A data center designed today may face very different demands tomorrow as AI workloads evolve. The engineering challenge? Preparing infrastructure to keep pace.
This article explores three critical design challenges emerging from the rise of AI workloads and how engineering professionals can prepare data center infrastructure to meet growing computational demands.
AI workloads require powerful computers to process large amounts of information. These computers use more electricity and generate more heat, increasing the demands on data center power supplies, cooling systems, and facility infrastructure.
According to Gartner, a global research and advisory company, global data center electricity consumption is projected to rise from 565 TWh in 2026 to 702 TWh in 2027. Over the same period, electricity consumption from AI-optimized servers is expected to increase from 175 TWh to 258 TWh. As AI requires more computing power, data centers must adapt to keep up. Here are three ways AI workloads are changing data center design.
Graphics Processing Unit (GPU) based servers, such as those equipped with NVIDIA AI GPUs, and specialized AI accelerator systems like Google Tensor Processing Units (TPUs), which are used to train and run AI models, can deliver substantial computing capacity within compact rack configurations. Their high power and cooling requirements increase the space needed for electrical distribution equipment and maintenance access.
According to the International Energy Agency, AI server power density increased 11-fold between 2020 and 2025, with a further fourfold increase projected by 2027. This rapid change means that data center designs based on older server configurations may no longer accommodate the demands of newer AI systems.
NVIDIA’s DGX SuperPOD Data Center Design Guide highlights the importance of accounting for power and cooling requirements when planning server layouts. The key is for the server capacity to be planned alongside the infrastructure needed to support it. by aligning rack density and layout with available power, cooling capacity, and airflow requirements to help data canters use space more efficiently while allowing room for future expansion. This integrated approach can also reduce the risk of costly redesigns as computing demands grow.
AI workloads, including training large language models (LLMs) and running inference across thousands of GPUs, can consume substantial amounts of electricity. As more servers operate simultaneously, the facility’s overall power demand increases, which can potentially exceed the capacity of its existing electrical infrastructure.
The U.S. Department of Energy’s Best Practices Guide for Energy-Efficient Data Center Design recommends accounting for both current and future electrical loads when designing and selecting data center power equipment. Applying this approach involves estimating the facility’s expected peak electricity demand across servers, storage, networking, and cooling equipment, then comparing it with the available capacity of transformers, switchgear, busways, and other electrical distribution components. If projected demand exceeds available capacity, infrastructure upgrades may be necessary before additional servers can be installed.
Adding servers is not just a matter of making more electricity available. The power distribution and Uninterruptible Power Supply (UPS) systems must also accommodate the additional power requirements of the new servers. Assessing these requirements against projected demand helps determine whether a facility can expand using its existing infrastructure or needs upgrades before adding computing capacity.
AI processors generate heat as they perform intensive calculations. As rack power density increases, conventional air cooling may struggle to remove heat from concentrated areas while keeping equipment within its specified operating temperature range.
A 2024 study titled “Experimental Evaluation of Direct-to-Chip Cold Plate Liquid Cooling for High-Heat-Density Data Centers” highlights how direct-to-chip liquid cooling can remove heat at its source by circulating coolant through cold plates attached to processors. This approach captures heat at the source, reducing the burden on air-based cooling systems.
Even when a data center has sufficient electrical capacity, its cooling system may limit how many high-density servers it can support. The choice of cooling technology therefore affects how much computing capacity can be installed within the facility. Evaluating heat-removal requirements early in the design process can help avoid investing in server capacity that the cooling infrastructure cannot support.
Ultimately, designing data centers for AI requires looking beyond the computing power of individual servers.
Space, electrical capacity, and heat removal all influence how much computing infrastructure a facility can support and how easily it can adapt to future demands. At the Engineering Institute of Technology (EIT), courses in Electrical and Mechanical engineering, as well as in Industrial Automation, Instrumentation and Process Control help learners develop the technical foundations to understand these interconnected design challenges.
References
Data Center Electricity Consumption to Grow
This article was published November 5th, 2026 and the content is current as at the date of publication.
Back to news