Reg.Oblako: generative AI has entered its 1997 internet phase, and the window for entry is narrowing

Generative AI has entered a stage where the market is transitioning from curiosity to systematic implementation, and delays are becoming costly. In a column on Habr, CIO of Reg.cloud Evgeny Martynov draws a direct analogy with the internet of 1997 and argues: 2026–2027 will determine who manages to embed AI into their product and who remains stuck at the pilot stage.

Why This is 1997

The main thesis of the material is simple: mass first contact with generative AI has already happened, but real competitive advantage will not go to those who "played around with prompts," but to those who turn experiments into repeatable processes. Martynov cites the Stanford AI Index 2026: in three years, generative AI reached 53% of the world's population. For comparison, the internet took about seven years to reach the same milestone, and personal computers took more than ten.

That is, the speed of adoption is higher than previous technological waves. But the author deliberately separates adoption speed from implementation depth. By his logic, the market is currently at the middle of the second wave: users and companies already know the tool, but have not yet embedded it into key processes.

It is precisely in this window that an opportunity appears for late players: there is still time to enter, but it is rapidly narrowing.

"Technology matures together with those who implement it."

This analogy is important because in the internet of the late 90s, the winners were not those who discussed the technology the most, but those who built processes, interfaces, and distribution channels first. The author applies the same logic to AI: when technology becomes truly everyday, late arrivals will have to pay not only in money but also in time to accumulate data, train teams, and integrate into products.

What the Numbers Say

For the Russian market, the column provides another argument: in 2025, the generative AI segment reached 58 billion rubles according to Just AI and Onside estimates, representing 4.5x growth in a year. The forecast for 2030 is 778 billion rubles at an average annual rate of 68%. But the key caveat concerns the immediate period: such growth depends on whether the 2026–2027 pilots turn into real services or remain presentations and internal demos.

At the same time, the market remains surprisingly raw. According to MTC Web Services data, only 26% of companies that already have AI budgets have a clear implementation strategy. The rest are either testing individual scenarios or have not even left the planning stage. Hence the second thesis of the article: the era of "post-hype" has already begun. There are many experiments, but winners will be determined not by the number of pilots but by the ability to connect data, build security, and achieve measurable business results.

The author separately discusses the topic of models. Russian companies, by his assessment, do not build frontier systems at the level of global leaders from scratch and therefore rationally bet on open-source bases — primarily Qwen, Llama families and their derivatives. Competition is shifting from a race of capital expenditures to fine-tuning, work with domain data, applied integrations, and product quality. For the corporate segment, this is more important than the idea of "your own large model."

How Implementation is Built

From this conclusion, the infrastructure part of the article emerges. If companies cannot send sensitive data to public services, and API stability and pricing predictability are more important than "magic out of the box," then a controlled loop in Russian jurisdiction is needed: owned or leased GPUs, internal knowledge bases, and clear provider responsibility. Against this backdrop, Reg.cloud established a separate AI direction and listed scenarios that can already be launched without a large team of researchers.

• Search across internal documents and RAG on closed knowledge base
• Corporate assistant with access to sensitive information
• First-line support assistant based on ticket history and documentation
• Contract generation and verification and other documents according to local compliance

As a basic stack, the company names GPU bare metal and virtual machines, inference through vLLM, internal chatbot on Open WebUI and Ollama, automation via n8n, collaborative environment JupyterHub, S3 storage, and an autonomous agent OpenClaw. Essentially, this is an attempt to assemble not one "magical model," but a practical platform where business can quickly go from idea to working scenario — from intelligent search across a knowledge base to support automation and document processes.

What This Means

The Reg.cloud article is simultaneously a market signal and a presentation of its own AI stack. The main idea sounds convincing: the window for comfortable exploration of generative AI is closing, and in 2026–2027 companies will be divided into those who have built working processes around models and those who are still discussing pilots. For the market, this means a shift in interest from "which model to choose" to questions of data, integration, security, and speed of implementation.