China includes 9 domestic AI accelerators in the first official registry of approved products for critical infrastructure — this is an effective ban on NVIDIA in the public sector. In parallel, the US has tightened export controls to a level where China’s DeepSeek was purchasing millions of NVIDIA GPUs through third countries. This is a structural rift in the AI infrastructure market, and it directly concerns Russia.
What is happening in the global AI infrastructure market is not a round of technological rivalry, but a fundamental shift that finally cements the trend toward de-globalization of critical technologies. We are witnessing how a multi-year dependence on a single supplier, in the face of NVIDIA, is being replaced by the forced construction of parallel, sovereign ecosystems.
This process, catalyzed by geopolitical tensions, turns artificial intelligence from an object of scientific research into a key element of national security and economic sovereignty. In this new landscape, a state’s ability to ensure its infrastructure with computing power independent of external political decisions becomes the main measure of technological and strategic potential.
Green Light for Chinese GPUs
The Cyberspace Administration of China (CAC) — the main cybersecurity regulator of the PRC — has published the first list of AI chips recognized as safe and reliable for critical infrastructure procurement. The list includes 9 products:
Huawei Ascend 910B (DaVinci) — Flagship for training and inference;
Huawei Ascend 310 (DaVinci) — Edge, IoT inference;
Moore Threads MTT S4000 (MUSA) — CUDA compatible, enterprise;
Moore Threads MTT S3000 (MUSA) — Budget segment;
MetaX (Xiyun) C500 (GPGPU) — Compute, HPC;
MetaX (Xiyun) C500 (GPGPU) — 64GB HBM2e, LLM inference;
Iluvatar BiV100 (Iluvatar Core) — Deep learning training;
Biren Technology BR100 (Biren) — Previously under BIS licenses;
Enflame CloudBlazer T20 (Enflame) — Training + inference.
Previously, Chinese state procurement formally did not prohibit NVIDIA, but practically discriminated against foreign products through «recommended catalogs.» Now the mechanism has become rigid: inclusion in the CAC registry equals access to contracts in the public sector, finance, telecom, and defense.
For vendors, this means:
Guaranteed demand — state contracts without competition from NVIDIA;
Production scaling — an incentive for SMIC and Samsung Foundry to ramp up output;
Ecosystem maturity — CANN, MUSA, and analogs receive massive workloads for debugging.
The Russian Ministry of Digital Development’s registry for domestic software operates on similar logic. The only difference is the scale. If in Russia we are talking about dozens of positions, in China the annual volume of AI infrastructure is measured in tens of billions of dollars.
This strategic move by Beijing to create a whitelist for its AI infrastructure is a direct response to the tightening of export controls by the US. If previously Chinese state structures only de facto preferred domestic solutions, now the mechanism has become truly rigid: inclusion in the CAC registry is the only key to the public sector, finance, and defense markets, which automatically creates guaranteed demand for national manufacturers. Such rapid institutionalization of support became necessary precisely after the BIS closed loopholes for sanctions evasion, betting on total control over re-export and extending bans to all Chinese legal entities, a lesson vividly taught by the DeepSeek scandal. As a result, Beijing is not just stimulating its vendors but is building a layered defense, striving to completely eliminate dependence on Western technologies in critical areas.
The US tightens control — the DeepSeek lesson
The US Bureau of Industry and Security (BIS) made clarifications to export controls following the DeepSeek scandal. The company purchased tens of thousands of NVIDIA H100 GPUs through intermediaries in Singapore and the UAE, bypassing direct restrictions.
New rules:
The ban applies to Chinese legal entities regardless of their country of registration;
Control over re-export through third countries is strengthened;
Licenses are required even for «indirect» supplies.
It is in response to this total control from the US, which has blocked all legal supply channels, that the golden hour has arrived for Chinese GPU manufacturers.
Three key factors:
Colossal domestic demand, backed by state investments in AI infrastructure;
Domestic chips reaching technological maturity and, as a consequence, a regulatory vacuum;
A unique moment created by US sanctions.
As a result, companies like Moore Threads, MetaX, and Huawei receive a guaranteed 3–5 year window of opportunity to scale up and capture the critical market without direct competition from NVIDIA, which is now completely cut off from this segment. Thus, external pressure has not just failed to stop the industry’s development, but has created greenhouse conditions for it, turning sanction pressure into the main growth driver.
Confluence of factors — the golden hour for Chinese GPUs
Three drivers converge into one moment:
Demand: Chinese state contracts ($50+ billion investments in AI infrastructure by 2027 according to the MIIT plan).
Supply: Domestic chips have reached a «good enough» level for inference and partially for training.
Regulatory: US bans block the alternative.
Moore Threads, MetaX, and Huawei Ascend get a 3–5 year guaranteed window to scale up without NVIDIA competition in the critical market.
What this means for Russia
Strengthening the positions of Chinese vendors Chinese GPU manufacturers now have a stable domestic market to amortize R&D. This reduces the cost price and makes export shipments to Russia more predictable in terms of price and delivery times.
Maturity of the software stack CANN 7.0 (Huawei) and MUSA SDK (Moore Threads) are being tested on millions of nodes in Chinese data centers. Russian integrators are getting more stable software than 18 months ago.
Risk of SMIC capacity shortage Simultaneous growth in demand in China + export orders = potential overload of production capacities. Recommendation for Russian buyers: plan purchases with a 3–6 month horizon.
Parallel evolution of sovereign stacks China and Russia are forming compatible but independent ecosystems:
China: CANN/MUSA → PyTorch/TensorFlow → LLM/VLM;
Russia: the same stack + localization (ARKA, GigaChat, Yandex GPT). This creates opportunities for cross-compatibility without binding to CUDA.
Price competitiveness vs NVIDIA By 2027, Chinese GPUs for inference (300+ tokens/sec on Llama 3 70B) will cost 2–3 times cheaper than NVIDIA analogs with comparable memory bandwidth.
Geopolitical hedge For Russian companies with critical infrastructure (finance, energy, public sector), the Chinese stack becomes the only real alternative to the aging stock of NVIDIA A100s and problems with servicing devices without official support.
Conclusions and Conclusion
The Chinese sovereign AI stack has moved from the phase of pilots and statements to the phase of industrial scale. The CAC state registry and the tightening of US sanctions are not disparate events, but two poles of one trend: the world of AI infrastructure is splitting into spheres of influence.
For Russia, this means:
Window of opportunity: 3–5 years while Chinese vendors are scaling up and ready for export;
Logistics risk: possible, direct supply channels need to be built, avoiding gray schemes;
Investments in localization: Platforms like ARKA are critically important for adapting Chinese hardware to Russian tasks.
The market is not waiting. Those who start migration now will gain an advantage in cost of ownership and independence. Those who wait for maturity will pay a premium for haste in 2–3 years.
In the long term, this split leads to the final formation of two incompatible technological paradigms, where the key factor for success is not so much the peak performance of the chip as the maturity and adaptability of the software stack. For Russian developers, this means the need to shift focus from simply acquiring «hardware» to deep integration of domestic software platforms, such as ARKA, with Chinese hardware. Creating a seamless, sovereign environment for AI model operation, independent of proprietary Western technologies like CUDA, becomes the main task determining the competitiveness of the national economy in the next decade. Success on this path will depend on the ability to form a unified ecosystem combining Russian LLMs (such as GigaChat and Yandex GPT) with Chinese silicon optimized for them.
Thus, we are not observing a crisis, but a powerful evolutionary push that is forever changing the rules of the game in the AI market. Sanction pressure and retaliatory protectionist measures have paradoxically accelerated the development of alternative solutions, turning the threat of isolation into a stimulus for accelerated growth. For Russia, this historical moment provides a chance not just to replace imported technologies, but to make a qualitative leap and occupy a niche in the new, multipolar technological reality. The main thing is to use the provided window of opportunity not to create temporary «patches,» but to lay the foundation for a truly independent and scalable digital infrastructure that will become the basis for future innovations.
Contact our experts to select the optimal hardware configuration and calculate rental pricing for your AI workloads.