Martin Petr for Lupa.cz: How to Build an Effective AI Infrastructure

Lupa.cz has published an interview with our colleague Martin Petr, who discusses the current state of the AI hardware market, its availability, and the challenges companies face today when building their own AI infrastructure.

High demand for GPU accelerators and other components is affecting not only prices but also delivery times. For specific configurations and platforms, such as NVIDIA HGX or DGX systems, delivery times can reach up to 24 weeks in extreme cases. According to Martin, this makes it all the more important to plan infrastructure well in advance and to take into account not only the required performance but also the availability of individual components when designing the system.

AI infrastructure isn’t just
about GPUs

In this interview, Martin explains why it’s not enough to focus solely on selecting a powerful accelerator when designing AI infrastructure. Network connectivity, fast NVMe storage, power supply, and cooling also play an important role.

The limitations of existing data centers are among the areas that companies often underestimate. Modern GPU systems place significantly higher demands on power consumption and cooling, which is why it is necessary to evaluate the infrastructure as a whole.

From training models
to inference

The way companies use artificial intelligence is also changing. Instead of training their own models from scratch, they are increasingly working with pre-trained models and adapting them to their own data. As a result, investment is gradually shifting from large-scale training clusters toward systems optimized for inference.

In the interview, you’ll also learn more about the NVIDIA AI Factory concept, the differences between on-premises infrastructure and the cloud, and what to consider when planning AI infrastructure to ensure it is ready for future generations of hardware.

You can read the full interview with Martin Petr on Lupa.cz.