The normalisation ladder · rung 1

US$ per megawatt

What a megawatt of AI datacentre capacity costs to hold or rent. The denomination the power and datacentre layer quotes in — interconnection queues, leases, builds — and the rung furthest from any price The Grid collects.

Reference — stated, not assessed · notice 2026-08-20 §2

1 · US$ per MW→2 · US$ per GPU-hour→3 · US$ per million tokens→4 · US$ per task

Ⅰ · The honesty frame

an empty rung is stated, never faked

The honesty frame

The Grid collects no capacity-cost-per-megawatt series. No lease rate, no build cost, no power-contract price enters any Grid store today. Nothing on this page is a print, an assessment, or an indicator — there is no number here to cite as one.

What renders instead. The arithmetic that connects this rung to the next one ($/GPU-hour), with every input cited to its public source, and one worked conversion from the latest GAP-H100 print — labelled a worked conversion. If The Grid ever collects a $/MW series, it arrives by methodology change notice first, never by this page quietly gaining a number.

Ⅱ · Megawatts to GPUs — the arithmetic

cited public inputs · the factor the arrow hides

GPUs per MW = 1,000 kW ÷ (kW per GPU × facility overhead). Every term below is stated; none is assessed.

Input	Value	Source of the figure
GPU board power (TDP)	700 W — H100 SXM	Vendor specification (NVIDIA H100 SXM published TDP). The board alone — no server, no cooling.
Server all-in power	~1.28 kW per GPU	Vendor datasheet (NVIDIA DGX H100: ~10.2 kW system maximum ÷ 8 GPUs) — adds CPUs, memory, interconnect switches, fans and supply losses.
Facility overhead (PUE)	1.1 – 1.5	Industry-published range: ~1.1–1.2 for hyperscale design figures; ~1.5 for fleet survey averages (Uptime Institute). Climate-dependent — this is why regional factors genuinely differ.

Reading	GPUs per MW	Why it is stated
Naive (TDP only)	1,429	1,000 kW ÷ 0.70 kW. The number a brochure divides to — wrong for capacity planning, shown because it anchors most public claims.
Facility, efficient	~713	1,000 kW ÷ (1.28 kW × PUE 1.1). Hyperscale-class design.
Facility, conservative	~523	1,000 kW ÷ (1.28 kW × PUE 1.5). Survey-average overhead.

The factor is contested, and that is the point — A megawatt houses somewhere between ~523 and ~713 H100-class GPUs depending on server configuration and facility efficiency — a spread wide enough to swallow most $/MW comparisons made without stating these terms. The ladder states the terms; it does not publish the factor as a series.

Ⅲ · A worked conversion — not a print

derived from the latest GAP-H100 print · labelled, dated

Input: GAP-H100 (spot) print of 2026-09-07 — 3.2000 US$/GPU-hr — the indicator-class series specified at /standards/instruments/gap-h100.

Input print provenance (travels with the derived figure): Observations: 8 bookable asks across 3 provider(s) (hyperstack 13d unchanged, lambda 13d unchanged, vast.ai 1d unchanged) · excluded from panel: runpod BENCHED 112d unchanged · transactions: 0 · judgment: none (arithmetic) · tier-2 bookable asks only

Facility reading	On-demand revenue-equivalent of 1 MW
~713 GPUs (PUE 1.1)	≈ US$54,758 per day at the printed ask, at 100% utilisation
~523 GPUs (PUE 1.5)	≈ US$40,166 per day at the printed ask, at 100% utilisation

Worked conversion — Arithmetic on the stated inputs above and one dated Grid print — nothing more. Real utilisation is below 100%, real contracts are not the on-demand ask, and none of the inputs is assessed for this purpose. The conversion shows the shape of the bridge between rung 1 and rung 2; it is not a $/MW value and is never stored as one.

Rung 1 of the normalisation ladder (notice 2026-08-20-normalisation-ladder, 2026-08-20). The Grid collects no $/MW series; a future one arrives by notice before it affects any surface. Vendor figures cited above are the vendors’ published specifications, stated as inputs to arithmetic, not endorsed as market values.