Planning for data center construction projects in 2025 and beyond looks significantly different than a decade ago. However, the planning and building strategies that have successfully supported the industry through a period of dramatic growth can still provide a path forward.
The emergence of ChatGPT in late 2022 sparked an unprecedented race among tech companies to develop AI solutions, fundamentally reshaping data center infrastructure and energy markets. At the core of this transformation are AI workloads, which consist of two main operations: training and inference. These operations rely heavily on graphics processing units (GPUs), which have proven far more effective than traditional central processing units (CPUs) for handling the parallel computations essential to AI processing.
AI training operations require immense computational power, utilizing synchronized GPU arrays to process vast datasets. These training systems impose significant infrastructure demands, particularly in terms of power consumption, which typically ranges from 90 to 130 kW per rack. Such intensive energy use necessitates robust cooling systems to maintain optimal operating conditions. By comparison, inference operations, where trained models execute specific tasks, consume considerably less power – typically between 15 and 40 kW per rack. To put this in perspective, while a standard Google search uses about 0.28 watt-hours of energy, a ChatGPT query consumes roughly four times that amount.
The scale of data center infrastructure has evolved dramatically to meet these demands. Modern facilities now require individual buildings consuming 100 MW of power, with entire campuses approaching 1 GW of power consumption – a stark contrast to previous facilities that distributed 100 MW across multiple buildings. The increasing power density of GPUs has also necessitated a shift from traditional air-based cooling to liquid cooling solutions, which dissipate heat more efficiently directly from the GPU units.
Given this state of play, future data center development must consider several critical factors. Understanding whether a facility will primarily handle training or inference operations is crucial for proper design. Power infrastructure must accommodate extremely high initial requirements exceeding 100 MW per building, with the capability to scale up to 1 GW per campus. Higher voltage systems are becoming necessary to manage increased power demands while addressing thermal limitations in power cables.
Cooling systems must evolve to handle greater demands across buildings and data halls, while IT environments grow more complex with their mix of GPUs, CPUs, storage, and networking components. This complexity requires a hybrid approach to cooling, combining traditional air-based systems for certain components with liquid cooling for GPU hardware. Additionally, fiber requirements are increasing significantly, impacting facility space and weight considerations.
Data halls themselves are evolving, requiring greater vertical space to accommodate additional infrastructure layers above racks. These layers include busways, cable trays, fiber raceways, fire protection systems, and primary cooling systems incorporating water piping and technical water infrastructure.
Speed is a feature of the current race, and as such, the design and construction cycle will need to be further reduced, leveraging prefabrication not only for the electrical and mechanical layers but also for the building as a whole. This is key to reducing further headwinds for construction planning, activities and workforce safety.
Existing data centers face challenges adapting to new AI requirements, particularly for inference workloads. This adaptation often involves electrical system modifications and retrofitting for liquid cooling capabilities, reminiscent of the data center evolution in the early and mid-2000s. Training facilities, however, typically require new sites to handle massive power requirements and strict networking specifications.
While recent Nvidia GPU iterations have shown impressive improvements in cost and performance for both training and inference operations, overall electrical power consumption continues to rise proportionally with usage, following Jevons Paradox. This trend demands ongoing development in power and cooling technologies and design approaches.
The AI industry’s evolution parallels Moore’s Law, emphasizing tightly networked racks to minimize energy waste and optimize data processing speed. This transformation effectively turns AI data centers into large-scale GPU units themselves.
The rapid growth of AI has created a dramatic shift in energy market dynamics, moving from steady yearly increases to a sharp exponential rise. This surge has led to several adaptations in the industry, including:
The expansion of data center infrastructure faces additional challenges due to constraints in the construction industry. These include limitations in manufacturing capacity, shortages of builders and specialty subcontractors, and a lack of skilled workers capable of meeting the technical demands of modern data centers.
Despite these significant challenges, the industry maintains an optimistic outlook, recognizing AI’s transformative potential and embracing the opportunity to innovate and adapt to these new demands.
The evolution of data center infrastructure is a critical factor in AI’s broader development, requiring ongoing collaboration between technology companies, utility providers, and construction specialists to meet the growing demands of this rapidly expanding sector.
