The Evolving Data Center Landscape

White paper by Carrie Goetz, Principal/CTO, StrategITcom

Power, AI, and the Future of Mission-Critical Infrastructure

A short year ago, the two words "data center" were relatively unknown to anyone outside of the data center industry. Today, those two words are echoed everywhere, generally followed in short order by the word “power.” These words are at the head of most news cycles, and here’s why.

The growth in the data center industry hasn’t slowed. While predictions for AI depict an exponential increase in power, that is only one factor driving the increase in power across the mission-critical sector.

In 2023, there were 5,375 data centers in the US, roughly half of the 10,655 worldwide. Of the 17,000 AI companies worldwide, about 12,000 are in the United States alone.

How AI pulls power in multiple ways

While Artificial Intelligence is often discussed as a singular topic, there are multiple types of AI. Depending on the AI tasks involved, different design and power requirements are necessary. There is also a difference in power demand for training versus processing responses. Regardless, one thing is certain: the overall data center power demand for AI is significantly higher than traditional computing.

We have seen an increase in rack density from 3-15kW per cabinet to full generative AI, with demands reaching up to 150kW per cabinet. These loads necessitate enhanced heat rejection/cooling methods. Whether implemented across the entire white space or simply for high-density zones, liquid cooling is here to stay.

Simply interacting with AI centers draws more power than traditional search engine search methods. To highlight the fluctuations and different power demands, a comparison of a simple web search is approximately 0.0004 Wh. Predictive analytics draws about .45Wh per transaction, healthcare diagnostics draws about 67.5Wh, and generative AI draws roughly 18Wh. When applying AI to energy management and sustainability, the demand increases to nearly 9Wh per search. With these increases, it’s no wonder that power is dominating headlines.

Building sustainability and efficiency in data centers with AI

On the upside, we can also use AI to help build more sustainable and efficient data centers. Machine learning (ML) and AI can help us process the most minute details that might otherwise go unnoticed.

As power companies grapple with power supply, AI can help shift loads to other facilities or consolidate workloads to shed power demand. AI can rapidly report fluctuations outside of norms for actionable intelligence.

In fact, AI may be the perfect bridge between facilities and the information technology teams. The actionable intelligence gleaned is incredibly valuable.

Helping to solve the problem of stranded power

Perhaps one area that could benefit from immediate improvement is stranded power. With states like Ohio adding the burden to consume at least 85% of delivered power, others are likely to follow suit.

Currently, the data center industry strands about 65% of delivered power. Stranded power is power that is allocated but remains unused.

Secondary and Tertiary power feeds may never energize (go live) outside of routine maintenance. Allocations per cabinet may exceed cabinet demand. Paperwork errors can cause some issues. DCIM intelligence may be under or not utilized. Servers, switches, and storage devices may draw less power than anticipated. Colocation facilities earmark power to cages throughout their lease, often missing opportunities to free up some amount of power for other areas. There are a number of reasons for stranded power.

Can AI alleviate power fluctuations in data centers?

Data centers have struggled with correctly addressing power fluctuations, instead relying on maximum demand numbers for average workloads. With intelligence, users can triage high-power draw events and help determine if they are, in fact, compute increases, hardware failures, errant applications, or some other issue.

Enabling IT teams with power information per device increases strategic decision intelligence. Energy demands, coupled with processing information, enable additional efficiency calculations. Performance per Watt, efficiency ratios, capacity planning, and others, all while providing enhanced troubleshooting intelligence.

Peak shaving, load shedding and building efficiencies

When batteries join the intelligence pool, additional operational savings are realized through peak shaving, load shedding, and orchestrations to work around demand fees and peak charges. It further enables owner/operator data centers to gain better intelligence across their tenants’ spaces while providing actionable insights and guidance to those tenants.

To build efficiencies across the entire facility, we need systems that can analyze our load fluctuations more effectively.

Data-driven decision-making for data centers

Let’s look at an example. Suppose we take server A at location 1 and server B at location 2. By examining power usage in granular detail, server teams can investigate the actions the servers are processing in real-time alongside power data.

• What processes can be moved to alternative (lower power) times?

• What server performs the best as a ratio of process to power?

• Are there processes that could be scheduled during off-peak hours or processed at alternative lower-power sites?

What's ahead for data centers and the grid

This year, we are likely to see an uptick in data centers in locations with alternative energy sources, such as small modular reactors, natural gas production sites, and geothermal solutions.

The grid is not expected to support all of the new workloads. New sites may generate their power and supplement the grid or operate in island mode, autonomously from the grid. Power generation and power storage will vary from traditional grid-based designs. Operators and tenants will have greater intelligence and power source variations to optimize their computing and sustainability efficiencies.

The divide between facilities and compute is rapidly disappearing, leaving both the IT teams and power teams more intertwined than ever.

Control cooling costs and capital spend in 2026

Entities must plan for today’s workloads while incorporating designs for AI power and cooling for future demands.

Temporary facilities may become a factor for model training. By implementing temporary data centers, loads needed to train models can be located where required and then moved to the following location, saving multisite capital expenditures. This methodology may ease talent strains as multiple sites won’t need continuous staffing for temporary needs. Likewise, we may see the resurrection of decommissioned facilities as power struggles continue.

Data centers in 2026 and beyond will require diligent power management from procurement, to storage, and throughout operations.

Author & Data Center Expert Carrie Goetz

Carrie Goetz, Principal/CTO, StrategITcom, is a philanthropist, international keynote speaker, globally published, and best-selling author of her book series, “Jumpstart Your Career in Data Centers.” The educational series teaches the internet and mission critical digital industry to inspire all ages.

Carrie's accolades include: Top 25 Women in Mission Critical, Data Center World Lifetime Achievement Award, 30 Most Influential Women in Tech, Top 10 Most Influential Tech Women, among others. She has worked in the industry for over 40 years and is currently a technical writer and fractional CTO while creating opportunities through apprenticeships, scholarships, and other learning avenues.