# SOURCES.md — citation ledger
**Phase 1 verification complete: 2026-06-09.** Rule: no Tier-3 source may be the sole support for any on-screen number. Filings/agencies > major outlets > secondary. `value_type` must be honest (reported | estimate | projection).

Tiers: **1** = primary (filing, agency, company release, originating research firm) · **2** = major outlet · **3** = secondary.

| # | Claim (as it may appear on screen) | Source | Date | URL | Tier | Verified |
|---|---|---|---|---|---|---|
| 1 | Global data-centre electricity: ~415 TWh (2024, ~1.5% of global) → ~945 TWh by 2030 (Base Case, ~3%); ~15%/yr growth | IEA, *Energy & AI* | 2025-04 | https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai | 1 | **Y** |
| 2 | Industry average PUE 1.54 (2025) — flat six consecutive years | Uptime Institute, 15th Global Survey | 2025 | https://uptimeinstitute.com/resources/research-and-reports/uptime-institute-global-data-center-survey-results-2025 | 1 | **Y** |
| 3 | Transformer lead times: standard power xfmrs avg **128 wks**; GSU avg 144 wks; large units 80–210 wks (Q2'25 survey) | Wood Mackenzie | 2025 | https://www.woodmac.com/news/opinion/supply-shortages-and-an-inflexible-market-give-rise-to-high-power-transformer-lead-times/ | 1 | **Y** |
| 4 | HV substation lead times 3–5 yrs; grid equipment scarcity delaying US projects | Bloomberg (HSBC research) | 2026-04 | https://www.bloomberg.com/news/newsletters/2026-04-01/us-data-center-boom-relies-on-hard-to-find-electrical-equipment | 2 | **Y** |
| 5 | NVIDIA holds **>90%** of data-centre AI GPUs (analyst estimates vary 80–94%) | Multiple analyst trackers (present as ">90%", never a fake-precise 92%) | 2025 | https://www.visualcapitalist.com/charted-the-battle-for-ai-data-center-revenue-2021-2025/ | 2 | **Y** |
| 6 | Chips & servers ≈ **40–60%** of all-in AI build cost; ~$5.2T investment needed for AI DCs by 2030 | McKinsey, *The cost of compute: a $7T race* | 2025 | https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-cost-of-compute-a-7-trillion-dollar-race-to-scale-data-centers | 1 | **Y** |
| 7 | Construction labor gap: ABC — 439K needed (2025) → **349K (2026)** → 456K (2027); ">500K annually" the two years prior | Associated Builders & Contractors | 2026 | https://www.abc.org/News-Media/News-Releases/abc-construction-industry-must-attract-349000-workers-in-2026-despite-macroeconomic-headwinds | 1 | **Y** |
| 8 | Electricians: **818,700 jobs (2024)**; median **$62,350** (May 2024); **+9%** 2024–34; **~81,000 openings/yr** | US BLS, Occupational Outlook Handbook | 2024–25 | https://www.bls.gov/ooh/construction-and-extraction/electricians.htm | 1 | **Y** |
| 9 | Data-center trades wage premium **~25–30%**; some workers >$200K | CNBC | 2026-03 | https://www.cnbc.com/2026/03/18/ai-data-center-buildout-jobs-salary-skilled-traders-worker-shortage.html | 2 | **Y** |
| 10 | Nadella: "It's not a supply issue of chips… I don't have warm shells to plug them into" | Bg2 Pod, via TechCrunch | 2025-11 | https://techcrunch.com/2025/11/03/altman-and-nadella-need-more-power-for-ai-but-theyre-not-sure-how-much/ | 2 | **Y** |
| 11 | Hyperscaler capex (AMZN+MSFT+GOOGL+META): ~**$200B (2024)** → ~**$400B (2025, reported)** → ~**$650–700B (2026, guided)**. Per-co 2025: AMZN $131.8B, META $72.2B, GOOGL $91–93B | CNBC; company 8-Ks/calls (AMZN Q4'25; Alphabet Q3'25 8-K; Meta) | 2026-02/04 | https://www.cnbc.com/2026/02/06/google-microsoft-meta-amazon-ai-cash.htm + https://fortune.com/2026/04/30/big-tech-hyperscalers-will-spend-700-billion-on-ai-infrastructure-this-year-with-no-clear-end-in-sight-eye-on-ai/ | 1–2 | **Y** |
| 12 | Microsoft FY26 capex ≈ **$190B** (raised +$25B on memory prices; ~$40B/qtr) | MSFT FY26 Q3 call / CNBC | 2026-04 | https://www.cnbc.com/2026/04/29/microsoft-msft-q3-earnings-report-2026.html | 1–2 | **Y** |
| 13 | **~$80B Azure backlog** unfulfillable until new powered capacity; constraints "at least through June 2026" (Hood); total MSFT commercial RPO ~$625B | MSFT earnings coverage (Investing.com analysis; Fortune) | 2026-01 | https://fortune.com/2026/01/28/microsoft-stock-drops-azure-growth-slows-capex-spending-q2/ + https://www.investing.com/analysis/microsofts-power-grid-reality-check-that-just-wiped-out-357-billion-200674200 | 2 | **Y** |
| 14 | **30–50% of the 2026 pipeline** may not come online on time; ~16 GW announced for 2026, only ~5 GW under construction | Sightline Climate (primary) + Bloomberg | 2026 | https://www.sightlineclimate.com/research/data-center-outlook | 1 | **Y** |
| 15 | McKinsey: **+125 GW incremental AI capacity 2025–2030** (global; total 219 GW by 2030); **US: 25 GW (2024) → 80+ GW (2030)**; US DC demand 147 TWh ('23) → ~606 TWh ('30) | McKinsey | 2025 | https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/ai-power-expanding-data-center-capacity-to-meet-growing-demand | 1 | **Y** |
| 16 | **300,000+ new electricians** needed for AI-driven demand; **~20,000/yr retiring** (~200K/decade); Oracle slipped completions 2027→2028 partly on labor | Fortune | 2026-03-02 | https://fortune.com/2026/03/02/ai-data-centers-electrician-shortage-gen-z-training-careers/ | 2 | **Y** |
| 17 | Electrical systems ≈ **40–50% of conventional DC construction cost**; Fortune cites **45–70%** for electrical in DC builds; **MEP up to ~75% of GMP in AI-focused builds**. Always state the denominator. | Industry cost guides (dgtlinfra et al.) + Fortune | 2025–26 | https://dgtlinfra.com/how-much-does-it-cost-to-build-a-data-center/ + row-16 URL | 2 (multi) | **Y** |
| 18 | Brad Smith: electrician shortage = "the **single biggest challenge** for data center expansion in the US"; Microsoft estimates **~500K new electricians needed** over the decade | Microsoft (Smith commentary; Voices for Innovation) | 2025-03 | https://www.voicesforinnovation.org/executive_briefing/our-nation-needs-more-electricians/ | 1 | **Y** |
| 19 | PE has bought **nearly 800** HVAC/plumbing/electrical companies since 2022 (PitchBook); Alpine techs +20% pay yr-1 | WSJ (via AIC summary; PitchBook data) | 2024-10 | https://www.investmentcouncil.org/icymi-wsj-highlights-how-private-equity-transforms-plumbing-and-hvac-small-businesses-boosting-wages-and-growth/ | 2 | **Y** |
| 20 | Siemens USA: train **200,000** electricians & mfg experts by 2030; Google: **$10M initial** to ETA/IBEW/NECA targeting **100K workers + 30K new apprentices** | Siemens (BusinessWire, Dec 2025); Google (blog.google, May 2025) | 2025 | https://www.businesswire.com/news/home/20251202226785/en/Siemens-USA-Commits-to-Train-200000-Electricians-and-Manufacturing-Experts-by-2030 + https://blog.google/company-news/outreach-and-initiatives/google-org/electrical-workers-ai-training/ | 1 | **Y** |

### Tokenomics page (added 2026-06-09)

| # | Claim | Source | Date | URL | Tier | Verified |
|---|---|---|---|---|---|---|
| 21 | Constant-quality LLM inference cost falls ~**10×/yr**; GPT-3-quality: $60/Mtok (Nov 2021) → $0.06 (late 2024) = **1,000× in 3 yrs** | a16z, *LLMflation* (Appenzeller) | 2024-11 | https://a16z.com/llmflation-llm-inference-cost/ | 2 | **Y** |
| 22 | Per-milestone price declines range **9×–900×/yr (median 50×; 200× post-Jan-2024)**; GPT-4-equivalent: $20 → **$0.40/Mtok** (~62× since launch); BUT cost of running **frontier-level** models has **risen ~18×/yr** (more inference tokens per result) | Epoch AI, *LLM inference price trends* | 2025–26 | https://epoch.ai/data-insights/llm-inference-price-trends | 1 | **Y** |
| 23 | GPT-3.5-level inference: $20/Mtok (Nov 2022) → **$0.07 (Oct 2024) = 280×**; hardware cost −30%/yr, energy efficiency +40%/yr | Stanford HAI, AI Index 2025 | 2025 | https://hai.stanford.edu/ai-index/2025-ai-index-report | 1 | **Y** |
| 24 | Silicon Data **LLM Token Expenditure Index** (usage-weighted market price/Mtok): roughly doubled Dec 2025→May 2026, recent downtick on mix-shift toward cheaper models; Jan-2026 snapshot: GPT-4o $2.50/$10 per Mtok in/out; frontier reasoning tier to **$21/$168**; median out:in price ratio ≈ 4× | Silicon Data | 2026 | https://www.silicondata.com/blog/llm-cost-per-token | 2 | **Y** |
| 25 | Bifurcation framing: adoption increasingly governed by price/scarcity of inputs; elastic demand pushes users toward cheaper models where frontier isn't required; "asset prices periodically forced to reconcile ambition with physical constraint" | Citadel Securities, *Tokenomics* (Macro Strategy, F. Flight) | 2026-06 | (client note; chart: Silicon Data index) | 2 | **Y** |

### Desk page (added 2026-06-10)

| # | Claim | Source | Date | URL | Tier | Verified |
|---|---|---|---|---|---|---|
| 26 | GPU rental spot ("price of inference"): H100 **$1.49–2.99/hr** specialist (Vast $1.87, RunPod $1.99, Lambda $2.99) vs **$3–4/hr** hyperscaler; **−64–75% since Q4 2024**; B200 from $2.12 | Provider price pages + IntuitionLabs/Spheron trackers | 2026-06 | https://intuitionlabs.ai/articles/h100-rental-prices-cloud-comparison + https://vast.ai/pricing | 2 (multi) | **Y** |
| 27 | IPO pipeline: SpaceX/xAI priced ~$1.75T, debut Jun 12; Anthropic S-1 filed Jun 1 (Polymarket 81% by Oct 31); OpenAI confidential S-1 Jun 8, Sept/Q4 window, up to ~$1T (GS/MS) | Yahoo Finance / Built In / Polymarket / Wall Street Horizon | 2026-06 | https://finance.yahoo.com/markets/stocks/articles/three-hotly-anticipated-ai-ipos-122556863.html | 2 (multi) | **Y** |

### Money Map (added 2026-09-02)

| # | Claim | Source | Date | URL | Tier | Verified |
|---|---|---|---|---|---|---|
| 28 | **RETIRED 2026-09-03** (removed from the site at the editor’s direction; the fund is understood to have wound down) — Situational Awareness LP Q2-2026 13F: **26 positions, $20.24B** 13F value; **$8.46B semiconductor put wall fully exited**; SanDisk $5.67B + Micron $5.57B = **56%** of book; Nebius **$1.23B new**; TSMC $1.27B long; SharonAI **21.1%** Schedule 13D (Aug 27) | SEC EDGAR 13F-HR acc. 0000935836-26-000418 (period 2026-06-30, filed 2026-08-14); Schedule 13D (filed 2026-08-28); corroborated by 13f.info | 2026-08-14 | https://www.sec.gov/Archives/edgar/data/2045724/000093583626000418/ | 1 | **Y** (2 independent EDGAR verifications, 2026-09-01) |

### The Atlas — campuses & strata (added 2026-09-02)

| # | Claim | Source | Date | URL | Tier | Verified |
|---|---|---|---|---|---|---|
| 29 | Campus atlas (re-counted 2026-09-03): **176 real AI/hyperscale data-center campuses** in **29 countries**, **126.4 GW planned** vs **7 GW energized** (5.5%) — 20 campuses added and 21 rows corrected after a six-batch adversarial re-check of every row (see data/campuses.json `checks`). Per-row operator, tenant, coordinates, MW planned/operational, status and sources travel with each mark | Operator releases, utility/ISO filings and trade press (DCD, Reuters, Bloomberg), compiled per row; **154 rows are VERIFIED** — every original row was re-checked 2026-09-03 by an independent adversarial pass (existence/location + capacity/status) against at least one opened tier-1/2 source; **22 stay REPORTED** (the 20 campuses added that day are single-pass, and rows whose only evidence was tier 3 or whose capacity had to be nulled keep the lower mark) | 2026-09 | data/campuses.json (each row carries its own source URLs) | 1–2 (per row) | **Partial — 14/156 independently re-verified; the rest carry the REPORTED mark** |
| 30 | Industry strata: the buildout as **13 layers** from energy generation to model labs, with 121 named players and a bottleneck flag per layer (4 CHOKE · 5 TIGHT · 4 LOOSE) | Composed from ledger rows 1–28; each layer's flag figure cites its own row (12 of 13 ledger-linked, 1 press-only) | 2026-09 | data/strata.json | 1–2 (inherited) | **Y** (flags trace to existing verified rows) |

### Hardware & memory — re-verified 2026-09-02 (supersedes row 26)

| # | Claim | Source | Date | URL | Tier | Verified |
|---|---|---|---|---|---|---|
| 31 | GPU rental, Sept 2026: neocloud H100 index **$2.57 → $2.75 (+7.0%) May→Jul 2026** — *rising*; hyperscaler H100 **~$5.00–7.50** (AWS P5 $5.19, CoreWeave on-demand $6.16) and falling ~26–54% by region since Q1'25; hyperscaler premium compressed **3.0–4.5× → 2.0–2.5×**. H200 index +14.4%; B200 **$5.98–12.36**; GB200 NVL72 ~$10.50/GPU-hr | Silicon Data H100/H200 rental indices + provider price pages (RunPod, Lambda, CoreWeave, AWS Capacity Blocks) | 2026-07/09 | https://www.silicondata.com/blog/h100-rental-prices-by-region-q3-2026 + https://www.coreweave.com/pricing + https://aws.amazon.com/ec2/capacityblocks/pricing/ | 1 | **Y** (adversarially re-verified; **supersedes row 26** — see corrections) |
| 32 | NVIDIA Q2 FY2027 (qtr ended 2026-07-26): revenue **$96.2B** (+106% YoY), Data Center **$89.0B** (+117% YoY); Q3 guide **$108B ±2%**. Vera Rubin in full production from Aug 2026. AMD Q2'26 Data Center **$6.7B** (+107% YoY) — the merchant ratio corroborates NVIDIA >90% | NVIDIA and AMD quarterly releases | 2026-08 | https://nvidianews.nvidia.com/ | 1 | **Y** |
| 33 | Memory is the squeeze: Micron FQ3'26 (ended 2026-05-28) revenue **$41.5B** (+346% YoY), non-GAAP gross margin **84.9%**, FQ4 guide $50B ±$1B at ~86% GM; SK hynix Q2'26 revenue **KRW 79.3T** (+257% YoY), ~**76% operating margin**; HBM4 in high-volume shipment. TrendForce: conventional and server DRAM contract prices **+13–18% QoQ in 3Q26**; HBM rises to **30% of DRAM wafer input by 2027** | Micron and SK hynix company releases; TrendForce press releases | 2026-06/07 | https://investors.micron.com/ + https://news.skhynix.com/en/q2-2026-business-results/ + https://www.trendforce.com/presscenter/news/20260709-13140.html | 1 | **Y** |

### September 2026 re-verification — capex, backlog, tokens, grid, turbines (added 2026-09-03; supersedes rows 11, 13, 22, 24 where noted)

| # | Claim | Source | Date | URL | Tier | Verified |
|---|---|---|---|---|---|---|
| 34 | Big-4 2026 capex after the late-July calls: **AMZN ~$220B** (raised from ~$200B), **MSFT ~$175B** CY2026 (FY26 cash PP&E $115.9B; restated on 25-yr useful lives and operating leases; Q1 FY27 >$50B), **GOOGL $195–205B** (raised from $180–190B; Q2 $44.9B), **META $130–145B** (narrowed) → **~$720–745B combined** guided. NVIDIA CFO Kress (Aug 26): top-five hyperscaler capex "nearly $800 billion in 2026 and $1.3 trillion in 2027"; Huang: all-in cost of a data-center gigawatt ~$30B → ~$60B in five years | Amazon, Microsoft, Alphabet, Meta Q2-2026 releases/calls; NVIDIA Q2 FY27 call | 2026-07/08 | https://ir.aboutamazon.com/ + https://www.microsoft.com/en-us/investor + https://abc.xyz/investor/ + https://investor.atmeta.com/ + https://nvidianews.nvidia.com/ | 1 | **Y** (supersedes row 11) |
| 35 | Backlogs: Microsoft commercial **RPO $678B** (+84% y/y; RPO ex-OpenAI +25%); Azure **>$100B** FY26 revenue, Q4 +43%, Q1 FY27 guide 45% cc; Hood: "customer demand continues to exceed available capacity" (the $80B "unfulfillable" figure was not restated — label it Jan 2026). **AWS backlog $496B**; **Google Cloud backlog $514B**; **Oracle RPO $638B** (+363%), FY27 net capex ~$70B, +1.2 GW DC capacity in FY26 | Company Q2-2026 / FY releases and calls; CNBC | 2026-06/07 | https://www.microsoft.com/en-us/investor + https://ir.aboutamazon.com/ + https://investor.oracle.com/ | 1–2 | **Y** (supersedes row 13) |
| 36 | Token prices, September 2026: Silicon Data usage-weighted index **$0.97/Mtok** (31 Aug) from $1.62 (3 Jul) — falling; flagship list prices rising by generation: GPT-5 $1.25/$10 → GPT-5.6 Sol **$5/$30** (promo **$4/$20** through 21 Nov); Terra $2/$12; Luna $0.20/$1.20; **GPT-5.5-Pro $30/$180**; Claude Opus 5 $5/$25, Sonnet 5 $2/$10, **Fable 5.1 $10/$50** (cache reads $0.25); Gemini 3.7 Flash $0.75/$3.75 → $1.50/$7.50 from 1 Jan 2027; Grok 4.6 $2/$6; DeepSeek V4-Pro $1.32/$3.96 peak, half off-peak | Silicon Data index page; OpenAI, Anthropic, Google, xAI, DeepSeek rate cards | 2026-07/09 | https://www.silicondata.com/products/silicon-index/llm-token-expenditure-index + https://developers.openai.com/api/docs/pricing + https://platform.claude.com/docs/en/about-claude/pricing | 1 | **Y** (supersedes row 24) |
| 37 | Frontier running cost rising **3–18×/yr** (bigger models, more reasoning tokens) while price at a fixed benchmark level falls **~5–10×/yr** at the frontier — MIT FutureTech, *The Price of Progress* (arXiv 2511.23455v2, Mar 2026). Epoch AI's inference-price page (last updated Mar 2025): **~40×/yr** at GPT-4-level, **$37.50 → $0.18/Mtok** (GPT-4 → Gemini 2.0 Flash) | MIT FutureTech; Epoch AI | 2026-03 / 2025-03 | https://arxiv.org/abs/2511.23455 + https://epoch.ai/data-insights/llm-inference-price-trends | 1 | **Y** (corrects row 22: the "$0.40" and "200×" figures are not on Epoch's page) |
| 38 | PJM 2028/29 Base Residual Auction (14 Jul 2026): cleared at the FERC cap **$325/MW-day** RTO-wide, **138,318 MW** UCAP procured, **6,831 MW short** of the reliability requirement, $16.4B; uncapped price would have been $554.72. Board-directed **Reliability Backstop** (≤15-yr contracts, cap $555/MW-day, window **30 Sep–21 Oct 2026**, results Dec); Large Load Registry (≥50 MW); new large loads without own generation by 1 Jun 2027 curtailable; Cycle 1 interconnection: **201.5 GW** accepted, gas 99.8 GW | PJM auction results release; PJM Inside Lines; FERC filing | 2026-07/08 | https://www.pjm.com/-/media/DotCom/about-pjm/newsroom/2026-releases/20260714-pjm-capacity-auction-procures-138318-mw-of-generation-resources.pdf + https://insidelines.pjm.com/ | 1 | **Y** |
| 39 | Grid queues: **ERCOT large-load queue 474.7 GW**, of which **420.8 GW (90.2%) data centers**, >5× the 91,089 MW record peak (22 Jul 2026); ~205 GW in "Batch Zero" (study results due 9 Apr 2027); **paused 3 Aug 2026** by Gov. Abbott's audit order. **FERC** show-cause orders to all six RTOs (18 Jun; responses 17 Aug; comments 16 Sep). **LBNL Queued Up 2026**: 2,061 GW active, gas 253 GW (+86%), median 61 months request-to-COD; ERCOT 408 / MISO 382 / PJM 144 GW | ERCOT, FERC, LBNL; POWER, Troutman, Baker Botts | 2026-06/08 | https://www.ferc.gov/news-events/news/ferc-launches-aggressive-targeted-action-speed-large-load-integration + https://emp.lbl.gov/queues | 1–2 | **Y** |
| 40 | Gas-turbine queue: **GE Vernova 116 GW** backlog + slot reservations (53 firm + 63 reserved; ≥125 GW by YE2026; output 20 GW/yr → 30 GW by 2030; reservations for 2031 delivery) — 8-K 22 Jul 2026; **Siemens Energy 69 GW firm + 26 GW reservations**, lead times "three years or more"; **MHI 35 GW** (80 turbines), 2028–30 delivery. Transformers: no new Wood Mackenzie survey since Q2'25 (128 wk); PwC puts high-capacity units at up to 4 years. **Fervo–Google 396 MW** enhanced-geothermal PPA (1 Sep 2026, online 2028) | GE Vernova 8-K; Siemens Energy Q3; MHI; Fervo | 2026-07/09 | https://www.gevernova.com/investors + https://www.siemens-energy.com/global/en/home/investor-relations.html | 1 | **Y** |

### Physical constants for Plates 11–12 — Scale and The Machine (added 2026-09-03)

| # | Claim | Source | Date | URL | Tier | Verified |
|---|---|---|---|---|---|---|
| 41 | Physical constants (EIA FAQs): the average US residential customer bought **10,791 kWh in 2022** (~899 kWh/month — purchases, not consumption, so solar homes read low); transmission and distribution losses averaged **~5%** of electricity transmitted and distributed, **2018–2022** | EIA, Frequently Asked Questions #97 and #105 | 2022 data · fetched 2026-09-03 | https://www.eia.gov/tools/faqs/faq.php?id=97&t=3 + https://www.eia.gov/tools/faqs/faq.php?id=105&t=3 | 1 | **Y** |
| 42 | NVIDIA Blackwell: **208 billion transistors** on a custom **TSMC 4NP** process; **GB200 NVL72** = 36 Grace Blackwell Superchips = **72 Blackwell GPUs + 36 Grace CPUs** on fifth-generation NVLink; NVIDIA claims up to 30× LLM-inference performance vs the same number of H100s. Rack power (~120 kW) is **not** stated in the release — the site carries it as REPORTED. Per-generation "NVIDIA content per gigawatt" (Hopper ~$18B / Blackwell ~$25B / Rubin ~$40B) is likewise REPORTED, not verified against a primary transcript | NVIDIA newsroom, "NVIDIA Blackwell Platform Arrives to Power a New Era of Computing" | 2024-03-18 · fetched 2026-09-03 | https://nvidianews.nvidia.com/news/nvidia-blackwell-platform-arrives-to-power-a-new-era-of-computing | 1 | **Y** |
| 43 | US nuclear fleet, March 2026: **96 operating reactors** at 57 plants in 28 states, **98,441 MW** net summer capacity → **~1.03 GW per reactor** (derived); largest reactor Grand Gulf **~1,400 MW**, smallest Prairie Island **~520 MW** each; average reactor age ~44 years | EIA, *Nuclear explained: US nuclear industry* | 2026-03 · fetched 2026-09-03 | https://www.eia.gov/energyexplained/nuclear/us-nuclear-industry.php | 1 | **Y** |

### Capital markets, operators, labor and pipeline — June to September 2026 (added 2026-09-03; supersedes rows 8, 27 where noted)

| # | Claim | Source | Date | URL | Tier | Verified |
|---|---|---|---|---|---|---|
| 44 | The AI capital stack went public: **SpaceX (SPCX, xAI inside)** priced the largest IPO on record at **$135** (638.9M shares, **$85.7B net**, closed 15 Jun 2026); 2 Sep close **$140.71**, Nasdaq market cap **~$1.91T**, 52-week range $104.83–225.64; Q2 AI segment revenue $2.6B, AI capex $15.8B, compute 1.4 GW. **Lock-ups**: 911.5M released 6 Aug; the price-conditioned 455.8M accelerator was **not** triggered; dated tranches **9 Sep (319.0M), 10 Sep (59.1M affiliates), 24 Sep / 9 Oct / 24 Oct (328.4M each)**, ~1.3B two days after Q3 results, 797.6M on **8 Dec**; Musk's 6.4B shares locked to 12 Jun 2027. **Anthropic**: no public S-1 on EDGAR as of 3 Sep; last hard mark **$965B Series H** (28 May). **OpenAI**: no S-1; CFO Friar "public company in 2027" (CNBC, 19 Aug — tier 2) | SEC 424B4, 8-K, 10-Q (CIK 1181412); Nasdaq quote API; EDGAR company search; Anthropic newsroom | 2026-06/09 | https://www.sec.gov/Archives/edgar/data/1181412/000162828026042639/spaceexplorationtechnologi.htm + https://www.anthropic.com/news/series-h | 1 (OpenAI timing 2) | **Y** (supersedes row 27) |
| 45 | **~$100.5B of investment-grade AI paper in twelve weeks**: NVIDIA **$25.0B** (15 Jun; 4.25–5.625%, T+20–65), SpaceX **$25.0B** (23 Jun; 5.35–6.65%, refinancing the X/xAI bridge), Amazon **$24.75B** (7 Jul; 4.60–6.25%, T+40–125, after a C$13.75B Maple), Alphabet **$25.75B** (6 Aug; 4.50–6.50%, T+33–130, after its 2 Jun **$84.75B** equity package incl. $10B Berkshire placement) — spreads widening deal by deal. No Meta or Microsoft bonds in the window. **Oracle 10-K**: borrowings **$129.5B** (from $92.6B), FY26 free cash flow **−$23.7B**, RPO $638B, ~$40B to raise in FY27. **SEC interpretive letter (29 Jul)**: data-center securitizations are not Exchange Act asset-backed securities | SEC FWPs/8-Ks; Oracle 10-K; SEC Division of Corporation Finance | 2026-06/08 | https://www.sec.gov/Archives/edgar/data/1652044/000119312526338750/d159970dfwp.htm + https://www.sec.gov/rules-regulations/no-action-interpretive-exemptive-letters/division-corporation-finance-no-action/certain-data-center-securitizations-072926 | 1 | **Y** |
| 46 | **The neocloud credit ladder** (cost of money by offtaker quality): Lambda **$926M** TLB, **Baa2, SOFR+300** (27 Aug, IG offtaker); IREN **$3.6B at 6.0%** for the Microsoft contract vs **$2.4B at 9.0%** Blue Owl/PIMCO for non-IG customers; Applied Digital **$1.59B 7.0%** project notes (16 Jun; 1,410 MW contracted, ~$36B lease revenue); CoreWeave **9.625% USD / 8.5% EUR** notes (18 Jun) and **$2.6B DDTL at SOFR+550**, Ba2/BB+ (10 Aug); Nebius **$5.75B converts at 0.5%/4.5%** (24 Aug), Q2 revenue **$582M (+454%)**, >$40B IG-contracted revenue. **CoreWeave backlog ~$104B (+246%)** at 30 Jun, +$25B early Q3, capex guide **$35–39B**, active power 1.5 GW. Nscale $900M revolver (7 Jul); Fluidstack **$830M Series A at $7.5B** (20 Jul). Crusoe ~$3B at ~$30B is Bloomberg-only (tier 2, unconfirmed) | Lambda, IREN 8-K/FY26 release, Applied Digital 8-K, CoreWeave 8-K + Q2 release, Nebius releases, Nscale, Fluidstack | 2026-06/08 | https://www.sec.gov/Archives/edgar/data/1769628/000176962826000362/coreweave2q26earningspress.htm + https://lambda.ai/blog/lambda-closes-926-million-senior-secured-term-loan-b-facility | 1 | **Y** |
| 47 | **Off-balance-sheet gigawatts**: Meta–BlackRock **El Paso** venture (28 Jul): 1 GW online 2028, **~$14B**, BlackRock 80% / Meta 20%, **$12.5B debt**, Meta residual-value guarantees up to **~$13B** — the Blue Owl/Hyperion template repeated; **Hyperion expanded to 5 GW, >$50B** (13 Jul); Meta Q2 capex $31.1B, 2026 guide $130–145B. **Cerebras (CBRS)** round-tripped: IPO $185 (May), 2 Sep close **$184.52**, market cap **$43.25B**, cash $6.7B. **Q2 13Fs**: NVIDIA holds SpaceX **$21.0B** and CoreWeave 47.2M shares (unchanged); Alphabet's SpaceX stake **$94.2B**; Appaloosa initiated CoreWeave; Atreides/Coatue/Harvard concentrated in Cerebras | Meta newsroom + Q2 release; Cerebras 10-Q; Nasdaq; SEC 13F-HR filings | 2026-07/08 | https://about.fb.com/news/2026/07/meta-announces-new-venture-with-blackrock-to-develop-data-center-in-el-paso/ + https://www.sec.gov/Archives/edgar/data/1045810/000104581026000065/xslForm13F_X02/information_table.xml | 1 | **Y** |
| 48 | **Capacity is still the constraint**: Microsoft Q4 FY26 revenue **$90.0B (+18%)**, Azure **+43%**, Hood: demand "continues to exceed available capacity", Q1 FY27 Azure guide ~45% cc. **Equinix** record Q2 bookings **$424M (+23%)**, xScale **134 MW** leased, guidance raised, 2027–29 capex **$5–7B/yr**; **Digital Realty** record backlog **$1.9B**, renewal spreads 25.4% GAAP / 32.0% cash. REPORTED (tier 2, Bloomberg/CNBC 26 Aug): **Anthropic–Nscale ~$45B / six years for ~460 MW** of Vera Rubin at Monarch, Mason County WV — capacity Microsoft had reserved under a **1.35 GW LOI** and walked away from over the summer; Stargate **Abilene capped ~1.2 GW** after the scrapped expansion | Microsoft release; Equinix, Digital Realty Q2 releases (SEC); Bloomberg, CNBC | 2026-07/08 | https://news.microsoft.com/source/2026/07/29/ + https://www.sec.gov/Archives/edgar/data/1297996/000110465926086270/dlr-20260723xex99d2.htm | 1 (WV/Abilene 2) | **Y** (reported items flagged) |
| 49 | **Labor, re-based**: BLS Employment Projections 2025–35 (27 Aug 2026): electricians **821,000 (2025) → 896,900 (2035)**, +9%, **~72,700 openings/yr**; BLS OEWS May 2025 survey count **757,220**, median wage **$63,190** — a *different series*, never mix. ABC 2026 outlook: **349,000** net new construction workers needed (from ~439,000 for 2025). Electrician wages in data-center markets **+6.6–9.9% y/y vs +2.4%** economy-wide (Bloomberg via Insurance Journal, 8 Jul 2026 — tier 2). Brad Smith: the electrician shortage is the "No. 1 problem" (Fortune, Mar 2026). **Google.org $50M** skilled-trades commitment (11 Jun 2026); Meta "Americas Workforce Academy" **$115M** is tier-3 sourced — REPORTED. Build cost: Cushman & Wakefield **$17.6M/MW** (+21% y/y, range $8.9–23.3M); JLL/Turner & Townsend **$11.3M/MW** global 2026 (+6%) | BLS, ABC, AGC, Google.org, C&W, JLL | 2026-01/08 | https://www.bls.gov/news.release/ecopro.nr0.htm + https://www.cushmanwakefield.com/en/united-states/insights/data-center-development-cost-guide | 1–2 | **Y** (supersedes row 8) |
| 50 | **Pipeline and pushback**: Currence (ex-Sightline): **~16 GW** slated for US 2026 delivery, only **~5 GW under construction**, **26%** of 2025 capacity slipped; **SemiAnalysis rebuttal (18 Jun)**: hyperscaler self-build forecast moved ~1% in six months, the 5 GW baseline "off by multiples" — carry both. **EPRI *Powering Intelligence 2026*** (Feb): data centers **9–17% of US electricity by 2030** vs 4–5% today (+60% vs its 2024 view). IEA 415 → 945 TWh unrevised. **New York**: first statewide one-year moratorium on ≥50 MW data-center permits (Gov. Hochul EO, 14 Jul 2026; ~12 GW in NYISO queue). Cancellations **2 (2023) → 6 (2024) → 25 (2025)**; 188+ local opposition groups in 40 states; Gallup: ~70% oppose a nearby data center. Uptime 2026 survey (28 Jul): only "minor" PUE improvement; >50% report hiring difficulty. Widely-repeated Microsoft pauses (WI, OH) date to **2025**, not this window. Texas grid audit and state ratepayer bills (FL, TN, OK, NJ, CA pending) are tier-3 sourced — REPORTED | Currence; SemiAnalysis; EPRI; NY Governor's office; Construction Dive; Governing; Uptime | 2026-02/09 | https://www.currence.ai/blog/data-center-outlook + https://www.governor.ny.gov/ + https://powering-intelligence.epri.com/load-growth.html | 1–2 | **Y** (reported items flagged) |

### September 24 re-verification — prices, backlog, grid, labor, atlas (added 2026-09-24; supersedes rows 2, 29, 36 where noted)

| # | Claim (as it may appear on screen) | Source | Date | URL | Tier | Verified |
|---|---|---|---|---|---|---|
| 51 | Oracle Q1 FY2027 (quarter to 31 Aug; released 10 Sep): **RPO $664B** (+$209B y/y, ~+46%; >$30B of new AI cloud contracts in the quarter); OCI IaaS revenue **$7.4B, +121%**; total revenue $19.3B (+30%); FY27 capex guided **$90–95B gross, ≤$70B net**; Q1 capex $28.5B; **850 MW** delivered in Q1; borrowings $125.3B. Updates the Oracle figure in row 35 | Oracle 8-K Ex. 99.1 and 10-Q (Q1 FY27); earnings call | 2026-09 | https://www.sec.gov/Archives/edgar/data/0001341439/000119312526387905/orcl-ex99_1.htm | 1 | **Y** |
| 52 | Compute and token prices, 24 Sep 2026: Silicon Data neocloud **H100 index $2.63** (7 Sep), **−4.4%** vs $2.75 in July — the rise reversed; 6-month reserve $2.52, 12-month $2.30. Usage-weighted token index **$0.96/Mtok** (22 Sep). OpenAI **GPT-6 Astra $10/$50** (3 Sep, new top standard tier); **GPT-6 Sol $2/$10** and Luna $0.10/$0.50 (22 Sep, ~half the 5.6 prices); Anthropic **Claude Opus 5.5 $4/$20** (22 Sep, −20%), Sonnet 5 $2/$10, Fable 5.1 $10/$50; GPT-5.5-Pro $30/$180 unchanged. B200 range $5.98–12.36 unchanged | Silicon Data index pages and blog; OpenAI and Anthropic rate cards | 2026-09 | https://www.silicondata.com/products/silicon-index/llm-token-expenditure-index + https://developers.openai.com/api/docs/pricing + https://platform.claude.com/docs/en/about-claude/pricing | 1 | **Y** (supersedes row 36) |
| 53 | Capital markets, September: **SpaceX (SPCX) $148.36** close 23 Sep, ~**$1.95T**; the 24 Sep lock-up tranche (~328M) released, next 9 Oct and 24 Oct. **Nscale S-1** (18 Sep): Anthropic agreements worth **up to ~$44.6B** (Vera Rubin, Monarch WV), contracted TCV $103.4B, 1.37 GW. **CoreWeave**: ~**4.2 GW** contracted power (11 Aug), $3.7B of **2.875%** converts due 2033 (closed 22 Sep); backlog $104B unchanged. **NVIDIA–Hugging Face $12.93B** definitive agreement (2 Sep), close expected 1H27. Anthropic: no public S-1; press reports an IPO pushed to November at a ~$2T target (WSJ, tier 2). OpenAI: talks at >$1.2T (Bloomberg, tier 2) | SEC filings (Nscale S-1, NVIDIA 8-K); CoreWeave IR; Yahoo quote; WSJ; Bloomberg | 2026-09 | https://www.sec.gov/Archives/edgar/data/0002110365/000119312526395475/ck0002110365-20260918.htm + https://investors.coreweave.com/ | 1–2 | **Y** |
| 54 | ERCOT Batch Zero, conditional classification (3 Sep; board deck 11 Sep): **base load 204 projects, 66.4 GW**; **studied 158 projects, 127.9 GW**; excluded 373 projects, 302.2 GW. The 9 Apr 2027 study date is **no longer firm**; eligibility and community-impact reports go to the PUCT by 10 Dec; energization of new ≥75 MW data centers stays paused. **GE Vernova** (16 Sep): now selling **2032** turbine slots; output 5 GW/quarter in 2H26, 6 from 2H27, then 7; 116 GW slots-plus-backlog stands until the Q3 print; total backlog $176B | ERCOT Board, Batch Zero Update; GE Vernova at the Morgan Stanley Laguna conference | 2026-09 | https://www.ercot.com/files/docs/2026/09/11/14-Batch-Zero-Update.pdf + https://www.gevernova.com/sites/default/files/ge-vernova-morgan-stanley-laguna-conference-2026-transcript.pdf | 1 | **Y** |
| 55 | Demand and efficiency, re-based: IEA *Key Questions on Energy and AI* (Apr 2026) — data centres **~485 TWh in 2025 → ~950 TWh in 2030** (~3% of global demand) → ~1,200 TWh in 2035 (range 700–1,700). Uptime Institute 2026 Global Survey: average annual **PUE 1.52** (capacity-weighted 1.36) — a slow decline, no longer flat | IEA; Uptime Institute | 2026-04 / 2026-07 | https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary + https://intelligence.uptimeinstitute.com/resource/growing-pue-advantage-larger-data-centers | 1 | **Y** (supersedes row 2) |
| 56 | Labor, September: AGC–NCCER 2026 Workforce Survey (n=1,830) — **electricians are the hardest craft to fill** (81% of firms with openings report difficulty); **28%** of firms did data-center work in the past year; 42% say worker shortages delayed projects. ABC Construction Backlog Indicator (Aug): **8.5 months**; **1 in 6** members on data-center contracts, a record; data-center contractors carry 9.9 months of backlog | AGC–NCCER; Associated Builders and Contractors | 2026-09 | https://www.agc.org/sites/default/files/users/user21902/2026%20Workforce%20Survey%20Analysis%20(4).pdf + https://www.globenewswire.com/news-release/2026/09/15/3362305/0/en/abc-survey-1-in-6-members-perform-data-center-work-backlog-rebounds-in-august.html | 1 | **Y** |
| 57 | Pipeline and pushback, September: **>60%** of capacity planned for 2027 completion has not begun construction (JPMorgan via Bloomberg); Google–Blackstone 'Project Braid' executives put on-time odds at 50%, down from 90%. Thailand suspended **49** data-center builds and froze 117 approvals (4 Sep). AWS says two Gulf zones hit in March cannot be restored (15 Sep) | Bloomberg; CNBC | 2026-09 | https://finance.yahoo.com/technology/ai/articles/google-blackstone-venture-faces-delays-202510079.html + https://www.cnbc.com/2026/09/15/aws-cant-restore-service-to-bahrain-uae-6-months-after-iran-strikes.html | 2 | **Y** |
| 58 | Campus atlas re-counted 2026-09-24: **182 real AI/hyperscale campuses** in **30 countries**, **134.0 GW planned** vs **7.0 GW energized** (5.2%). Added 24 Sep: Western Downs QLD (1.44 GW), Project Phoenix PA (2 GW, under construction), HyperVault Hyderabad (1 GW), Fortune Kandukur (1.6 GW), Bell Saskatchewan (1.2 GW), PPC–AWS Kozani (0.3 GW). Project Tembo (Cheyenne) re-classed to announced after Google took it over from Crusoe | data/campuses.json (operator releases, filings, trade press) | {T} | data/campuses.json | 1–2 | **Partial** — {C['verified']}/{C['campuses']} rows independently checked (supersedes row 29) |

**Corrections vs the original seeds (recorded per Standing Rule 2):**
- "$380B 2025 / $650B 2026" → **$400B reported / $650–700B guided** (row 11).
- "MSFT ~$120B FY26" → **$190B** (row 12).
- "770K electricians" (used in explorer) → **818,700** (row 8); explorer corrected same-day.
- "439–499K shortage" → ABC's current trajectory **439K ('25) → 349K ('26) → 456K ('27)** (row 7).
- "45–70% electrical" → multi-denominator truth (row 17).
- "92% NVIDIA" → **">90%"** (row 5).
- Fortune's "$15M Google pledge" → Google's own **$10M initial** (row 20; primary beats press).
- **September 2026 re-verification:** capex row 11 → **row 34** ($650–700B → $720–745B guided); backlog row 13 → **row 35** ($625B RPO → $678B; $80B labelled Jan 2026); token prices row 24 → **row 36** ($21/$168 → GPT-5.5-Pro $30/$180, market index $0.97/Mtok); row 22 corrected by **row 37** (Epoch: $37.50 → $0.18, ~40×/yr; "$0.40" and "200×" withdrawn); row 28 **retired**.
- **September 2026 research pass (rows 44–50):** IPO row 27 "~$2T still queued" → **superseded**: SpaceX listed (row 44); no public S-1 for OpenAI or Anthropic as of 3 Sep; OpenAI guides to 2027. Row 8's **818,700** electricians (EP 2024 base) → BLS EP 2025 base **821,000** (row 49); the OEWS survey count (757,220) is a different series and is not a correction. Row 29 re-counted **156 → 176 campuses, 102.1 → 126.4 GW, 6.0 → 7 GW**; 21 verdicts corrected (e.g. Project Jade → Google "Project Tembo" 2.7 GW; Rainbow Energy Coal Creek already operating; Google Pryor North MW/capex nulled as Epoch-only). One Sept-2026 catalyst ("OpenAI listing window opens Sept–Q4") **withdrawn** as contradicted by the CFO's 2027 guidance.
- **GPU rental (row 26 → row 31, 2026-09-02):** hyperscaler H100 "~$3.50/hr" → **$5.00–7.50**; B200 "$2.12/hr" → **withdrawn, unsupported by any provider page or index**; "−64–75% since Q4 2024" → true of 2024–early 2026 but **no longer the direction**: the neocloud tier rose +7.0% May→Jul 2026 while hyperscalers cut. Row 26 is superseded.
- **September 24 date check (rows 51–58):** Oracle RPO $638B → **$664B** (row 51); neocloud H100 +7.0% May→Jul → **−4.4% Jul→Sep, $2.63** (row 52); token index $0.97 → **$0.96**; mainstream flagships **cut 20–50% on 22 Sep**, so 'list prices rising by generation' now holds only for the top tier (row 52); PUE 1.54 → **1.52** (Uptime 2026, row 55); IEA 415 → 945 TWh (2024 → 2030) re-based to **485 → 950 TWh (2025 → 2030)** (row 55); ERCOT Batch Zero re-classed and its April 2027 date withdrawn (row 54); atlas 176 → **182 campuses**, Project Tembo re-classed from under construction to announced (row 58).