Working group proposes allowing foreign GPUs in Russian government systems and critical infrastructure

A Russian working group within a government subcommission has proposed admitting foreign AI solutions to trusted hardware and software systems (PAK). These systems run significant critical information infrastructure (CII) facilities and government information systems. According to CNews, the proposal concerns GPUs, and the motive is stated plainly: there are no domestic accelerators that could replace foreign ones. No decision has been made yet. The Ministry of Digital Development stresses that the proposal must be coordinated with all relevant agencies and must not weaken cybersecurity requirements.

What exactly the working group proposed

The wording is short, but it carries a lot of weight. Until now, Russian CII regulation has rested on import substitution: Federal Law No. 187-FZ on CII security, and Decree No. 166, which since 1 January 2025 has restricted purchases of foreign software at significant facilities. Now one of the working groups acknowledges that, at the level of compute hardware for AI, this logic has hit a ceiling.

Here is what the published information says:

• initiator: a working group of the government subcommission;
• subject: foreign AI solutions, including GPUs, in trusted PAK;
• where they would be used: significant CII facilities and government information systems;
• main argument: there are no domestic equivalents;
• stated risk: compatibility of foreign hardware with domestic software;
• Ministry of Digital Development position: agency coordination and compliance with cybersecurity requirements are required;
• status: a proposal, not a rule. The announcement gives no timelines and no list of permitted solutions.

Why the problem comes down to GPUs

Training and inference of modern models depend on accelerators with high memory bandwidth and a mature software stack. Russia has neither mass-produced GPUs of that class nor an ecosystem comparable to CUDA. Domestic processors and accelerators exist, but public-sector tasks, from document analysis to processing large datasets, increasingly require AI workloads specifically.

If an agency builds a language-model system inside a trusted perimeter, it needs hardware. In practice, giving up foreign accelerators means giving up the projects themselves or falling back to weaker solutions. The working group essentially acknowledges this choice and proposes a managed exception instead of a ban.

Compatibility: the proposal's weak spot

The group itself points to the problem of compatibility between foreign hardware and domestic software. This is not a formality. Russian operating systems, virtualization tools and security products are certified for specific configurations. Drivers and libraries for foreign accelerators are usually not tuned for them. Every driver update can change system behavior, and any closed binary component inside a trusted perimeter raises questions for regulators.

The second layer of the problem is security. Foreign accelerators ship with proprietary firmware, which cannot be verified the way source code can. That is why the Ministry of Digital Development calls cybersecurity a mandatory condition. The announcement does not say which verification procedures are possible for closed hardware.

What customers and integrators should do

While the initiative is not yet a rule, it is too early to buy hardware for it. But you can start preparing now.

1. List the AI workloads in your trusted perimeters and split them into those that can run on domestic hardware and those that do not work without GPUs. 2. Check which configurations your OS and security products are certified for. That is where incompatibility will surface. 3. If a project depends on foreign accelerators, build a fallback into the plan in case approval is denied. 4. Watch for decisions from the Ministry of Digital Development and related agencies: the verification requirements they set for firmware will determine the cost of deployment.

The main question remains open. Will admitting foreign GPUs be a temporary measure with a clear end date, or will it become permanent? The answer determines whether Russian accelerator developers have an incentive to catch up, and a market in which it makes sense for them to compete.