AI Is Rewriting the Data Center From the Power Line In
Updated: 6 days ago
Joaquin "Jack" Fagundo, writing from Parkland, FL, on how AI's appetite for compute is changing the way data centers are designed, powered and cooled.
I've seen plenty of "next big workloads" over my career. Virtualization, big data, the move to cloud. Each one changed the data center at the margins. AI is different, because it changes the building itself.
Start with the rack. A legacy enterprise rack drew 5 to 10 kW. AI GPU racks now need 60 to 100 kW or more. Training runs need networking fabrics with very low latency and enormous bandwidth, and storage measured in petabytes on fast SSD and NVMe systems. Change the rack that much and everything around it has to change too: the electrical distribution, the cooling, the floor, the size of the site.
The forecasts reflect it. McKinsey projects AI ready capacity growing about 33% a year through 2030, with roughly 70% of all demand driven by AI by then. Goldman Sachs expects data center power demand to rise 165% by 2030, much of it from AI workloads.
POWER FIRST, EVERYTHING ELSE SECOND
Data centers were already among the largest consumers of electricity before AI arrived. Now AI is expected to double or triple the power needs of a single facility. That's why so many operators are turning to on site generation, partnerships with utilities and long term renewable power agreements. In practice, a project's real schedule is set by when the power shows up.
In practice, a project's real schedule is set by when the power shows up.
Cooling follows close behind. Air alone can't keep up with dense GPU clusters, so liquid cooling loops, immersion tanks and hybrid designs that blend air and liquid are moving from pilot to standard practice. There's a nice irony here, too. AI is also helping run these buildings, through predictive load balancing, real time cooling adjustments and machine learning that catches failures before they happen.
Then there's the community around the site. Regulators and neighbors are watching closely, which makes carbon free energy, water conservation and honest emissions reporting part of the license to operate rather than an optional extra.
HOW THE BIGGEST PLAYERS ARE PREPARING
Google runs its fleet at a power usage effectiveness of about 1.09, among the best in the industry, builds its own TPUs for AI and keeps expanding renewable sourcing to support the training behind Gemini. Oracle has joined OpenAI and SoftBank in the Stargate initiative, which plans five new AI campuses in the United States targeting 10 GW of compute, and Oracle has committed about $40 billion to Nvidia GB200 chips. Microsoft and AWS are each spending billions on AI ready regions and investing in nuclear partnerships, battery storage and renewables to keep those regions powered.
McKinsey's estimate of where all this leads is staggering: about $6.7 trillion in global data center capital spending by 2030, with $5.2 trillion of it for AI infrastructure.
WHAT COULD SLOW IT DOWN
Growth this fast carries real risk, and I'd keep five on any planner's radar. Grid constraints and permitting delays. Water scarcity, which decides which cooling designs are even possible in a given place. Chip supply chains that can swing without warning. Regulatory pressure on carbon and emissions. And AI workloads themselves, which can spike in ways nobody forecast. These risks are already shaping design and site selection, well before any of them makes the news.
The infrastructure going up now is the backbone of whatever the AI economy turns out to be. The operators who get power density, cooling and sustainability right will define the next decade of this industry, whether they're building in Silicon Valley or here in South Florida.
I've since updated the numbers in Follow the Megawatts: What the 2026 AI Spending Numbers Are Really Saying.
Joaquin "Jack" Fagundo writes about data centers, cloud and the business of AI from Parkland, FL. He has spent 20+ years leading enterprise infrastructure and advises companies through JFagundo Consulting Services. Read more about Jack or find him on LinkedIn.
