Wiki/Reference
Glossary & Methodology
Definitions for the load-bearing terms in AI capex coverage. Words that get used loosely in headlines but have specific meanings in 10-Ks and FERC filings.
Capacity & Power
MW IT — megawatts of IT load: the power actually delivered to servers. The number quoted on rack-level capacity sheets.
MW critical — synonym for MW IT in most operator disclosures.
MW grid / MW gross / MW total — total facility power including cooling, lighting, losses. Always larger than MW IT by the PUE factor. A "1 GW datacenter" can mean either, and the gap is ~30–40%. Most press releases use the larger number.
PUE (Power Usage Effectiveness) — total facility power ÷ IT power. A PUE of 1.3 means 30% of power goes to non-compute (cooling, distribution). Hyperscale air-cooled: ~1.2–1.4. Modern liquid-cooled AI: claimed ~1.1; realized often higher.
WUE (Water Usage Effectiveness) — liters of water consumed per kWh of IT energy. Direct WUE counts on-site cooling water; indirect WUE includes water used to generate the electricity. Hyperscaler-disclosed direct WUE is often 0.1–0.5 L/kWh; industry-wide direct + indirect averages ~1.8 L/kWh.
Density (kW/rack) — power per server rack. Traditional enterprise: 6–10 kW. Hyperscale air-cooled: 15–30 kW. Liquid-cooled H100/H200: 60–80 kW. Blackwell GB200 NVL72: 120–132 kW. Future Rubin generation: 200+ kW.
MWh — megawatt-hour. Energy, not power. 1 MW running for 1 hour = 1 MWh. 1 GW datacenter at full utilization for a year = 8,760 GWh ≈ 8.76 TWh.
Money
Capex (capital expenditure) — cash spent acquiring/building long-lived assets. On the cash-flow statement under investing activities. Hyperscaler AI capex is dominated by GPUs (~60%) and the buildings/power/cooling that house them (~40%).
Opex (operating expenditure) — recurring costs to run the asset: power, water, staffing, maintenance, insurance, property taxes. Hits the income statement directly.
NOI (net operating income) — revenue minus opex, but before depreciation, interest, and tax. The standard real-estate metric. A REIT lives or dies by NOI growth.
Depreciation — non-cash allocation of capex over the useful life of the asset. Servers depreciated over 6 years cost ~17% of their purchase price per year on the income statement. Useful-life assumptions are the single biggest accounting lever in AI capex coverage — see economics/gpu-depreciation.md.
EBITDA — earnings before interest, tax, depreciation, amortization. NOI for non-REITs, roughly.
Yield on cost (YoC) / stabilized yield — NOI ÷ all-in development cost, at the point the asset is fully leased. The number datacenter REITs target on new builds. Currently ~11–12% at DLR; ~25%+ cash-on-cash on legacy gross book at Equinix.
Cash-on-cash return — annual cash flow ÷ initial equity invested. Higher than YoC when leverage is used.
Payback period — capex ÷ annual cash flow. Inverse of yield. A 10% YoC implies a 10-year nominal payback. A 50% gross-revenue/capex ratio implies a 2-year top-line "payback" but is not the same as economic payback because it ignores opex and depreciation.
IRR (internal rate of return) — discount rate that sets NPV of cash flows to zero. The number a private infrastructure fund actually optimizes for. Datacenter infra funds typically underwrite 10–14% unlevered IRRs.
Capex-to-FCF ratio — capital expenditure ÷ free cash flow. A ratio >1 means the business is consuming more cash on capex than its operations generate. Several hyperscalers crossed this threshold in 2025.
Lease structures
Turnkey — fully built-out datacenter space (power, cooling, M&E). Priced per kW/month. Tenant brings only servers.
Powered shell — slab, roof, power, and conduit; cooling and M&E supplied by tenant. Priced per square foot. Cheaper for the operator, gives the tenant more control over AI-specific build-out.
Colocation / colo — multi-tenant facility where customers rent rack space or cages. Equinix's core business.
Hyperscale lease — single-tenant or near-single-tenant arrangement, typically ≥1 MW per contract, 10–15-year term, with annual escalators of 3–4% or CPI.
$/kW/month — the unit price for turnkey space. Northern Virginia wholesale rates reached ~$215/kW/month in Q2 2025.
Pre-leased % — share of an under-construction facility already contracted before completion. Currently running ~75–100% at the major REITs.
Backlog — signed-but-not-yet-commenced leases. Future revenue locked in. DLR ended 2025 with ~$1.4B at 100% share / $852M at its share.
Escalator — annual rent increase clause. Pre-2022: 2–3% typical. Post-inflation: 3–4% or CPI-linked on >85% of new bookings.
Churn — share of revenue lost to non-renewals. Hyperscale > 1 MW churn is currently near zero at the public REITs.
Operators
Hyperscaler — Microsoft, Meta, Alphabet/Google, Amazon, plus increasingly Oracle and Apple. Operate their own clouds and consume large fractions of compute internally.
Colocation REIT — Equinix, Digital Realty, Iron Mountain. Public real-estate investment trusts that build and lease datacenter space. Disclose more than hyperscalers do.
Neocloud — GPU-cloud companies sitting between Nvidia and end customers, typically heavily debt-financed and customer-concentrated. CoreWeave, Lambda, Crusoe, Nebius, Together, Applied Digital.
Miner-conversion — bitcoin-mining companies (Core Scientific, IREN, Hut 8, Cipher, TeraWulf) repurposing or co-locating AI hosting capacity.
Powered land developer — entities like Lancium, PowerHouse Data Centers, that acquire land and provision power, then sell or lease to operators.
Anchor tenant — the lead customer for a new development. Their lease commitment underwrites the financing.
Power & grid
ISO / RTO (Independent System Operator / Regional Transmission Organization) — entities running wholesale electricity markets. PJM (mid-Atlantic), ERCOT (Texas), MISO (midwest), SPP (plains), CAISO (California), NYISO, ISO-NE.
Interconnection queue — list of projects waiting to connect to the grid. PJM's queue exceeded 200 GW in 2024; ERCOT's large flexible load (LFL) queue is its data-center analog.
Capacity auction — annual auction run by some ISOs (PJM, ISO-NE) to procure resource adequacy. PJM's 2025/26 auction cleared at $269.92/MW-day — ~9x the prior year. Datacenter demand is the largest single driver.
Behind the meter (BTM) — generation co-located with load, not interconnected via the grid. Avoids transmission charges and queue wait. Talen-Amazon's Susquehanna deal was the test case (FERC rejected the amended ISA in November 2024).
Front of the meter (FTM) — generation interconnected via the grid; pays transmission and capacity charges.
Curtailment / flexible load — load that can be reduced or interrupted on instruction. ERCOT SB 6 (2025) imposes new curtailment requirements on large flexible loads.
PPA (power purchase agreement) — long-term contract for power from a specific generation source. Microsoft-Constellation Three Mile Island and Google-Kairos SMR are PPAs.
Cost shift — when large industrial loads avoid paying their share of transmission/distribution costs, residual costs fall on residential and small commercial ratepayers. The central political fight in Virginia, Georgia, and Texas.
Chips & systems
GPU (Graphics Processing Unit) — historically rendering chips, now the dominant training/inference accelerator. Nvidia ~85% share.
Hopper, Blackwell, Rubin — Nvidia datacenter GPU generations. H100/H200 (Hopper, 2022–2024); B100/B200/GB200 (Blackwell, 2025); Rubin (2026+); Rubin Ultra (2027).
GB200 NVL72 — Nvidia rack-scale system with 72 Blackwell GPUs. ~$3M per rack. Requires liquid cooling and ~130 kW per rack.
HBM (High Bandwidth Memory) — stacked DRAM mounted next to the GPU die. Critical bottleneck. SK Hynix is the dominant supplier; HBM3e and HBM4 are the current generations. Demand outstripped supply through 2025.
CoWoS (Chip-on-Wafer-on-Substrate) — TSMC's advanced packaging process, used for Nvidia GPUs. CoWoS-L variant required for Blackwell. Capacity is the second bottleneck after HBM.
ASIC (Application-Specific Integrated Circuit) — custom chips. Google TPU, Amazon Trainium, Meta MTIA, Microsoft Maia. Cheaper per FLOP but only for first-party workloads.
InfiniBand / RoCE / Ultra Ethernet — high-speed networking fabrics connecting GPUs. Nvidia (Mellanox) dominates InfiniBand; Broadcom and the Ultra Ethernet Consortium are pushing Ethernet alternatives.
Demand-side terms
ARR (Annual Recurring Revenue) — annualized run-rate of subscription revenue. Often the trailing-month or peak-quarter figure × 12. Not the same as GAAP revenue, which is realized. AI labs (OpenAI, Anthropic) frequently report ARR; financial statements report a lower realized number.
Inference vs. training — training is the one-time compute spend to build a model; inference is the ongoing compute spend to run it. Inference is becoming the dominant share of cloud GPU demand.
Token economics — pricing per million tokens. Has deflated ~10–100x from GPT-4 launch (March 2023) to comparable-quality models in 2025. Compresses margins for closed-model providers.
Compute commitment — multi-year contract from a customer to a cloud provider. The MSFT-OpenAI ~$250B+ multi-year commitment is the largest disclosed; AWS-Anthropic, Oracle-OpenAI follow.
Reading the wiki: methodology
Citations. Every quantitative claim links to a primary source (10-K, transcript, ISO filing, official report) when possible, secondary trade press (Datacenter Dynamics, Bloomberg, FT, Reuters) when not, and named industry analysts (Semianalysis, Synergy, Omdia) when those add unique data.
Dates. Pages note when sources were pulled. AI capex moves fast enough that even six-month-old numbers may be stale; check the most recent earnings cycle.
Where numbers come from REITs. Datacenter REIT 10-Qs and earnings transcripts disclose more about project economics than hyperscaler 10-Ks do (because hyperscalers bundle datacenter capex into larger infrastructure line items). REIT data is therefore the best public proxy for unit economics.
Where numbers come from hyperscalers. Quarterly capex is in the cash flow statement. Useful-life policy is in Note 1 / "Summary of significant accounting policies." Property and equipment composition is in the property note (typically Note 4 or 5). Lease commitments are in the leases note.
Where numbers come from ISOs. Interconnection queue dashboards (PJM "New Services Queue," ERCOT "Large Flexible Load" registry, MISO queue). Capacity auction results are published immediately. Load forecasts are in annual planning reports (PJM "Load Forecast Report," ERCOT "Capacity, Demand & Reserves").
Where numbers come from local government. County board agendas, planning commission filings, EPA permitting records (Title V, NSR, Section 401), and state PUC dockets. Loudoun County (VA), Mesa (AZ), Memphis/Shelby County (TN), and Newton County (GA) are the highest-density sources.
See reference/primary-sources.md for direct URLs.