# Health Check

## Coverage Gaps

- Need the underlying BG2/Altimeter chart for the SpaceX, Anthropic, and Google revenue/GW numbers.
- Need exact active power by quarter for CoreWeave and Nebius to convert actual or guided revenue into realized revenue/GW.
- Need more primary support for Oracle OCI AI margins if used in the article.

## Contradictions

- BG2 terrestrial cost stack says about $60B/GW all-in, split as $35B silicon and $25B land/shell/power/cooling. The same discussion suggests orbital compute could be about $30B/GW with $5B/GW of launch/non-silicon cost, which does not reconcile if silicon is still $35B/GW.
- Epoch AI estimates all-in up-front cost at $38B/GW, well below the BG2 $60B/GW terrestrial figure. The likely difference is system generation, scope, ownership model, and whether all site/power scarcity costs are included.

## Duplicate or Overlapping Concepts

- "Physical capex", "non-silicon capex", "landlord capex", and "plant capex" overlap but are not identical. Keep definitions visible.

## Weakly Supported Claims

- SpaceX/xAI $14B/GW/year, Anthropic $22B-$23B/GW/year, Google $50B/GW/year are podcast-model claims from the saved BG2 transcript unless the source chart is recovered.
- JLL/Core Insights and Alpha-Matica cost ranges came through GPT Researcher as secondary sources and should be treated as triangulation.

## Promotion / Merge Candidates

- None yet. The workspace is first-pass.

## Next Passes

- Recover the source chart or X post referenced in BG2.
- Build a small model with toggles for silicon capex, physical capex, rent, power price, utilization, PUE, contract term, debt share, and target IRR.
- Convert the final model into a calmer article section that explains the spread between CoreWeave-style base economics and SpaceX/xAI upside economics.