What a $105 billion AI data center actually needs, why PUE is the number to watch, and how an ops engineer reads a megawatt the way a pilot reads airspeed. No hype — just the machine and what it eats.
When I read that a data center deal landed at $105 billion, the first thing I did was stop thinking in dollars and start thinking in megawatts. That's the reflex of anyone who's ever cold-started a server room. Money is an accounting abstraction; power is physics, and physics does not negotiate.
A facility of that scale is, at its core, an energy infrastructure project wearing a computing costume. The machines are the point, sure — but every watt of compute has to be delivered, and every watt of heat has to be removed. At AI scale the numbers stop sounding like IT and start sounding like a small city's utility grid: hundreds of megawatts of sustained draw, cooling that runs around the clock regardless of the weather outside, and redundancy layered so thick that a single failure of any component is just noise.
“The industry metric I actually live by is PUE — power usage effectiveness. It's dead simple and it hides a decade of engineering: how much extra energy it takes to keep your compute alive.” — field note, DATA-CENTER OPS 002
For context on the base concept: a data center is formally a building or room where computer servers and related equipment are operated — a facility housing the server rooms and control rooms that your entire life quietly depends on. When it's done right it is boring, which is the highest compliment in this trade. A boring data center is one where the alarms are silent because the engineering already answered the questions before the machines could ask them.
Data centers chase three things, in order: (1) cheap, reliable, physically abundant power; (2) a climate that helps cooling do less work; and (3) land where you can build without fighting a city zoning board for a decade. Ohio has all three, and central Ohio specifically has become ground zero for the modern AI data center boom — enough cheap electricity, enough land, and enough political will to let a builder put steel in the ground. That's the whole story. The rest is marketing.
The real lesson for an aspiring sysadmin: the machine you operate is not the server. It's the whole plant. A ticket that says "node slow" is, more often than not, a cooling story, a power story, or a network story wearing a compute hat. Learn to read the room the way you read the log.
PUE (Power Usage Effectiveness) is the ratio of total facility power to the power your IT equipment actually consumes:
That single ratio tells you how efficient your whole plant is at the one job it exists to do. A perfect, impossible value is 1.0 — every watt in, every watt doing compute. Real facilities land in the 1.1–1.6 range. The gap is everything the plant burns that isn't compute: cooling fans and compressors, power distribution losses, lights, the water pumps, the UPS idling.
| PUE range | What it means | Where you see it |
|---|---|---|
| 1.0 | Theoretical perfection — every watt computes | Nowhere on Earth |
| 1.1–1.2 | Excellent — modern hyperscale, free-air or advanced cooling | New hyperscale builds |
| 1.3–1.5 | Good to solid — typical efficient facility | Most well-run facilities |
| 1.6–2.0+ | Poor — legacy cooling, mixed old gear | Aging enterprise closets |
Here's the thing that surprises people: at a PUE of 1.5, a facility drawing 50 MW of total power is spending roughly one third of everything it buys from the grid on cooling and overhead, not on the servers doing the work. Do that math at $105 billion of steel and electricity and you understand why cooling innovation — liquid cooling, free-air, immersion — is where the real money quietly lives.
This is the tool I wish I'd had when I was first handed a rack and told to "make it work." Slide the IT load and the PUE you think your plant actually runs at, and it tells you three things your boss will ask about: total facility draw, the exact megawatts being wasted on overhead, and the annual power bill. No mystery constants — the formula is right there under it.
FORMULA: total = IT × PUE · overhead = total − IT · energy = total × hours · bill = energy × rate.
Fill the same numbers in done right and the difference between a PUE of 1.2 and 1.6 at 50 MW is tens of millions of dollars a year — that's not trivia, that's the whole budget.
This page is grounded in the public knowledge graph and today's news — the numbers here are meant to be checked, not taken on faith. The base concept of what a data center is comes from Wikidata (identifier Q671224); the Ohio story is linked in the ticket at the top.
This page belongs to a connected body of work — my op log on graceful degradation, and my first film on Alarm 1201 as IT triage. Same discipline: read the tell before the red light decides for you.
A neighbor who thinks about resilience the way I do — treating a fault with patience and structure, not force — is worth a bench beside: