Sberbank is looking to Chinese-made artificial intelligence chips to power GigaChat, Russia’s flagship domestic large language model, in a sign that Moscow’s AI ambitions are becoming increasingly dependent on Beijing’s semiconductor ecosystem as Western sanctions restrict access to advanced U.S. and European technology.

Chief Executive German Gref said the bank hopes Chinese microchips can be used for GigaChat, Reuters reported on May 20, citing comments made to Russian state broadcaster Channel One during President Vladimir Putin’s visit to China. Gref did not identify the specific chips Sberbank wants to purchase, but the statement places Russia’s largest lender directly inside one of the most strategically important bottlenecks in the global AI industry: access to high-performance accelerators capable of training and running large models at scale.

The move matters because Sberbank is not only a financial institution. Since Russia’s invasion of Ukraine and the subsequent tightening of Western sanctions, the bank has become one of the country’s most prominent technology platforms, investing heavily in AI tools, cloud services, digital assistants and enterprise software. GigaChat, launched by Sberbank in 2023 as a Russian-language rival to ChatGPT, has become a centerpiece of that strategy and a symbol of Moscow’s effort to maintain technological sovereignty in a field dominated by U.S. and Chinese firms.

The hardware challenge is acute. Large language models require vast clusters of specialized processors, high-bandwidth memory, advanced networking equipment and mature software stacks. Nvidia’s most powerful AI chips remain the benchmark for much of the global industry, but Russia’s access to those systems has been sharply limited by sanctions and export controls. That has pushed Russian companies toward alternative suppliers, workarounds and domestic development efforts, none of which fully match the scale or efficiency available to leading U.S. cloud operators or China’s largest technology groups.

Sberbank’s interest in Chinese chips comes as China’s own AI ecosystem is under pressure from U.S. export restrictions. Washington has restricted the sale of advanced Nvidia processors to China, encouraging Chinese companies to accelerate procurement from domestic suppliers such as Huawei while also seeking limited access to approved U.S. chips where possible. The result is a constrained and politically sensitive market in which Russian buyers may have to compete with some of China’s largest technology companies for supply.

Reuters has reported that major Chinese internet groups, including ByteDance, Tencent and Alibaba, have been moving to secure Huawei’s Ascend 950 AI chips after rising demand for domestic alternatives. Those companies are building and deploying models across search, cloud computing, advertising, productivity software, e-commerce and consumer apps, giving them both scale and strategic priority within China’s own technology policy framework. Sberbank’s prospective demand would therefore enter a market where supply is already tight and where Chinese national priorities may take precedence.

That competition is important for GigaChat’s trajectory. Sberbank has continued to upgrade the system, including the March introduction of GigaChat Ultra, which the company described as a next-generation AI assistant based on a new flagship model. Sberbank said that version improved the assistant’s ability to personalize communication, retrieve information and produce text responses faster. Those kinds of model improvements typically depend not only on better algorithms and data, but also on access to enough compute to train larger models, fine-tune specialized systems and serve users with acceptable latency.

The bank has also framed GigaChat as more than a consumer chatbot. Its AI tools are being embedded into banking, small-business services, internal operations and broader digital products. Sberbank has previously said AI-driven services are reaching millions of users across its digital channels, while Reuters reported in 2025 that the bank planned to introduce GigaChat models with reasoning capacity. That type of capability is especially compute-intensive because reasoning-oriented models often require more complex training runs and heavier inference workloads than simpler conversational systems.

For Russia, the dependence on Chinese chips reflects a broader structural vulnerability. The country has retained strong software and mathematical talent, but advanced semiconductor manufacturing requires a deep industrial base spanning chip design tools, fabrication equipment, process technology, packaging, memory, networking and data-center integration. Russia lacks the high-volume domestic manufacturing capacity needed to produce world-class AI accelerators at scale, making imports critical for frontier AI development.

China is the most plausible supplier, but it is not an unconstrained one. Huawei’s Ascend line has become central to Beijing’s effort to reduce reliance on Nvidia, and Chinese cloud providers are working to adapt their AI software stacks to domestic hardware. Still, Chinese chips have generally been viewed as trailing Nvidia’s most advanced processors in raw performance, ecosystem maturity and developer adoption. Even when chips are available, large-scale deployment requires compatible software frameworks, optimized kernels, stable networking, sufficient memory capacity and technical support.

Sberbank’s challenge is therefore not simply buying hardware. It must determine whether GigaChat can be efficiently trained or served on Chinese accelerators, whether existing model architecture and inference pipelines can be optimized for those chips, and whether the bank can obtain enough units to support commercial-scale deployment. AI projects can be constrained by the weakest part of the stack: insufficient chips, immature drivers, limited software compatibility, weak networking between servers or bottlenecks in data-center power and cooling.

The timing also reinforces how the AI race has moved from model announcements to infrastructure control. In the first wave of generative AI competition, public attention focused on chatbots and model capabilities. The next phase is increasingly defined by supply chains, chip availability, data-center capacity and export policy. Companies that can secure reliable compute can iterate faster, train larger or more specialized systems, and support wider commercial usage. Those that cannot may be forced into smaller models, more targeted use cases or dependence on foreign technology partners.

For Sberbank, GigaChat serves both commercial and national objectives. Commercially, the bank can use AI to automate customer service, support small businesses, improve internal productivity and expand beyond traditional banking. Strategically, GigaChat gives Russia a domestic platform less dependent on Western model providers, which may be inaccessible or politically unacceptable under current conditions. A successful transition to Chinese hardware would strengthen that sovereignty narrative, even if it deepens technological dependence on China.

The development also signals an emerging hierarchy within the non-Western AI market. China has the strongest alternative semiconductor base among countries facing U.S. technology restrictions, while Russia has model developers and state-backed demand but lacks comparable chip production. That creates a relationship in which Chinese suppliers may become infrastructure gatekeepers for Russian AI. The arrangement could benefit both sides politically, but commercially it may be complicated by limited supply, pricing power and Chinese companies’ own urgent compute needs.

Huawei is central to that equation. The company has spent years building an alternative to U.S.-origin AI hardware after facing American restrictions on its access to advanced semiconductors and manufacturing technology. Its Ascend processors and Atlas computing systems are part of a broader Chinese push to create domestic AI infrastructure. If Sberbank can obtain and integrate such systems, it would show that Chinese AI hardware is not only serving domestic technology companies but also becoming an export pathway for countries outside the Western technology sphere.

However, the economics remain uncertain. High-performance AI chips are expensive, and the total cost of ownership includes servers, networking, storage, power, cooling, software engineering and maintenance. Sberbank has the balance-sheet capacity and strategic backing to invest, but hardware constraints can still slow deployment. If Sberbank receives only limited quantities, it may prioritize inference for selected services rather than full-scale training of frontier-class models. If supply improves, GigaChat could be expanded across more enterprise and government-linked applications.

There is also a performance gap to consider. Reuters reported that Huawei’s most advanced chips remain behind Nvidia’s H200, one of the leading U.S. accelerators used for demanding AI workloads. A lower-performing chip can still be useful if paired with software optimization, model compression, distributed training techniques and sufficient scale. But the gap matters when competing with global model developers that have access to enormous Nvidia-based clusters and sophisticated cloud infrastructure. Sberbank may be able to improve GigaChat materially without closing the distance with top U.S. or Chinese models.

From a regulatory and geopolitical perspective, Sberbank’s move illustrates the limits and side effects of technology controls. Sanctions have constrained Russia’s direct access to Western hardware, but they have also increased the incentive for sanctioned entities to build alternative supply routes and deepen ties with China. At the same time, China’s own exposure to U.S. export controls has accelerated its domestic semiconductor ambitions. The result is a more fragmented AI hardware market, with parallel ecosystems developing across geopolitical lines.

For Western technology companies, the immediate commercial impact may be limited because Sberbank and many Russian entities are already outside normal Western supply channels. The broader significance lies in market structure. If Chinese chips become credible substitutes in sanctioned or restricted markets, they could gradually erode the global reach of U.S.-controlled AI infrastructure. If they remain supply-constrained or technically difficult to deploy, Nvidia and its ecosystem will retain a major advantage even where direct sales are politically restricted.

The announcement also lands during a period of intensifying China-Russia economic coordination. Putin’s visit to China provided the political backdrop for Gref’s comments, reinforcing the degree to which technology cooperation has become part of the broader bilateral relationship. For Russia, AI is tied to productivity, defense-adjacent capabilities, education, finance and state administration. For China, supporting Russian demand could expand influence, but it must be balanced against its own domestic compute shortage and the risk of further Western scrutiny.

Sberbank has not disclosed the scale of chips it hopes to secure, the timing of potential deliveries or whether it is negotiating directly with any named Chinese supplier. Those details will determine whether the initiative is a near-term infrastructure plan or a strategic signal. Without confirmed volumes, the market should treat the announcement as evidence of direction rather than proof that GigaChat will quickly receive a major compute upgrade.

Still, the direction is clear. Sberbank is trying to keep Russia in the generative AI race by aligning its flagship model with the most viable non-Western hardware supply chain available. That effort will test whether Chinese AI chips can support major foreign model developers while meeting heavy domestic demand. It will also test whether Russia’s AI ecosystem can adapt to hardware designed around China’s technical and industrial priorities.

The most likely near-term outcome is incremental rather than transformative. Chinese accelerators could help Sberbank expand GigaChat’s inference capacity, support more Russian-language services and continue model improvements despite sanctions. But the bank will still face constraints in chip availability, system integration and performance relative to the frontier systems used by leading U.S. and Chinese AI labs. In that sense, Sberbank’s chip search captures the central reality of the current AI market: model ambition is increasingly limited by compute access, and compute access is increasingly shaped by geopolitics.