MTS

Wiki

AI Capex & Datacenters — Research Wiki

A working corpus on the economics, physics, and politics of the 2024–2026 AI datacenter buildout. Built from primary sources (10-Ks, FERC orders, ISO queue data, DOE/LBNL reports, local government records) rather than commentary.

The wiki has no thesis. It is structured to let a reader form one — by surfacing the numbers that decide the argument and the disclosures that get hand-waved in tweet-length takes.

The motivating dispute

David Sacks tweeted back-of-envelope math for a 1 GW AI datacenter: ~$50B capex, ~$25–30B/yr enterprise revenue, ~$1–2B/yr power, ~2-year payback. "The boom is real."

A pushback in reply: that's gross revenue, not NOI. Strip out opex, staffing, maintenance, taxes, and you get to a ~10% yield on cost — closer to a 10-year payback. QTS, Equinix, Digital Realty, and the other big developers are underwriting in this ballpark.

The wiki exists because both numbers can be true, depending on what you're counting and who's doing the spending. See economics/yield-vs-payback.md for the reconciliation walk-through.

The single most load-bearing primary datapoint we've found: Digital Realty Q4 2025 disclosed an 11.9% expected stabilized yield on $10B / 769 MW under construction — essentially flat against the 11–12% range it has guided to for years. AI demand has moved scale, capital structure, and tenant concentration; it has not moved project-level yield-on-cost.

Sections

1. Economics

2. Physical layer

3. Operators

  • operators/colocation-reits.md — Equinix, DLR, Iron Mountain, plus the privatized cohort (QTS, Aligned, CyrusOne). What they actually disclose.
  • operators/hyperscalers.md — MSFT, META, GOOG, AMZN, ORCL quarterly capex, useful-life policy, off-balance-sheet SPVs, debt issuance.
  • operators/neoclouds.md — CoreWeave, Lambda, Crusoe, Nebius, miner-conversions. The leveraged middle layer.

4. Sites

5. Demand

6. Risks

7. Reference

  • reference/primary-sources.md — where to find the underlying data: SEC EDGAR, FERC, ISOs, state PUCs, county records, the small set of industry analysts worth reading.
  • reference/glossary.md — PUE, WUE, yield-on-cost, NOI, MW-IT vs. MW-grid, neocloud, BTM, etc.

8. Documents — primary sources, annotated

These pages take the most consequential documents in the AI-capex record and read them honestly, with extended verbatim excerpts. Each one is a tour through a public file most coverage references without quoting. Where the PDF is freely accessible, we link to it and save a local mirror under data/source-pdfs/.

  • documents/ferc-er24-2172-talen-amazon.md — FERC's November 1, 2024 rejection of the Amazon-Talen behind-the-meter ISA, the dissent (Phillips) and concurrence (Christie), the June 2025 restructuring as front-of-meter PPA, and the December 2025 follow-up order directing PJM to rewrite its co-location rules.
  • documents/jlarc-virginia-2024.md — Data Centers in Virginia, JLARC Report 598, December 9, 2024. The most-cited but least-read state-level analysis of the industry. Headline numbers walked through (74,000 jobs, $928.6M FY23 exemption, $33/month residential rate impact by 2040), methodological caveats surfaced, the 2025 General Assembly veto cycle, and the AI-bubble stranded-cost risk JLARC explicitly names.
  • documents/dominion-gs5-order.md — Virginia SCC final order PUR-2025-00058 (November 25, 2025) creating the GS-5 rate class for ≥25 MW customers, with verbatim excerpts of the cost-of-service findings (GS-4 at 20.75% rate of return vs. GS-5 at 6.28%), the 85/85/60 minimum-charge structure, and the 14-year contract term.
  • documents/pjm-capacity-auctions.md — every PJM Base Residual Auction clearing price 2007/08 through 2027/28, the cap-binding 2026/27 and 2027/28 results ($329.17 and $333.44/MW-day), and the Independent Market Monitor's finding that data center growth drove ~45% of $47.2B in cleared cost across the last three auctions.
  • documents/coreweave-filings.md — CoreWeave's FY25 10-K with verbatim disclosures: 67% Microsoft concentration, the DDTL 1.0 through 4.0 debt ladder, $21.6B total indebtedness, useful-life change 5→6 years, $60.7B RPO, and the rarely-noted Item 9A admission that internal controls were not effective at December 31, 2025.
  • documents/hyperion-spv.md — the $27.3B Blue Owl / Meta Hyperion SPV (Beignet Investor LLC), priced October 16, 2025. The 6.581% A+ bond, the 4-year initial lease coupled with a 16-year residual value guarantee, and the implied yield-on-cost that ties tightly to Digital Realty's disclosed 11.9%.
  • documents/lbnl-2024-vs-2016.md — DOE/LBNL's 2024 datacenter energy report compared verbatim with its 2016 predecessor. Why the 2016 forecast was not spectacularly wrong (until GPU servers ate the bottom-up model), and what the 80% spread between the 2028 low and high scenarios in 2024 actually means.
  • documents/xai-shelby-permits.md — xAI Colossus Memphis: Shelby County Health Department air permit 01156-01PC, the NAACP/SELC administrative appeal (July 2025), the federal Clean Air Act complaint filed April 14, 2026 in N.D. Miss., and EPA's January 2026 NSPS Subpart KKKKa rulemaking that closes the "nonroad engine" loophole xAI relied on.

9. History

  • history/buildout-comparisons.md — AI capex measured against prior infrastructure cycles: 1990s telecom ($500B+ aggregate, ~$120B peak 2000), 19th-century railroads (peak 10–20% of GDP, two bankruptcy waves), early-20th-century electrification (Paul David's 20-year productivity lag), semiconductor fab buildouts, and the under-discussed AWS analog (2014 $4B → 2019 $20B → $40B+ operating income). The pattern that emerges: most infrastructure overbuilds eventually get utilized; the first wave of operators usually goes bankrupt.

How to read this wiki

  • Every quantitative claim should have a citation in-line. If you find a number without one, it's a bug — flag it.
  • Numbers are dated. AI capex is moving fast enough that even six-month-old figures may be stale. Each page notes when its sources were pulled.
  • Coverage of the debate is balanced by design. Where there's a bull and a bear case, both get sourced. Where there's a third view (e.g., "hyperscalers self-consuming compute" rather than "selling it externally"), it gets surfaced.
  • The wiki favors under-covered data corpora. REIT 10-Q backlog disclosures, FERC docket ER24-2172, JLARC Virginia datacenter report, ERCOT large-load registry, Hynix HBM allocation commentary — these get more attention than another think-piece on Nvidia.

Status

This is a snapshot built in May 2026 from sources current through Q1 2026 earnings cycles. The data corpora it references are continuously updated; the synthesis here is not. See reference/primary-sources.md for how to refresh.