Datacenters & Digital Infrastructure

Maps the AI-infra supply chain (DOE supercomputers, neoclouds, power/cooling vendors) on GB300 NVL72; Solstice 100k Blackwell, xAI Colossus 2 500k+ GPUs

First-party central-bank confirmation that AI data-center construction is now a primary driver of US business fixed investment and GDP growth: business fixed investment +11% annualized Q1 2026 vs +5.5% in 2025, explicitly attributed to AI-infrastructure buildout

ACTUAL Q2 2026: Google Cloud rev $24.8B (+82% YoY), Cloud op margin 35.6%, Q2 capex $44.9B; free cash flow NEGATIVE $(5.9)B as capex outran OCF, funded by $49.6B equity raise + $20.3B notes. FY capex guide and RPO backlog absent from the exhibit (call/10-Q items).

ACTUAL Q2 2026: capex $5.79B (+142% YoY) turned FCF NEGATIVE $(1.1)B; AI Training Compute table moved AND switched units to MW (Cortex 1 >90MW, Cortex 2 >115MW, both Production; onsite Texas compute more than doubled H1); self-funded by drawing $1.2B cash, no capital raise.

ACTUAL Q2 2026: revenue $5.46B beat the $5.00-5.40B guide top, 'data center' named a growth leader; analog upcycle confirmed (Analog +26% YoY, Q3 guide steps up to $5.65-6.15B). Supply-side mirror: TXN capex falling ($4.9B->$3.3B TTM), FCF +271% YoY as its 300mm buildout crests.

First primary vendor spec in the KB for GB300 NVL72 rack power: 135 kW rack TDP, up to 155 kW peak (50V DC), 72 B300 GPUs + 36 Grace CPUs, ~90% liquid / 10% air. Fills the topic's OLDEST unfilled gap — kW/rack — absent across the prior 13 sources. As-of doc date 2026-07-15.

First neocloud PRIMARY source in the KB. Q1 2026 (qtr ended Mar-31-2026): revenue $2.08B, revenue backlog $99.4B, active power >1 GW / contracted >3.5 GW, Q1 capex $7.70B (~3.7x revenue), operating loss $(144)M, net loss $(740)M; Meta $21B commitment named. Fills the CoreWeave/neocloud-financials gap AND feeds backlog-double-counting (its backlog is contracted to the same hyperscalers whose own backlogs the KB tracks). Q2 2026 not yet released (mid-Aug).

Second hyperscaler fleet PUE/WUE on a stated scope boundary, two comparable fiscal years: global PUE 1.16 (FY24) -> 1.17 (FY25) — WORSE — while WUE improves 0.30 -> 0.27 L/kWh. Regional: Americas 1.16/0.34, APAC 1.28/0.25 (WUE up from 0.03), EMEA 1.16/0.03. Scope EXCLUDES leased capacity and any site under 12 months operational.

Data centres are about half of US electricity demand growth to 2030 (>420 TWh added); at least 150 GW of data-centre projects sit in grid queues in advanced stages and one fifth of the global build is at risk of grid delay; only ~20% of US data-centre connection requests materialise. CORRECTION: the proposal's 485 -> 950 TWh figure is NOT in this report (it is from IEA 'Key Questions on Energy and AI', Apr 2026, not ingested).

SELF-REPORTED survey of 800+ operators: industry average PUE 1.52 in 2026 (n=644; 1.54 in 2025), seven years of near-stasis, held up by legacy sites; average modal rack density passed 11 kW (9 kW in 2025; 7.8 kW excluding a few high-density sites); 24% now run some racks >=30 kW (19% in 2025); AI inference matches AI training as a driver of high-density racks (21% each).

Google fleet TTM PUE 1.09 for 2025 and still 1.09 at Q2 2026 (quarterly 1.10); flat at 1.09-1.12 for a decade; campuses 1.04-1.15 TTM, newest Mesa AZ at 1.28 quarterly. Widest stated overhead boundary (substations, offices, cafeterias). Against Microsoft 1.17 (FY25) and industry 1.52, overhead per kWh of compute is 0.09 vs 0.17 vs 0.52.

2026 hyperscaler capex ~$725B (+77% YoY): Amazon ~$200B, MSFT ~$190B, Google $175-185B, Meta $115-135B; ~50% compute / ~35% plant / ~15% power

Power is the binding constraint; 2026 capex >$600B; 770 facility pipeline; 36+ projects ($162B) stalled; Meta >10GW target

Names the $660-690B aggregate 2026 hyperscaler capex range (vs ~$380B 2025) with per-company breakdown; hyperscalers now supply-constrained not demand-constrained; power flagged as the binding deployment constraint

TSMC Q2 2026 record revenue (NT$1.27T, +36% YoY), full-year AI growth outlook raised above 40%; CoWoS advanced-packaging capacity and N3 node confirmed sold out through end-2026 with lead times into 2027 — the binding constraint on datacenter GPU supply

YTD 2026 data-center construction spending ($58.1B through May) is 4x the 2025 record pace; data centers now >20% of all US nonresidential building starts; power-infrastructure construction forecast +30.8% in 2026

Tesla's in-house 500MW AI-training data center (Cortex 2) at Giga Texas — a vertically-integrated, non-hyperscaler/non-neocloud example pairing GPU clusters directly with on-site Megapack power storage

Samsung preliminary Q2 2026 guidance: operating profit ~19x YoY on HBM/DRAM pricing tightness — memory-side confirmation of AI-datacenter demand pull, ahead of the Jul-23/Jul-30 print window

Pre-print framing ahead of Alphabet's Wed Jul-22 print: capex guidance revision + Google Cloud backlog ($462B) are the watched signals, not headline EPS; FY2026 capex guide $180-190B

Quantifies the grid-interconnect queue the power thesis was missing: 3-4 year waits in data-center load zones, >300 GW combined PJM+ERCOT queue across ~1,500 projects (as-of May 2026); shortfall is queue throughput/deliverability, not raw MW (queue holds enough on paper through 2030). Fills the June-flagged interconnect-queue-wait-time gap.

Fills the long-absent 'capex per MW' gap — the master unit of the buildout: ~$59M/MW all-in (~$30M IT + ~$29M infra), ~$15M/MW/yr lease revenue, ~$1M/MW opex, ~4yr payback. Unit economics as-of ~April 2025 (Crusoe CEO pricing), republished May 2026; single-operator estimate. Note divergent bases: shell-only ~$11.3M/MW, AI shell+fit-out ~$37.3M/MW.

Puts a capacity trajectory on the CoWoS bottleneck (genuinely new vs the existing 'sold out' source): TSMC ~120-140K wafers/mo + OSAT 50-60K = ~200K industry in 2026; supply-demand gap ~20% -> ~10% by end-2026 (narrows but does not close); >80% CAGR 2022-2027. Fills 'CoWoS realized capacity vs the 130K target' gap. Secondary (not a TSMC filing). As-of 2026-06-15.

Eleven-chapter primer on power-plant economics and US ISO markets. Carries the two implied-price anchors for AI power deals: Anthropic-TeraWulf $19B / 400MW / 20yr from H2 2027 => ~$271/MWh WITHOUT GPUs, vs SpaceX-Reflection ~$5000/MWh WITH GPUs. Also LMP as the Lagrange multiplier of the location balance constraint, the spark spread formula, and BTM as queue-avoidance bought with siting cost.

The mechanism behind the water number: a 2024 closed-loop direct-to-chip design that rejects heat to AIR and 'consumes zero water for cooling during operations' — i.e. a water-for-power substitution. WUE series 2.3 -> 0.27 L/kWh by 2025; 2030 target 40% intensity cut vs a 2022 baseline with 25% achieved; FY25 first water-positive year; alt-water share Quincy 74% / Singapore 99% / San Antonio 79%.