The Russian AI market has officially moved past the phase of naive, back-of-the-envelope experimentation. It is time to face the facts: the chatbot hype has been replaced by cold, hard calculation. A study by Selectel and Apple Hills Digital confirms that nearly half of companies have already migrated their AI workloads to the cloud or plan to do so within the next 12 months. In a reality where top-tier hardware becomes obsolete in just two or three years, building an on-premise circuit is a gamble for the ultra-wealthy or the reckless; only 11% of respondents have opted for this path.
The Inference Economy: A Strategic Shift
Businesses have become noticeably more discerning. While the cloud was previously seen merely as a way to rent scarce GPUs, priorities have shifted. Access to specific accelerators is now critical for only 17% of customers. Far more important is the availability of pre-trained models (23%) and mature data management tools. The most popular strategy is the deployment of open-source solutions, chosen by 37% of companies. This approach allows for data security control and inference cost management without falling into the trap of proprietary system lock-in.
"The cloud allows businesses to rapidly scale successful use cases and test different GPU types without the risk of getting stuck with redundant hardware," explains Alexander Tugov, Director of the AI Vertical at Selectel.
The Infrastructure Maneuver: From Pilots to Budgets
The mechanics of this shift are clear: companies are no longer buying "digital promises"; they are investing in flexibility. Nearly half of the market has already allocated real budgets for cloud-based AI, ensuring a surge in demand through 2024–2025. At the application level, this is manifesting as a transition from simple scripts to full-fledged AI agents and fine-tuned language models—each currently used by 27% of survey participants.
The only remaining question is whether domestic providers can handle the load when an army of pilot projects suddenly scales into full production. For now, the infrastructure race is gaining momentum, and the winner will not be the one who implemented a neural network first, but the one who learned to scale it effectively without breaking the bank.