2026: A government working group proposed allowing the use of foreign chips and AI solutions at critical facilities — retelling of Forbes, temporary nature of the exception.
- Federal Law of 07/26/2017 No. 187-FZ: “On the Security of the Critical Information Infrastructure of the Russian Federation” — definitions of critical information infrastructure, entities, and significant facilities (Article 2).
- President of Russia: Decree No. 166 of 03/30/2022 — bans on purchasing and using foreign software on significant critical information infrastructure facilities.
- ConsultantPlus: Government Resolution No. 1912 of 11/14/2023 — procedure for transitioning to trusted hardware-software systems.
Limitations: the text of the amendments has not been published, and details may change during coordination; expert opinions reflect the positions of their organizations; this article is not a substitute for advice on categorization and requirements for a specific facility.
How AIrassvet helps plan AI in a protected environment
If you are planning an AI project on a site with strict equipment and data requirements, the main risk is choosing a model and architecture for which no permissible computing capacity can be found. This should be checked before procurement. On the AIrassvet website the following areas are described that are directly relevant to this task:
- Deployment and fine-tuning of language models on the customer’s servers, including inference optimization for available equipment.
- RAG systems and enterprise knowledge bases for practical use cases such as searching regulations.
- Computer vision solutions and ML forecasting for production processes.
- Process auditing, data preparation, MVP development, and testing before scaling.
The first step is to choose one AI task, describe the current process, the data and its security requirements, equipment constraints, and the criterion by which you will consider the pilot result sufficient.
Bottom line: plan under the current rules, prepare for an exception
The working group’s proposal shows that the shortage of high-performance domestic accelerators has been acknowledged at the government level, and for AI on critical information infrastructure sites, a temporary procedure with strict conditions may appear. But for now, this is only the beginning of coordination, and the existing restrictions of Decree No. 166 and Resolution No. 1912 remain in force.
The practical strategy is to divide AI tasks by power requirements, check available trusted options and compact models for applied scenarios, and prepare technical justifications in advance. Then the project will be feasible both if the restrictions remain and if an exception is introduced.
