AI Infrastructure Is a Physical Problem Now
Updated: 56 minutes ago
PARKLAND, FL | Joaquin "Jack" Fagundo on why power, cooling and the people who run them will decide the future of AI infrastructure.
EVERY GENERATION OF computing has changed the data center. AI is changing it faster, and in more physical ways, than anything I've seen in 20+ years of running infrastructure. The future of AI infrastructure will be decided less by software than by power, heat and the people who operate both.
It starts with density. Accelerated servers pack far more power into a rack than traditional enterprise equipment, and that single change ripples through the whole building: electrical distribution, floor loading, cooling design, even the size of the facility. AI optimized sites are now planned at several times the capacity of a traditional enterprise data center.
Air cooling has practical limits at those densities. Direct to chip liquid cooling is becoming the standard for the densest AI clusters, and many operators now design halls that support both methods, so they can add liquid capacity as density rises without tearing anything out.
And power has become the unit of planning. Racks and square footage used to drive the infrastructure plan. Now it starts with megawatts, interconnection dates and how quickly the utility can deliver. I walk through the numbers behind that shift in Follow the Megawatts. The cost of that shift is now showing up in power bills well beyond the data center, as I explain in my piece on this fall's AI warning signs.
* * *
The hardware gets the attention, but operations has to evolve just as much. Facilities teams need real experience with liquid cooling, higher voltage distribution and tighter thermal margins. Monitoring has to connect facility, power and IT data so problems surface before they turn into downtime. And the same technology driving all this demand is improving operations too, from predictive maintenance to cooling optimization.
If I were advising a leadership team today, I'd start with four things. Decide where your AI workloads will run, because cloud, colocation and your own facilities each fit different stages of an AI program. Ask every provider for dates, and get power delivery timelines and liquid cooling roadmaps in writing. Design new space to support air today and liquid as density climbs. And invest in the team, because operating skills deserve the same attention as the hardware.
The future of AI infrastructure is physical. Organizations that plan around power, cooling and operating skill will deploy AI on schedule. The ones that treat it as another server refresh will wait. I look at what that means for providers in What AI Buyers Actually Want From a Data Center.
About the author
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.
