Summary of Key Points
Russia may seem "invisible" in the global AI competition, but its domestic tech giants (Yandex and Sberbank) have established a self-sufficient AI ecosystem with practical models and applications. However, due to issues such as brain drain, limitations in computing power chips, and a small market size, Russia cannot compete at the top level in general AI. As a result, it has shifted its focus to military AI and "sovereign large models," adopting a differentiated development strategy.
I. Silence Does Not Mean Absence: The Local AI Ecosystem is Growing Quietly
Russia is not falling behind; it just doesn't stand in the spotlight. Two major players are supporting the local AI ecosystem:
- Yandex (the Russian version of Google): Has released YandexGPT Pro 5.1, which performs better than international high-end commercial models in Russian-specific tasks (such as local searches and content generation) in over 60% of cases. It has also developed a multimodal tool called Shedevrum that can create images and videos, as well as an intelligent assistant named Alice Pro. The Yandex browser, which includes Alice Pro, has 71 million monthly active users and ranks among the top ten mobile browsers globally.
- Sberbank: Has created GigaChat 3.5 Ultra, which uses a "hybrid expert architecture" to improve efficiency significantly (up to four times faster in long-text generation), as well as an image/3D generation model called Kandinsky.
These technologies have been applied in practical fields such as medical CT diagnosis assistance, financial risk management (fraud prevention), and 40% of its customers use intelligent voice systems. Western sanctions have actually helped, as Russian AI products were previously excluded from the market, allowing local solutions to fill the gap within two years.
II. Talent Retention: A Challenge
Russia was once a rich source of AI talent:
- Its mathematics education is world-class, with consistent victories in international math competitions. It has a strong foundation in computer vision and algorithms; there's even a saying in Silicon Valley that "for complex math problems, turn to Russian engineers." Ilya Sutskever, former chief scientist at OpenAI, comes from the former Soviet Union.
- However, it struggles to retain talent: Top algorithm experts in Moscow earn only a few hundred thousand dollars per year, while similar talents in Silicon Valley earn much more and receive equity (for example, early employees of OpenAI have seen their stock options appreciate significantly).
As a result, Russia has become a "training ground," with its talented individuals moving to large companies in the UK and US.
III. Computing Power: A Bottleneck
AI model training requires massive amounts of high-performance computing power, which is a major obstacle for Russia:
- Its supercomputers are outdated: The supercomputers built before sanctions (such as Yandex's Chervonenkis) used NVIDIA A100/V100 GPUs, which are now out of the top 100 in global rankings compared to the thousands of GPUs used in Silicon Valley and China.
- Chip supply shortages: Western bans on latest AI chips (e.g., NVIDIA H100) have affected Russia. Although it tried to purchase Huawei's Ascend 950PR, demand from Chinese companies is high, and local semiconductors can only produce 180-90 nanometer chips (while current AI chips require 3-5 nanometers). Russia plans to start producing 28-nanometer chips by 2030, which is several generations behind.
Scattered chip supplies are sufficient for small-scale models, but not for large-scale training—this solves the problem of "having" chips, not "enough" chips.
IV. Changing Course: Focusing on Military AI and Sovereign Large Models
Since Russia cannot compete in general AI, it has chosen two more practical approaches:
- Military AI: Applying AI to military applications such as autonomous drone swarms for target identification, real-time enemy recognition from satellite images, electronic warfare interference, and battlefield decision support. These areas do not require the same level of investment as general AI and leverage Russia's military strengths.
- Sovereign Large Models: Legislation requires that "sovereign AI" be developed by domestic companies with data and servers located in Russia, emphasizing data security (to prevent sensitive information from being shared abroad). This approach is being marketed to BRICS countries as an alternative solution, particularly those concerned about data breaches. Although its general capabilities may not match GPT, security and compliance are key selling points.
V. New Rules of the AI Race: It's About the Entire Industrial System
In the past, AI relied on individual geniuses (e.g., the AlexNet model was developed by a team of just a dozen people in 2012, and the Attention paper had only eight authors in 2017). Today, training top-tier models requires hundreds of thousands of GPUs, nationwide data centers, high-quality electricity, and billions of dollars in investment—similar to building a high-speed train, which depends on the entire industrial infrastructure of a country.
What Russia lacks is this comprehensive industrial system. It has the technology and talent but lacks the capital, chips, and large market (with only 200 million Russian speakers, compared to the millions in English-speaking countries). Therefore, it cannot compete in general AI. It hasn't withdrawn from the AI arena; rather, it has chosen a different path that suits its capabilities and circumstances.
In summary: Russia's AI efforts are not ineffective; it simply cannot afford the resource-intensive and globally integrated nature of general AI. Instead, it focuses on areas where it has advantages, such as military applications and sovereign large models. This highlights that in the AI era, success requires a complete industrial ecosystem, beyond just technology and talent.