Economics of a Megawatt of AI Data Center

Summary

Analysis piece (NextBigFuture, published 2026-05-02) that decomposes the cost and revenue of one megawatt of AI-data-center capacity. It fills the 'capex per MW — the master unit of the buildout, still absent' gap that has been open since the topic was created. The underlying unit economics are attributed to Crusoe CEO Chase Lochmiller's contract-pricing breakdown, dated ~April 2025 — so the figures are as-of ~April 2025, republished/analyzed May 2026. This is analysis-tier and rests on a single operator's pricing; treat the numbers as a labelled analyst estimate, not fact.

Key figures (unit economics as-of ~April 2025, per Crusoe; republished May 2026)

Capex — all-in ~$59M per MW of developed capacity

• IT (GPUs, CPUs, networking, storage): ~$30M/MW — i.e., roughly half the all-in is the compute inside the building, not the building.
• Infrastructure (datacenter shell, power plant/generation, other): ~$29M/MW.

Revenue — ~$15M+ per MW per year

• Pure infrastructure lease / IaaS layer: ~$15M/MW/yr
• Managed-services / cloud layer: an additional ~$15M+/MW/yr possible → combined ~$30M+/MW/yr

Opex & returns

• Annual opex (power + other): ~$1.0–1.1M/MW
• Annual net revenue from infrastructure alone: $1M+/MW
• Payback: ~4 years on the infrastructure investment.

Cross-check against other 2026 $/MW estimates (from search, NOT ingested)

The $/MW figure is contested and definition-dependent — capturing the spread so the number isn't quoted as a point:

• Aggregator benchmarks put standard DC construction at ~$11.3M/MW (2026, shell) and AI-optimized all-in at ~$37.3M/MW (shell + ~$25M/MW tech fit-out).
• This source's ~$59M/MW all-in is higher because it bundles on-site power generation and a fuller IT load.
• The definition drives the number: shell-only ($10–12M) < AI shell+fit-out ($37M) < all-in incl. on-site generation & full IT ($59M). Always state which basis when quoting.

On-thesis relevance

The capex/MW figure is what lets the whole buildout be modelled end-to-end — it converts the hyperscaler capex aggregates ($610–725B) and the neocloud capex (CoreWeave $31–35B/yr) into deliverable megawatts, and it ties to the grid-queue and rack-power sources (135 kW/rack → ~7.4 racks/MW). The ~half-is-IT split is the key insight: it means capex tracks GPU shipments (hence CoWoS/HBM bind capex directly), and the ~4yr payback / ~$15M/MW/yr lease is the direct input to frontier debate #2 (do neoclouds earn their cost of capital, or is it a depreciation mirage?).

Provenance / caveats

• Single-operator pricing (Crusoe), analyst-republished — the weakest provenance among this ingest batch. Numbers are an estimate labelled as such.
• The revenue/payback figures assume full utilization and the operator's own contract pricing; they are best-case unit economics, not a realized cohort average.
• As-of basis is ~April 2025 for the unit economics despite the May 2026 publication — GPU prices and $/MW have moved since.

Source: Economics of a Megawatt of AI Data Center, NextBigFuture, 2026-05-02 (unit economics per Crusoe CEO, ~April 2025).