---
license: cc-by-nc-4.0
language: [en]
pretty_name: Macrostack Datacentre Energy
size_categories: [n<1K]
tags: [datacenter, energy, power, cooling, nuclear, grid, infrastructure]
extra_gated_prompt: >-
  This dataset is free to use for evaluation and research. Tell us who you are
  and what you are evaluating, and access is granted.
extra_gated_fields:
  Name: text
  Company or institution: text
  Work email: text
  What are you evaluating?: text
  I will attribute Macrostack in any published use: checkbox
configs:
  - config_name: default
    data_files: [{split: train, path: data/energy.csv}]
---

# Macrostack — Datacentre Energy

**21 options for powering and cooling AI infrastructure**, compared on the
field that actually decides them: **lead time**.

Maintained at [macrostack.net/energy](https://www.macrostack.net/energy).
Verified `2026-09-06`.

## Why this exists

The binding constraint on AI in 2026 is not chip allocation. It is megawatts.

- US grid interconnection queue: **2,600+ GW**, waits **approaching 5 years**
- Projects that withdraw before energising: **nearly 80%**
- Data centre share of US electricity: **6-12% of total US electricity by 2026, up from 4% in 2024**
- A GB200 NVL72 rack draws **120–132 kW** against a 2026 average of ~27 kW

**No neutral comparison of these options exists**, because every organisation
qualified to write one sells one of the answers. Turbine vendors publish on
turbines. SMR developers publish on SMRs. This is the version written by
somebody selling none of them.

## The consequence nobody states plainly

A queue position is not a power supply. With four in five projects withdrawing,
**announced gigawatts and delivered gigawatts are different numbers** — and
press releases quote the first. If your interconnection date is 2031, your 2026
chip decision is really a 2031 chip decision.

## Licence

**CC-BY-NC-4.0** — free for research and internal evaluation, attribution
required. For commercial redistribution, contact macrostack.net.