AI doesn't start with GPUs: what businesses need to implement artificial intelligence

Artificial intelligence is increasingly penetrating business processes, but alongside initial successful pilots, companies face a more complex question: what to do next? Build your own infrastructure, move to the cloud, prepare data, or first figure out what problem AI is supposed to solve?

This was discussed at the workshop "From GPU Infrastructure to Secure AI Implementation" within the framework of the AI & Digital Bridge 2026 forum in Astana by Farukh Bakiev, Head of Data Management at QazCode / Beeline Kazakhstan, Wang Huaxin, Director of Cloud Business at Huawei, Mohammed Karim, Regional Sales Director and Country Manager at Cyberhaven, Kirill Ilganaev, Client Manager at Check Point, and representatives of other technology companies. The focus of the discussion was infrastructure, corporate data, and security—issues that become critical when an artificial intelligence experiment turns into an industrial solution.

First the task, then AI

Farukh Bakiev from QazCode admitted that one of the most common requests from business today sounds extremely simple: "We want AI." But in his opinion, you should start not with choosing a model or purchasing computing power, but with understanding why the company needs artificial intelligence at all.

The problem is especially noticeable in large companies. There, AI must be integrated not into a clean technological environment, but into an existing architecture—with outdated systems, complex business processes, and varying data quality. 

Farukh Bakiev, Chief Data Officer QazCode

Farukh Bakiev, Head of Data Management QazCode / Beeline Kazakhstan
For a small startup that clearly understands its task, the optimal solution may be using ready-made infrastructure, on top of which the application is built. In large business, the approach is different: first, it is important to define business tasks and understand which technologies are truly necessary, and only then make a decision on serious investments in GPU power.

GPUs, or graphics processing units, are powerful computing chips capable of performing a huge number of operations simultaneously. This is why they are used for training and running AI models: the more complex and large-scale the system, the more such computing resources it may require.

As for data, even the most modern model cannot compensate for low-quality initial information.  

Farukh Bakiev, Head of Data Management QazCode / Beeline Kazakhstan
The quality of the result directly depends on the quality of the initial data. If they are poorly structured or contain errors, we will get the same low-quality result at the output—just in a more beautiful package.

Own Data Center or Cloud

One of the most lively debates revolved around infrastructure. Should a company invest in its own data centers and GPUs, or is it easier to use cloud power?

For small companies and startups, the answer from the QazCode representative is unambiguous. Purchasing even relatively compact infrastructure can turn out to be too expensive, especially considering how quickly LLMs and computing requirements are changing. 

Photo: Beeline Kazakhstan

Cloud services allow you to start experimenting with AI without large initial investments and scale up power as needed.

For companies considering cloud infrastructure as an alternative to their own data center, Beeline Business is developing the HyperCloud direction—a cloud platform for hosting corporate IT systems and services with the ability to scale resources to business tasks.

For large organizations, in his opinion, the choice is more complex. If localization, infrastructure control, and storage of sensitive data within the country are critical, own power may be justified.

Farukh Bakiev, Head of Data Management QazCode / Beeline Kazakhstan
At the current stage, for small companies and startups, the use of cloud infrastructure may be the most rational option.

Other participants in the discussion generally agreed that there is no universal answer. A representative of Huawei also spoke about the advantages of the cloud model for quick solution deployment, while for stable large loads, local infrastructure or local cloud power can be used.

Check Point Client Manager Kirill Ilganaev emphasized a hybrid scenario: large companies often combine their own infrastructure and the cloud. However, this approach simultaneously increases the attack surface—both circuits must be protected.

A representative of Cyberhaven added to the discussion, noting that modern protection tools allow controlling the handling of confidential data both in external services and within corporate infrastructure.

AI creates new risks, but the basics of security do not change

A separate topic was so-called Shadow AI—a situation where employees independently use public AI services and may upload work information there without company control.

A representative of Check Point recalled new classes of attacks on AI systems: prompt injection, model poisoning, data and infrastructure compromise. However, participants agreed that the emergence of AI does not cancel traditional information security rules.

Farukh Bakiev suggested starting with basic practices: separating access, using firewalls, analyzing code for vulnerabilities, collecting logs, building a SIEM contour, separating test, pre-production, and production environments, and using isolated "sandboxes" for AI agents.

Farukh Bakiev, Head of Data Management QazCode / Beeline Kazakhstan
No matter how modern the protection tools are, they do not replace basic digital hygiene and adherence to the main principles of information security.

Any system remains vulnerable as long as people interact with it. The situation with AI agents is similar: despite their autonomy, people still participate in the system's work both internally and externally. Therefore, basic information security practices remain the foundation of protection—regardless of how modern the technologies become.

As a result, the discussion returned to the same question with which it began: businesses should not start AI transformation by purchasing the maximum number of GPUs or building their own data center. First, you need to define the task, understand what data is needed for it and where it can be stored, and then select infrastructure and a protection system for a specific scenario.

Then artificial intelligence stops being an experiment for the sake of technology and becomes a working business tool.