AI in Central Asia is moving beyond pilot projects: what is preventing companies from scaling its implementation

We spoke with Yuriy Kapelko, COO of Servercore, about the current stage of the AI market in the region, why it is no longer enough for businesses to simply connect a ready-made model, and what role local infrastructure plays in the transition to industrial AI.

Yuriy Kapelko, Almaty, COO of Servercore, LinkedIn

About Me

I am the COO at Servercore and have over 10 years of experience in creating and scaling infrastructure and cloud products for real business needs. In my work, I connect people, business goals, and technologies, relying on the principles of a human-centric approach.

What stage is the AI market in Central Asia at today

The AI market in Central Asia is gradually transitioning from familiarizing itself with generative models to their practical implementation in business processes. This transition is uneven: large banks, telecom companies, government structures, FinTech, and e-commerce have advanced significantly further than most traditional industries.

What is becoming the main advantage in AI

Access to powerful models is no longer a differentiator — everyone gets it roughly at the same time. The difference is now created by other factors:
— data quality, including in the national languages of the region
— a company’s ability to build a technological foundation for AI once — connecting to corporate data, access rights, quality control — and then using it for new tasks, rather than building every solution from scratch
— the ability to demonstrate economic impact at the level of a specific business operation.

Data quality is becoming one of the factors that determine the difference between AI projects. For the region, materials in national languages are especially important, as well as industry-specific texts and other sources necessary for training and evaluating modern AI models.

Most global language models are trained primarily on English and other languages that are well represented in open digital sources. Kazakh, Uzbek, and other languages of Central Asia have a smaller volume of high-quality digital texts, industry materials, and other data necessary for training and evaluating modern AI models.

This can affect not only the naturalness of dialogue. The difference is visible in the most mass tasks: speech recognition in call recordings, where two languages and industry terms are mixed, works noticeably worse than in English; extracting data from contracts and invoices hits a wall with professional terminology that is scarce in open sources. The same limitations appear in the analysis of legal, regulatory, and government texts.

Government and academic initiatives show that the development of high-quality digital data in national languages is becoming a separate direction of AI infrastructure.

In Kazakhstan, this is the development of KazLLM, as well as the project by Al-Farabi KazNU, which started in 2024. A significant part of the work in it fell not on the model itself, but on forming a corpus of texts in the Kazakh language — collecting, cleaning, and structuring materials for training. Similar programs are developing in Uzbekistan.

What is needed to scale AI in a company

However, the development of local data does not mean that every company needs to create its own AI model.

Most companies in the region do not need to train their own foundational model. It is much more important to create a technological environment that allows securely connecting different models, providing them access to corporate data, managing security, controlling quality, and monitoring costs.

As AI models evolve, they will change, but corporate data, processes, and integrations will remain. Therefore, the value of such an environment is determined by how quickly a company can scale successful AI scenarios to new business processes, rather than creating every solution from scratch.

In practice, this is evident in how subsequent projects are launched noticeably faster than the first one. A company that has once built a search system for its internal knowledge base: connected data sources, configured access rights, and set up answer quality checks, uses the same foundation for processing support requests and assisting new employees. The set of documents and the formulation of the task change, not the solution architecture.

The spread of such solutions also changes the role of market participants. Integrators, cloud providers, developers, and industry consultants begin to solve a new class of tasks and increasingly become technological partners: connecting AI to corporate data, ensuring secure integration with internal systems, establishing operational processes, and being responsible for the results of implementation.

Where AI really pays off for business

When AI projects move from pilots to industrial operation, companies increasingly evaluate them not by the technological capabilities of models, but by the expected economic impact.

If previously attention was focused on the number of pilots and model selection, today companies calculate the economics of the process itself: the cost of performing an operation before and after AI implementation, error rates, quality control costs, and the final financial effect.

The most promising processes are those that simultaneously feature:

— a large volume of repetitive operations
— a significant share of manual labor
— digital input data
— clear criteria for evaluating results
— measurable cost of errors.

Therefore, the first successful projects most often appear in document processing, customer support, analysis of standard inquiries, financial operations, logistics, and government services. It is in such processes that it is easier to prove the economic efficiency of AI and make a decision to scale.

A characteristic example is the transcription and analysis of call recordings: the volume is large, input data is already digital, the result is verifiable, and before implementation, such recordings were almost always listened to selectively or not at all.

The form in which AI is embedded in these processes is also changing. The transition from chatbots to AI agents means that the model not only generates text but also performs a sequence of actions: searches for information, works with documents, interacts with corporate systems, and passes the result to the next participant in the process.

However, predictable results are delivered not by universal agents capable of performing any tasks, but by narrow solutions embedded in specific processes: one checks document completeness, another prepares a draft response, and the third collects data for a report. With such an agent, you can clearly define authority, verify the result, and limit the consequences of an error — that is, meet exactly those conditions that make the economic effect measurable.

AI is gradually becoming a noticeable part of a company’s P&L, not just an IT department task. For business, the price of a single GPU card or token is not as important as the cost of performing a specific business operation, taking into account data preparation, integration, infrastructure, information security, support, and human involvement where necessary. At the same time, at scale, inference becomes a constant expense item — and the question of where and on what it is performed returns already at the operational stage. 

Why business needs GPUs when scaling AI

At the pilot stage, computing resources have indeed ceased to be a barrier: access to models via API covers most initial scenarios. The question returns at the stage of industrial operation, but in a different formulation.

First, the load profile changes. A pilot involves one-off experiments, while industrial implementation involves constant inference with predictable latency and costs that directly impact the cost of a business operation. Fine-tuning for industry terminology and local languages adds a regular, rather than one-off, need for computing power.

For business, the price of a single GPU card or token is not as important as the cost of performing a specific business operation, taking into account data preparation, integration, infrastructure, information security, support, and human involvement where necessary. At the same time, at large volumes, inference becomes a constant expense item.

Second, there is a requirement for where computations are performed physically. As soon as AI begins to work with customer inquiries, personnel and financial documents, and internal regulatory materials, the project is subject to data localization and protection requirements.

It is important that this requirement applies to the location of the site, not to the form of equipment ownership. Companies do not necessarily need to purchase GPU servers and maintain their own server room: the same requirements can be met by renting computing resources in a cloud located within the country.

Therefore, the availability of local cloud GPU resources is already not just a matter of cost, but also a matter of scenario permissibility: if data must remain in the country, an AI solution can only be launched where computations are also performed within the country.

For most scenarios, renting is more practical than buying. The load profile is uneven: fine-tuning creates short spikes, while inference creates a constant but moderate load. Equipment purchased for peaks sits idle most of the time, whereas a cloud model allows changing the configuration as the task grows and avoids freezing capital for several years ahead — which is especially important while application scenarios are still being refined.

This is where local cloud providers become part of business AI infrastructure. They provide access to the necessary power without capital expenditures and long procurement processes, allow changing the configuration according to actual load, and keep data within the country.

For example, Servercore provides GPU resources in the cloud at sites in Almaty and Tashkent. For companies, this is an opportunity to scale AI scenarios without the need to independently purchase and maintain GPU infrastructure.

What is needed for the development of AI in Central Asia

Access to powerful models is today obtained by everyone roughly at the same time. Therefore, the development of AI in the region is now determined not by the model itself, but by what surrounds it. I would highlight four components.

The first is data in national languages. Kazakh, Uzbek, and other languages of the region are less represented in digital sources, and this affects the most mass tasks. Therefore, forming text corpora — that is, collecting, cleaning, and structuring them — becomes a separate direction of AI infrastructure.

The second is the technological environment within companies. It allows securely connecting different models to corporate data and transferring successful solutions to new processes, rather than building each one from scratch.

The third is applied scenarios with measurable economic impact: document processing, customer support, financial operations, logistics, government services. It is in these that it is easier to prove that AI pays off and to make a decision to scale.

The fourth is computing resources within the country. With the transition to constant operation, inference becomes a regular expense item. And working with customer, personnel, and financial data falls under data localization and protection requirements.

The government strategies of Kazakhstan and Uzbekistan reflect exactly such a comprehensive approach. Along with computing infrastructure, they provide for the development of data, regulation, personnel training, and applied scenarios, i.e., the practical application of AI in specific industries and processes. Computing power is developing in parallel with regulation and data, not instead of them.

What will the next stage of AI in Central Asia be

The next stage of AI in the region will be associated with the transition from pilots to constant operation. As access to modern models becomes mass, the main value will be created by the ability of companies to quickly integrate and scale AI.

Data in national languages will play a separate role at this stage. Today, their shortage affects the most mass tasks. Speech recognition in calls where two languages and industry terms are mixed works worse. It is more difficult to extract data from contracts and invoices, as well as to analyze legal and regulatory texts. The more high-quality texts in Kazakh, Uzbek, and other languages of the region are collected, the smaller this gap with English will be. Therefore, such projects as KazLLM and the work of Al-Farabi KazNU are important not only for science but also for business.

It is important for companies not to create every solution from scratch, but to build a technological environment that can be reused for new AI scenarios, and to demonstrate economic impact at the level of a specific business operation. And when AI moves to constant operation, the question of inference cost and where computations are physically performed is added to this.