cooling-pue-ladder: heat out, priced in PUE, megawatts, and seconds

Cooling decides three things at once for an AI site: whether the rack can exist (density), how much of the feed reaches the GPUs (PUE), and how many seconds you survive a cooling fault (ride-through). This repository makes the cooling ladder executable — five rungs from legacy air to immersion, a density gate you can put in CI, a capacity model where a PUE point converts to megawatts, and an executable explanation — checked against both public series — of why the industry-average PUE "plateau" and hyperscale ~1.1 fleets [2][3] are both true at once.

TL;DR: run pue-ladder validate and five findings fall out. (1) Density, not efficiency, forces the ladder — an NVL72-class ~120 kW rack [10] leaves only direct-to-chip and immersion standing; the rear-door rung buys 2.5× density while cold-climate PUE gets slightly worse. (2) A PUE point is unqueued megawatts — a 10 MW feed moved from the survey-average 1.56 [2] to 1.2 frees ~1.9 MW (~1,500 H100s' worth [11]) behind an interconnection you already own, against a median ~5-year queue for new grid capacity [8]. (3) The plateau is two fleets — one adoption mechanism reproduces the 2014 (~1.7) and 2024 (~1.56) survey levels and the hyperscale ~1.1 fleets at once; energy-weighted PUE runs ~0.32-0.35 below the survey number, so ask buyers' questions in energy-weighted terms. (4) Efficiency sells ride-through — the direct-to-chip rung AI needs today is the ladder's thermal-buffer minimum (~11 s at 132 kW), making cooling-loss a first-class chaos-testing fault. (5) Water is a priced axis — evaporative assist buys PUE at ~1 L/kWh-class WUE [3]; dry rejection gives the water back for a modeled +0.015-0.08 PUE (+0.02-0.03 on direct-to-chip).

Part of the DIMAGGI series on turning GPU capital into usable compute — cooling sits upstream of nominal in usable = nominal × network × scheduling × recovery × placement: at a fixed grid feed, nominal IT power is feed / PUE. Full analysis in docs/study.md; all claims trace to REFERENCES.md.

The ladder

A PUE point is unqueued megawatts

The plateau is two fleets

Quickstart

pip install cooling-pue-ladder                # or: pip install -e .
pue-ladder ladder                             # the rungs, priced
pue-ladder gate --density 120 --rung contained-air   # exit 1: BLOCKED
pue-ladder capacity --feed-mw 10 --pue-from 1.56 --pue-to 1.2
pue-ladder fleet                              # the two-fleets demo
pue-ladder validate                           # model vs public record

gate exits non-zero when the rung can't cool the density — usable as a CI gate for infrastructure-as-code. Every constant (density ceilings, PUE floors and penalties, retrofit $/kW, thermal buffers) is an exposed field on cooling.ladder.RUNGS, deliberately easy to override and rerun.

Reproduce

make test        # 40 tests: invariants + the 10-point validation registry
make figures     # regenerates figures/ (needs matplotlib)

Python 3.10+, stdlib only; matplotlib only for figures. The validation registry separates calibrated points (published fleet PUEs and the 2014 initial condition the constants were tuned to [2-6][13]) from emergent ones (the 2024 survey level, the survey/energy-weighted divergence, and the density verdicts [2][10][11]), and the suite fails if either kind drifts.

Series — turning GPU capital into usable compute

• GPU Cluster Networking (network-vs-more-gpus) · GPU Cluster Scheduling (scheduler-vs-more-gpus)
• Chaos Fidelity Standard (ai-cluster-chaos-fidelity) · Reliability Economics (reliability-economics) · Governed Autonomy (governed-autonomy)
• Compute↔Power Placement (compute-power-placement) · Edge↔Neocloud Placement (edge-continuum-placement) · AI-RAN ↔ Neocloud Resiliency (airan-neocloud-resiliency)
• Cooling & the PUE Ladder (this work) — heat out: the factor that converts a grid feed into nominal IT power

Margaret (Maggie) Nanyonga — Founder & Principal Architect, DIMAGGI AI.