If 2023 was the year that excitement around AI hit a fever pitch, 2024 is the year when hype collides with reality. Industry leaders are touting the game-changing possibilities for AI to transform the way companies do business, yet the actual use cases in the enterprise are few and far between, mainly due to issues with cybersecurity, data confidentiality, and the labor shortages.

Meaningful progress is poised to take shape in the year ahead, as more companies get a handle on their strategies and begin to develop tangible business plans for real AI use cases in the enterprise.

As AI makes the shift from hype to reality, data center operators must be prepared to address a multitude of factors — chief among them being the evolving power requirements for AI infrastructure.

The growing need for power management

The rise of AI will only accelerate the need for data center power. According to research from the University of Washington, training a single AI model can use up to 10 gigawatt hours, roughly equivalent to the amount of electricity that 1,000 homes consume in a year in the U.S.

This has considerable implications for data center operators, as the amount of energy needed to power a GPU (~700 W) is nearly three times as much needed to power a standard CPU (~250 W). Since GPUs are typically grouped in clusters — up to eight within an NVIDIA DGX H100 server — this massive power draw has a sizeable impact on rack density, as it can demand a whopping 10 kW per server and 80-100 kW per rack compared to the 5-10 kW per rack of traditional CPU servers.

Taken together, the power requirements for AI data centers are exponentially greater than current data centers can support. For these reasons and others, the Dell’Oro Group published a report finding that the data center physical infrastructure (DCPI) market is set to grow to more than $46 billion by 2028, driven by increasing data center power and thermal management requirements.

Necessity fuels innovation

Changing power requirements require new approaches to power management, and this will be especially true as more AI deployments take shape. Following are some trends we can expect to see later in 2024 and beyond, as data center operators look to seize on new opportunities to transform the way their data centers manage power.

Investments in resiliency amid availability challenges

As grid power availability becomes less reliable due to increasing demand, data centers will need to prepare for more frequent potential power outages and disturbances. This will mean greater reliance on UPSs that are more efficient and provide higher power density in a smaller footprint, as data center space becomes more limited. Many operators may seek to replace less sustainable gen sets with battery energy storage systems (BESSs), which is becoming possible in modern UPSs, thanks to the inclusion of lithium-ion battery technology.

The newly recognized phenomenon in larger AI data centers, called power “bursting,” is a frequent and dramatic swing in utility power draw by as much as +/- 50%, repeating every few seconds continually. This has sharpened the focus on a recent innovation known as grid-interactive UPSs, which transform the UPS into a distributed energy resource (DER) for the grid. Using sophisticated digital capabilities and analytics, the UPS can create a fast-responding bidirectional flow of energy to and from the grid, offsetting peak energy usage and providing grid stabilization capabilities while creating a potential revenue source to offset high energy costs.

Higher density, new approaches to cooling — The IT rack helps organize and secure IT equipment in everything from high-density data centers to edge environments. The larger size of servers for AI will likely necessitate racks that are significantly deeper than current racks to accommodate for higher static weight capacity.

The higher power requirements of these servers will also likely have implications for rack PDUs, which will become higher in power density (50 kW or greater) and change in form factor as operators require fewer outlets per rack but more breakers. Traditional vertical PDUs will also likely need to be converted to horizontal to accommodate for larger power cables, fiber connectivity, and other space-hungry components.

Of course, the challenge with high density power in the data center is the ability to adequately cool it. Much has been made of the emergence of liquid cooling as a necessary innovation to meet the exponential increase in heat from AI servers and GPUs. Operators should also consider alternative solutions for managing cooling closer to the source through in-row cooling or containment solutions that can evacuate heat faster and more efficiently than currently available options.

Acceleration of digital transformation — While the digital transformation of the data center began prior to the AI boom, the embrace of digital tools will only accelerate as operators seek to optimize performance, resiliency, and sustainability in response to the growing complexity of AI infrastructure. The more data operators can collect and analyze, the better equipped they are to improve performance, mitigate potential power events, analyze and report key sustainability metrics, and improve asset utilization both in the centralized data center and across edge environments.

Historically, data centers have relied on disconnected “point” solutions, such as DCIM and EPMS software, to manage and monitor disparate elements of their power management systems — from power distribution to power quality. New software tools are emerging that combine these functions under a converged platform, allowing operators to leverage tools they need for today’s challenges, such as power availability issues, while laying a foundation to add new functionality as needs evolve.

Invest for tomorrow

The data center boom has only begun and will likely go through multiple cycles of evolution before reaching maturity. As operators deploy new infrastructure to support the growing demand for AI solutions, power management must play a key role in supporting the efficiency and resiliency of that infrastructure. By leveraging advancements in power management technology across backup power, cooling, digitalization, and other key areas, operators can help make their data centers ready to power the next generation of AI.