Having your own language model does not yet make artificial intelligence sovereign. Training and running large models requires accelerators, high-speed memory, servers, and a software environment—and in these areas the Russian industry largely relies on foreign suppliers. A full-fledged replacement will not appear quickly: it will take years and major investment.

RACIB Executive Director Alexander Brazhnikov assessed the real limits of technological independence in the field of AI and identified the key bottlenecks in the industry.

What Russia can produce on its own

The most noticeable gap is in the production of advanced accelerators. According to Deputy Prime Minister Denis Manturov, the country serially produces microsystems using 300, 200, 130, and 90 nanometer process nodes. Moving to finer process technologies will require domestic chip production equipment.

The process technology alone does not determine processor speed. But a modern accelerator used to train large models also needs high-speed memory, complex chip packaging, and efficient data transfer. If even one component is lacking, the device loses capability.

Russia's own accelerators are being developed by the Module Research Center, the Elvis Research and Production Center, and HiTech. At the same time, a specialized neural processor and a general-purpose graphics processor solve different tasks and are not always interchangeable. SNDGLOBAL CEO Olga Kvashenkina estimated that a niche accelerator on a 28–65 nm process needs two to four years to go from a mature prototype to a serial product, and that is with guaranteed orders and state support. A world-class general-purpose GPU is an order of magnitude more difficult.

Even a finished chip does not close the issue. What is needed is serial production, compatible servers, stable component supplies, and software that makes the equipment work under real workloads.

There is a separate debate about which chip should be considered domestic. The Ministry of Industry and Trade discussed a localization system for accelerators for inclusion in the register of Russian radio electronics, and one expert allowed that a Russian architecture manufactured in China or Southeast Asia could become grounds for such status. Brazhnikov emphasized that the place of development and the place of production are different things. An original architecture builds engineering competencies, but dependence on foreign fabs, equipment, and materials does not disappear, and a formal entry in the register does not remove it.

Chinese chips are in short supply even in China

In the spring, Sberbank CEO German Gref said the bank expects to move GigaChat to Chinese microsystems. No signed contract or specific chip models were reported. At the same time, China's largest internet companies—ByteDance, Tencent, and Alibaba—were themselves hunting for Huawei Ascend 950 accelerators.

China is ramping up AI chip output but cannot yet fully cover domestic demand. In the first half of the year, 2.008 million data center accelerators were shipped in the country. Nvidia had 678,000 devices and 34% of the market, while Huawei had 572,000 and 28%. Nvidia shipments fell 44% year on year, while Huawei's grew. The market structure is also changing: 26% fewer graphics processors were shipped, at 991,000, while other AI accelerators were up almost 80%, at 1.016 million. The number of devices says nothing about their total power, since their architecture and purpose differ.

Chinese manufacturers planned to produce about 3.5 million AI chips in 2026 against demand of more than 5 million. Companies filled the shortfall with Nvidia gaming and professional video cards. ByteDance, Alibaba, and Tencent ordered more than 3 million Nvidia RTX PRO 5500 cards with a total value of more than $35 billion, and Nvidia did not guarantee delivery of the full volume. Such cards are suitable for inference, that is, generating answers from an already trained model, but they do not replace server accelerators when training large models.

For Russia, Chinese equipment is another supply channel. But its availability depends on manufacturers' capacity and the appetites of the Chinese market itself. The pivot to the East changes the structure of dependence rather than removing it.

The ecosystem holds tighter than hardware

Nvidia's strength is not only in chip specifications. Sberbank Deputy Chairman Alexander Vedyakhin pointed this summer to the importance of CUDA—the platform on which most tools for developing and running AI workloads are built.

Over the years, an ecosystem of libraries, drivers, compilers, and development tools has formed around CUDA. Moving to another architecture means rewriting part of the code, rechecking compatibility, and tuning performance for the new equipment. Therefore, comparing price tags on accelerators does not show the real cost of migration.

Chinese companies are trying to remove this barrier. In September, DeepSeek announced joint development with Huawei of tools for Ascend accelerators, including open libraries for computation and data exchange, and demonstrated a solution on 128 Ascend 950 accelerators. But the mere appearance of such tools does not prove that they deliver comparable performance across a wide range of tasks. Testing on real models and workloads is needed.

How much it costs to move to another platform

The price of a chip is only part of the expenses. Added to this are engineers' work, code adaptation, verification of computational correctness, possible performance degradation, and system maintenance. Cheap equipment may turn out to be more expensive to operate if more devices are needed for the same result or additional development is required.

A preprint by researcher Zheng Yu showed what this looks like in practice. The author ran two inference tasks on a system of 16 Huawei Ascend 910 accelerators with the CANN and vLLM-Ascend software stack. To make the module work stably, 12 edits had to be made to the source code. Some functions that speed up operation were disabled for the sake of computational accuracy, and control and recovery mechanisms were added to protect against failures. Throughput in one scenario reached its optimum at about four parallel request streams, while at six or more the queue saturated and errors and timeouts increased. Preparing the system for the first request took several minutes.

These are results for one specific configuration, and they cannot be transferred to all Huawei accelerators or compared directly with Nvidia. But the example shows the amount of work hidden behind the words "move to another platform."

Platforms should be compared on the same models and workloads across a range of parameters:

  • time and labor costs for porting software;
  • throughput, response latency, and result quality;
  • power consumption and the cost of processing a given number of requests;
  • failure frequency and maintenance costs;
  • the price of moving again if the system has to be migrated once more.

Standardized MLPerf Inference tests help with some of these questions. In version 6.1, released by MLCommons in September, an end-to-end RAG test and an agentic inference test appeared. But such benchmarks measure performance on given workloads, not the cost of porting the stack and total cost of ownership. There is no universal estimate in open data of the cost of migrating from Nvidia to Ascend or Russian chips, and without comparable tests, any averaged percentages and sums say little.

Dependence can be reduced without replacing chips

The need for computation can be reduced through compact specialized models and optimization of their operation. For parsing documents, classifying requests, or a single banking operation, a general-purpose model is not always needed. A narrow solution runs on less powerful equipment and costs less, while model compression and inference optimization help squeeze more out of existing accelerators.

Training large models will still require powerful infrastructure. The result depends on the task and on how much computational expense can be cut without losing the required quality.

Another path is software that supports multiple hardware platforms. If a model runs on different architectures without lengthy code rework, the customer has more freedom in choosing equipment. In such a scheme, Russian accelerators can cover individual applied tasks even if they do not become a universal replacement for world leaders.

Sovereignty as the ability to choose

Russia does not have to produce every component of AI infrastructure on its own. But its own models will not provide technological self-sufficiency as long as their training and operation depend on equipment whose supplies may be interrupted for external reasons. Chinese accelerators expand choice, Russian developments can occupy niches, and compact models reduce the load. None of these solutions alone eliminates dependence.

It becomes critical at the moment when a halt in supplies or a change in terms by one manufacturer puts the entire system at risk. The question is not about rejecting foreign components but about having a real alternative. Whether the industry can move to another platform and continue operating without excessive costs and serious loss of performance—that is the measure of sovereignty.

Cryptalist analysis: The case demonstrates that in AI infrastructure the barrier is shifting from silicon to the software ecosystem: even with chips available, migration costs and software maturity determine real independence. The accelerator market remains effectively monopolized at the CUDA level, so betting on alternative platforms is justified only with parallel optimization of models and multi-platform software. For investors, this is a signal that capital in the region will flow into applied software and niche solutions rather than attempts to catch up with leaders in general-purpose GPUs.