AI’s Power Problem — and Why Boring Infrastructure Still Wins

AI’s Power Problem —
and Why Boring Infrastructure Still Wins

Justin Simonds, HPE Master Technologist

A few weeks ago, my own town got a preview of a fight that’s playing out across the country. Greenwood County Council (South Carolina) spent much of the summer wrestling with whether to allow data centers here at all — ultimately passing a one-year moratorium after residents and several of my neighbors (I missed this meeting unfortunately) attended a public hearing (organizers had to move it to the Performing Arts Center to fit everyone – LOL) to raise concerns about water use, power demand, noise, and what a facility like that would do to a rural county that has always been a little bit ‘under the radar’. There were no specific proposals yet for our county but many have been popping up in South Carolina. The county just wanted rules in place before a proposal came in. Proactive government – a new concept 😊

The debate wasn’t really about data centers per se — small towns have hosted server farms and call centers for years without controversy. It was about scale. People have started to understand, correctly, that the AI-era data center is different from what came before it, and that its appetite for power (and water, and grid capacity) is large enough to reshape a community before anyone’s even decided if they want it.

I have been reading more about this and they’re not wrong to be worried.

The information is becoming alarming. U.S. data center grid-power demand is expected to climb from roughly 62 gigawatts in 2025 to over 75 gigawatts in 2026, and past 130 gigawatts by 2030 — and the Electric Power Research Institute now estimates data centers could account for 9% to 17% of all U.S. electricity generation by 2030, a range far above what analysts assumed just two years ago. Gartner separately projects global data center electricity demand could top 1,000 terawatt-hours in 2026 alone — more than the entire country of Japan consumes. Yikes.

What’s driving this increase isn’t necessarily more data centers, it’s denser ones. A conventional server rack draws maybe 5 to 15 kilowatts. New AI-optimized racks are pulling 30 to over 110 kilowatts, and a single high-density AI rack can now consume as much power as 80 to 100 homes. Some hyperscale campuses coming online this year are designed for gigawatt-scale draw – the output of a full nuclear power plant, dedicated to one facility. It’s not hard to see why a county council would want to think that through before signing anything.

The grid isn’t keeping up, either. Analysts increasingly point to interconnection queues and transmission capacity – not chip supply or capital – as the real bottleneck, and Gartner has gone so far as to predict that power shortages could restrict as much as 40% of planned AI data center capacity by 2027. This isn’t a theoretical problem for utilities to plan around later. It’s already showing up in project timelines, in interconnection queues, and – as my Greenwood friends and neighbors worried about – potentially in local power bills and water tables.

So why am I writing about this in Nonstop Insider?

Because the industry’s answer to this problem has, so far, mostly been “throw more power at it” – bigger campuses, dedicated generation, nuclear deals, on-site fuel cells. That’s a reasonable response if you’re training frontier models. But it’s worth remembering there’s an older, quieter engineering tradition that solved a related problem decades ago: how do you get the most reliable, most available compute out of the least amount of hardware, without wasting cycles on redundant work the business doesn’t actually need? That’s been Nonstop’s design philosophy since long before “efficiency” was a boardroom talking point. Fault tolerance through smart architecture, not brute-force overprovisioning. Scale that’s linear and purposeful, not speculative.

As AI inference – not just training – starts getting embedded into transactional, mission-critical workloads (the theme I touched on in my agentic AI Connection article Trends & Wins July-August 2026), the question shifts from “how much AI can we build” to “how do we run the AI we actually need, reliably, without every inference call carrying the power footprint of a small city.” That’s an infrastructure conversation Nonstop and HPE are well positioned to be part of, whether through GreenLake’s consumption-based model or simply the discipline of designing for efficiency and availability at the same time, instead of treating them as a tradeoff.

The AI industry is currently in its “bigger is better” phase. My own county just spent a summer discovering the downside of that. History suggests the boring, efficient, reliable approach usually gets its turn to look smart again.