AI Means More Data Centers, But It Can Also Make Them More Efficient

Data center providers have been slow to adopt artificial intelligence solutions to make their facilities more efficient, even as AI creates a boom in demand for their product, but experts say that is beginning to change. 

Ever since the launch of OpenAI’s ChatGPT kicked off a Big Tech AI arms race late last year, the growing impact of artificial intelligence on the data center industry has emerged as one of the sector's focal points. While much of the attention from industry leaders has focused on AI’s role in driving record demand, the wave of innovation has spawned new AI tools that are starting to be used by data center providers themselves to improve operations in ways that wouldn’t have been possible previously.

Some of the largest data center providers are exploring using AI to manage the critical power and cooling systems within their facilities, with the aim of improving efficiency, driving down power costs and improving sustainability metrics. Experts say they expect AI to eventually become an integral part of how data centers are run, even if the pace of adoption has been frustratingly slow.

“There is an enormous opportunity and a wealth of areas where AI can drive tangible value for data centers today,” said Jim Gao, CEO and co-founder of Phaidra, a Seattle firm that creates AI-powered control systems for data centers and other industrial facilities.

“It's ironic that the data center industry is benefiting so much from surging AI demand, yet as an industry we're not always using the very tools that our customers are creating," he added. 

Adoption may be slow, but systems like Phaidra’s represent one of the most significant ways that the largest colocation data center providers are starting to incorporate AI into their operations. 

Phaidra’s AI draws data from thousands of sensors within a data center, analyzing it in real time to make recommendations for how the building's mechanical and power systems should be operated to optimize energy efficiency. AI can autonomously control the data center, and it learns and improves over time by analyzing the results of its recommendations.

“The AI can see thousands of sensors in real time across the entire data center facility, and it’s planning multiple hours into the future,” Gao said. “It understands complex trade-offs and interactions between things like pumps and chillers and fan coils, and it’s paying attention 24/7 and doesn't need bathroom breaks. It’s also constantly learning from its own actions and getting better at operating the facility.”

Improving energy efficiency — specifically by reducing the amount of power used for cooling a data center — has been the primary entry point for AI adoption among data center firms, experts say. Limiting the amount of power that a data center has to devote to cooling instead of powering IT equipment is one of the industry’s key performance metrics.

For providers, the ability of operations teams to constantly adjust mechanical systems to maintain these efficiency standards amid shifting conditions like the outside air temperature is central to their business model. 

Industry insiders say AI tools like Phaidra’s can make these adjustments more effectively than even the most experienced operations team, incorporating far more real-time data, considering far more possible actions and working far faster than the manual processes and procedures that have been the industry standard. 

“If it hits 105 degrees today in Dallas, we need more chilling capacity, and the normal operating procedure says to turn on another chiller. What we are seeing through some of the exercises we’re doing using AI is that maybe you just need better flow to the mechanical plant and you don’t need to turn on another chiller,” said Bob Crespi, CyrusOne senior vice president for portfolio management.

“In the normal data center world, you have your standard operating procedures, and you're going to go through them because that's the tried-and-true method,” he added. “AI is a tool that makes the process more intelligent.” 

Evidence suggests that implementing this kind of AI tool to improve efficiency can yield dramatic results.

Prior to founding Phaidra, Gao developed a similar AI tool at Google to improve cooling efficiency at the tech giant’s self-operated data centers. The result was an immediate 12% improvement in efficiency. Energy savings grew to 30% over the course of a year as the AI learned and collected data, according to Google. 

Even though hyperscalers like Google have had these kinds of AI systems in place for years, adoption among colocation providers has been slow. Industry insiders point to a number of barriers, both technical and cultural, that have slowed the sector’s wholesale embrace of AI. 

Not all data centers are AI-ready. In order to deploy real-time AI tools, a facility has to have the infrastructure in place to collect the necessary data and transmit it, a capability that many facilities don’t have, according to Gao. And even if that capability exists at a given facility, managing data collection, data reliability and analysis represents a new skill set for data center operators. Indeed, the fact that implementing AI tools requires changing well-established procedures and practices for operations staff needs to be taken into account when considering these tools, Crespi said. 

For other industry leaders, AI carries significant security concerns, with a perceived risk that the flow of data to a third party could leave proprietary information or even industrial controls exposed. These potential vulnerabilities have made some industry leaders hesitant to fully embrace AI in their data centers without significant precautions in place, according to panelists at Bisnow’s DICE South event last month. 

“There was a reported case a couple of weeks ago where some employees took a piece of code and they put it into a machine learning model looking to optimize it, but they didn't realize by putting that code up there that it exposed some of that intellectual property and put it out in the wild,” Phill Lawson-Shanks, chief innovation officer at Aligned Data Centers, said at DICE South. “It's very important that you establish boundaries well in advance, and you don't just let someone on the operations team who's doing night shift just start spinning up machine models because this kind of thing is bound to happen.”