---
type: workspace
topic: ai-data-center-unit-economics
status: active
created: 2026-07-05
updated: 2026-07-05
source_notes:
  - clippings/26-07-03  substack.com  CoreWeave Unit Economics, Margin Profile, and Leverage Analysis.md
  - clippings/26-07-04  youtube.com  The SpaceX IPO, Fable 5, AI Capex Update & Market Check w Gavin Baker, Andrew Fox & Clark Tang.md
derived_from:
  - gpt-researcher session research-c2d443210dce
confidence: medium
---

# AI Data Center Unit Economics

## Purpose

This workspace collects datapoints for estimating AI data center unit economics in dollars per gigawatt: revenue per GW, silicon capex per GW, physical plant capex per GW, rent/colo cost per GW-year, power opex, annualized TCO, margins, financing, and deployment speed.

## In Scope

- Local CoreWeave unit-economics clipping and BG2 SpaceX/xAI transcript.
- GPT Researcher session artifacts from `research-c2d443210dce`.
- Primary or near-primary company filings, investor relations releases, earnings transcripts, and research models for CoreWeave, Core Scientific, Applied Digital, Nebius, NVIDIA, and Epoch AI.
- Secondary datapoints only when clearly labeled as secondary or triangulation.

## Out of Scope

- A full investment recommendation on CoreWeave, SpaceX, Nebius, or NVIDIA.
- General AI adoption forecasts unless they affect revenue/GW or capex/GW.
- Writing polish for the eventual public article. This workspace feeds the writing workflow but is not the article itself.

## Current Thesis / Working Questions

- What does a 1 GW AI data center need to earn annually to clear its capital cost?
- How much of the cost stack is silicon versus shell/power/cooling/networking?
- Are the newest SpaceX/xAI and lab contracts implying structurally higher revenue per watt than the CoreWeave-style neocloud base case?
- How much of the apparent upside comes from better monetization versus lower deployment cost?

## Output Goals

- Maintain a single, cited datapoint inventory.
- Preserve a calculation scratchpad that can be reused in the article.
- Flag which datapoints are primary, model-derived, podcast-derived, or secondary.

## Promotion Candidates

- Revenue per gigawatt.
- Silicon versus non-silicon capex.
- Contract-backed infrastructure financing.
- Deployment speed as financial variable.

## Known Tensions

- "GW" can mean IT load, active power, critical IT load, contracted power, or gross utility power. The note should state which basis is being used.
- SpaceX/xAI podcast numbers are not public filings. They are useful as a market-implied or investor-model view, not as verified contract economics.
- Orbital compute numbers in the BG2 transcript are internally approximate: one part says terrestrial all-in cost is about $60B/GW with $35B silicon and $25B non-silicon, while another line rounds orbital all-in to about $30B/GW despite also using a $5B/GW launch/non-silicon figure.