---
id: "prereq-ai-energy-intensity"
type: "prerequisite"
source_timestamps: ["00:46:45"]
tags: ["artificial-intelligence", "energy"]
related: ["claim-ai-energy-bottleneck", "concept-space-data-centers"]
reason: "Grasping the massive power requirements for training and running modern AI models is essential to understanding the thesis behind Aetherflux."
sources: ["robinhood"]
sourceVaultSlug: "cardone-bhatt-robinhood-aetherflux-2026Jun25"
originDay: 10
---
# Energy Intensity of LLMs

## Why it matters

The premise of building data centers in space — see [[concept-space-data-centers]] — depends on understanding that:

- Training and inference for frontier AI models consumes **substantial electrical power**.
- Aggregate data-center energy demand is projected to rise sharply (media coverage cites figures like a **165% surge in AI power use** and headlines about a multi-hundred-billion-dollar industry power bill).
- Power-hungry workloads are concentrated geographically and stress local grids and interconnection queues.

## Why grid capacity is the binding constraint (per Bhatt)

See [[claim-ai-energy-bottleneck]] for the full claim. Securing real estate, getting utility connections, and constructing new substations and generation can take **more than half a decade** — a timeline that doesn't match the pace of frontier-AI scaling.

## Why this premise is necessary but not sufficient

Terrestrial paths remain plausible:

- Hardware/algorithmic efficiency gains (ASICs, model compression, neuromorphic compute) could blunt the demand curve.
- Nuclear, renewables, and storage build-out can expand the supply curve.
- Permitting reform can compress timelines.

A downstream agent should hold *both* the demand reality and the terrestrial-response possibility space before evaluating Aetherflux on its merits.