The Definitive Guide toAI Data Centers
Ask the GuideAccount

Numbers Provenance Register

The guide's date-stamped figures — 1,434 register entries (a figure quoted in several chapters appears once per mention), each with its as-of date and source label, tracked over time as values are revised.

1,434 entries253 with revision history56 flagged contested166 dated 2024 or older
Recently revised
How this register works

The volatile figures in the guide's 166 chapters land here automatically at publish time, with their as-of date and source label; a figure is flagged contested when credible sources materially disagree. Each publish snapshots values into a time series — the sparklines and the “recently revised” list above are that history, not editorial claims. Treat every number as a range to stress-test, not a constant: run irreversible decisions across the spread.

1,434 matches · showing 400
MetricValueTrendAs ofUpdatedWhereSource
rack power across the inflection: legacy → GB200 NVL72 (~132 kW) → Rubin Ultra Kyber (~600 kW, 2027 roadmap)~10–15 kW → 120–600 kW20260.1SemiAnalysis / NVIDIA roadmap
practical air-cooling ceiling per rack — the discontinuity that forces liquid and rewrites the building~41 kW20250.1ASHRAE TC 9.9; SemiAnalysis Datacenter Anatomy
inference share of AI compute in 2026 (½ in 2025, ⅓ in 2023); 80–90% of draw at large operators~2/320260.1Deloitte TMT Predictions 2026; McKinsey
US large-load grid interconnection lead time end-to-end; up to ~10 yr in the worst queues — the binding constraint~3–7+ yr20260.1LBNL Queued Up; ERCOT / PJM filings
HV/substation power transformer lead time (standard); up to ~60 months in constrained markets — often the schedule's long pole~128 wk20250.1Wood Mackenzie Q2 2025 survey / pv magazine
global data center capex in 2026 (~21% CAGR through 2029; GPUs ~1/3 of capex)approaching ~$1T20260.1Dell'Oro Group
cumulative global data center capex by 2030 (~$5.2T AI-capable) — the scale that makes mis-coordination catastrophic~$6.7T20250.1McKinsey, 'The cost of compute'
end-to-end electrical-chain efficiency, 800VDC/SST chain vs a typical AC path (utility-to-VRM) — a system gain only co-design captures~87% vs ~82% AC20250.1SemiAnalysis, Datacenter Anatomy Pt 1 / Inside the 800VDC Revolution
inference share of AI compute in 2026 (½ in 2025, ⅓ in 2023) — a fast-moving figure read as direction, not a fixed level~2/320260.2Deloitte TMT Predictions 2026; McKinsey
global data center capex 2026, approaching — volatile market figure; analyst estimates differ by capex-scope definition~$1T20260.2Dell'Oro Group
per GB200 NVL72 rack (shipping, ~115 kW liquid + ~17 kW air) — a semi-durable hardware spec you can design against~132 kW2026-06-29: 132 → 2026-07-07: 13220250.2115/17 split: HPE QuickSpecs (OEM); ~132 kW all-in: Vertiv/451
per Rubin Ultra Kyber-class rack — marked roadmap/announced, not shipping; do not budget as a level~600 kWH2 2027 (announced)0.2NVIDIA GTC (Jensen Huang); DCD, Tom's Hardware
practical air-cooling ceiling per rack — an engineering rule-of-thumb (~40–50 kW practical range; see 5.1), a planning constant rather than a law of physics~41 kW20250.2ASHRAE TC 9.9; SemiAnalysis Datacenter Anatomy
large-load grid interconnection lead time — volatile and region-dependent; up to ~10 yr in worst queues~3–7+ yr20260.2LBNL Queued Up; ERCOT / PJM filings
GPU economic vs book life — flagged CONTESTED; run irreversible decisions across the range, not a point estimate · contested2–3 yr vs 5–6 yr20260.2CNBC / SemiAnalysis synthesis
best-in-class vs industry-average training goodput — a GOODPUT-thread target, vendor-marketed upper bound~96% vs ~90%20250.2SemiAnalysis ClusterMAX / CoreWeave
industry-weighted average PUE, flat for a 6th year; best-in-class liquid 1.05-1.151.5420250.3Uptime Institute Global Data Center Survey 2025
WUE range: industry avg ~1.8-1.9; best-in-class 0.3-0.7; closed-loop ~0~0-1.9 L/kWh2026-06-29: 0.95 → 2026-07-02: 0.9520250.3Vertiv / NREL synthesis; Microsoft FY2025 fleet ~0.27
goodput (effective training time): industry average vs best-in-class~90% / ~96%20250.3SemiAnalysis ClusterMAX 2.0 / CoreWeave
scale-up (NVLink) domain size: HGX node, NVL72 rack, announced Rubin Ultra Kyber8 - 72 - 57620260.3NVIDIA NVLink / Rubin platform roadmap
scale-up (NVLink5/GPU) vs scale-out (per-NIC) bandwidth — roughly 18x apart per direction (900 vs 50 GB/s each way)~1.8 TB/s vs ~400 Gb/s20250.3NVIDIA / SemiAnalysis
self-operated TCO at 2048-GPU scale, 90% util; ~$1.03-3.50 rented (contested — single-source) · contested~$0.74/GPU-hr20250.3SemiAnalysis, 'How much do GPU clusters really cost'
inference cost per million tokens: self-hosted 70B worked example vs market average~$1.90-2.502026-06-29: 2.2 → 2026-07-10: 2.220250.3SemiAnalysis synthesis
capital premium for Tier IV (fault-tolerant) over Tier III (concurrently maintainable)~20–40%20250.3Uptime Institute Tier classification; industry cost syntheses
Uptime: concurrent maintainability vs fault tolerance; legacy ~99.982% (~1.6 h/yr) vs ~99.995% (~26 min/yr), now Uptime-disavowedTier III / IV20250.4Uptime Institute Tier Standard
TIA-942-C resilience scale; full-facility telecom + M&E standard, May 2024 (C) revisionRated 1–420240.4ANSI/TIA-942-C
EN 50600 / ISO/IEC 22237 Availability Classes (+ Protection Classes); basis of the EU DC sustainability schemeClass 1–420240.4CEN / ISO/IEC JTC 1
ASHRAE TC 9.9 air classes and liquid W-classes (5th ed. + 2024 liquid-cooling resiliency addendum)A1–A4 / W17–W4520240.4ASHRAE TC 9.9 Thermal Guidelines
OCP Diablo 400 (Mt. Diablo) sidecar-power spec; ±400/800 VDC, ~100 kW to ~1 MW racksv0.7.0March 1, 20260.4OCP (Google/Meta/Microsoft)
FedRAMP 20x Key Security Indicators replacing 325+ NIST 800-53 controls; Phase 3 opens to all Q3 202656–61 KSIs20260.4FedRAMP PMO (RFC-0006)
ISO/IEC 42001 (first AI management-system standard) from publication to operationalized certification bodies2023 → 202620260.4ISO/IEC; ANAB/BSI accreditation
industry-weighted PUE (flat YoY) — the ISO/IEC 30134-2 KPI that lands in leases and disclosures1.5420250.4Uptime Institute Global Data Center Survey 2025
Tier IV capital premium over Tier III for the fault-tolerance step (the two-independent-path guarantee)20-40%20250.5Uptime Institute; Turner & Townsend Data Centre Cost Index
of impactful data-center outages root-caused to power (most often UPS); IT/networking ~23%45%20250.5Uptime Institute Annual Outage Analysis 2025
of recent major outages cost over $100k / over $1M respectively~57% / ~20%20250.5Uptime Institute Global Survey
of human-error outages caused by staff not following procedures (up 10 pts YoY) — process, not topology58%20250.5Uptime Institute Annual Outage Analysis 2025
best-in-class H100 cluster failure rate; one failure restarts a synchronous job from checkpoint~1 failure / 512 GPUs / week20250.5SemiAnalysis (100k H100 clusters)
training goodput: industry average vs best-in-class; reliability tax ~6–21% of paid GPU-hours, by provider~90% / ~96%20250.5SemiAnalysis ClusterMAX 2.0 / CoreWeave
data-center load lost in the July 10 2024 N. Virginia 230 kV disturbance (six faults over 82 s; 60 sites, 25 substations) — later a driver of NERC's rare May 2026 Level 3 alert~1,500 MW20240.5NERC Incident Review (large load loss); NERC Level 3 Alert (May 2026)
per GB200 NVL72 rack (≈132 kW typical: ~115 kW liquid + ~17 kW air)~120–132 kW2026-06-29: 126 → 2026-07-02: 12620251.1115/17 split: HPE QuickSpecs (OEM); 120 kW cooling: NVIDIA; ~132 kW all-in: Vertiv/451
per Rubin Ultra Kyber rack (NVL144) on 800 VDC~600 kWH2 2027 (announced)1.1NVIDIA GTC (Jensen Huang); DCD, Tom's Hardware
practical air-cooling ceiling per rack; RDHx ~50–100 kW; DLC 200+ kW~41 kW20251.1ASHRAE TC 9.9; SemiAnalysis Datacenter Anatomy
inference share of AI compute in 2026 (½ in 2025, ⅓ in 2023); 80–90% of draw at large operators~2/320261.1Deloitte TMT Predictions 2026; McKinsey
active generation + storage in US interconnection queues (end-2025; >1.5x US installed capacity); large-load waits 4–7 yr in top hubs>2,060 GWend-20251.1LBNL, US interconnection queue data through 2025 (May 2026)
all-in cost per 8-GPU H100 server (excl. storage); ~$31k/GPU/yr enterprise all-in$283–318k20251.1SemiAnalysis AI Neocloud Playbook
TCO at 2048-GPU scale, 90% utilization; ~$1.03 small clusters; cloud H100 ~$1.49 (contested — single-source) · contested~$0.74/GPU-hr20251.1SemiAnalysis, 'How much do GPU clusters really cost'
accelerated economic life vs 5–6 yr book life; used GPUs retain ~20–40% residual after 3 yr2–3 yr20251.1Goldman Sachs; CNBC/secondary-market analyses
per dense training rack (GB200 NVL72 ~120–132 kW; GB300 ~142 kW)120–142 kW2026-06-29: 131 → 2026-07-10: 13120251.2NVIDIA OCP / SemiAnalysis
per Rubin Ultra Kyber rack (NVL144) on 800 VDC (announced roadmap)~600 kWH2 2027 (announced)1.2NVIDIA GTC (Jensen Huang); DCD, Tom's Hardware
practical air-cooling ceiling/rack; RDHx ~50–100 kW; DLC 200+ kW~41 kW20251.2ASHRAE TC 9.9; SemiAnalysis Datacenter Anatomy
GB200 NVL72 DLC coolant envelope (rack accepts up to 45 °C supply / 65 °C return, ASHRAE W45-class) and secondary-loop flow≤45 °C / ~170–235 L/min20251.2NVIDIA OCP contribution / Vertiv & vendor NVL72 reference designs
training back-end fabric non-blocking; 2:1 'optimized' cuts back-end cost ~31% (contested — single-source) · contested1:1 vs 2:120251.2SemiAnalysis AI Neocloud Playbook
NVLink5 per-GPU BW (1.8 TB/s bidirectional, ~900 GB/s/dir) vs ~400G scale-out NIC (~50 GB/s/dir) — keep collectives in scale-up~18x20251.2NVIDIA / SemiAnalysis
unplanned interruptions on 16,384 H100s (~1 / 3 hr); 78% hardware-caused419 / 54 days20241.2Meta Llama 3 paper (Table 5)
best-in-class mature H100 cluster MTBF; one failure restarts a synchronous job~7 days / 512 GPUs20251.2SemiAnalysis 100k-H100 clusters
training goodput: industry average / best-in-class; reliability tax ~6–21% of paid GPU-hours, by provider~90% / ~96%20251.2SemiAnalysis ClusterMAX 2.0 / CoreWeave
inference share of AI compute in 2026 (½ in 2025, ⅓ in 2023); 80-90% of draw at large operators~2/320261.3Deloitte TMT Predictions 2026; McKinsey
AI inference capacity to 2030 (~35% CAGR) vs training 23.1 → 62.2 GW (~22%)20.9 → 93.3 GW20261.3McKinsey, 'The next big shifts in AI workloads'
market for inference-optimized chips in 2026; most inference stays in data centers, not at the edge>$50B20261.3Deloitte TMT Predictions 2026
power-oversubscription headroom: inference (uncorrelated per-request peaks) vs training (synchronous peaks)~21% vs ~3%2026-06-29: 21 → 2026-07-03: 2120241.3Patel et al., POLCA (Microsoft Research), ASPLOS 2024
inference fabric oversubscription (vs 1:1 non-blocking for training); 2:1 cuts back-end cost ~31% (contested — single-source) · contested2:1-3:120251.3SemiAnalysis AI Neocloud Playbook; Juniper
HBM3E per Ironwood TPU v7 (inference-era ASIC); 9,216-chip pods, 42.5 FP8 ExaFLOPS, 4,614 FP8 TFLOPS/chip192 GiB / 7.4 TB/s20251.3Google Cloud; SemiAnalysis
self-hosted vs market-avg inference cost per million tokens; ~10x/yr token-price deflation (LLMflation)~$1.90 → ~$2.50/M tok2026-06-29: 1.9 → 2026-07-10: 1.920251.3NVIDIA synthesis; a16z
of wall-clock spent on rollout generation in agentic/reasoning RL post-training~80%20261.42025–2026 RL-systems papers (ROLL Flash, ROLLART)
of compute consumed by rollouts at 16K-token generation length (RLVR long-CoT)~70%20251.4RLVR / long-CoT RL-systems analyses (arXiv)
tokens per RL trajectory for reasoning/agentic tasks — the rollout that dominates cost10K–100K+2026-06-29: 10000 → 2026-07-10: 1000020261.4reasoning-model RL reports
wall-clock speedup of variance-controlled async RL vs synchronous at equal accuracy (~42h vs ~105h)2.5x20261.4Stable Asynchrony / VCPO (arXiv 2602.17616)
just to hold weights for a 70B PPO-RLHF stack (actor + reference + reward + critic), pre-optimizer8–16 GPUs2026-06-29: 12 → 2026-07-10: 1220251.4derived (guide analysis: actor + reference + reward + critic weights)
QLoRA fine-tune on a single 48 GB GPU; memory cut from >780 GB to <48 GB without quality loss65B on 48 GB20231.4QLoRA (Dettmers et al., arXiv 2305.14314)
share of parameters trained by a LoRA adapter vs full fine-tune (model-dependent)~0.1%20261.4LoRA (Hu et al.) / 2026 PEFT practitioner guides
GPU:CPU norm rebalancing toward more CPU per node as agentic RL adds rollout/tool/env load~8:1 → ~2:1 and below20261.4SemiAnalysis
one-way fiber latency from distance alone (~5 ms per 1,000 km); ~1.6 ms RT per 100 mi before any processing~0.80 ms / 100 mi20251.5M2 Optics fiber-latency analysis (4.9 µs/km, n≈1.47)
MEC round-trip at the access edge; under ~50 ms from a regional 5G URLLC breakoutsub-10 ms20251.5ETSI ISG MEC; arXiv 2504.03708 (telco-LLM latency)
perceptibility thresholds: hard real-time / interactive (AR-VR, agentic) / 'instant' conversational~30 / 50 / 100 ms20261.5Spheron hybrid edge guide; AR/VR latency literature
edge data center market, 2026 to 2033, ~14.9% CAGR; AI/ML inference the fastest-growing segment~$40B → ~$106B20261.5Grand View Research; Coherent Market Insights
micro data centers' share of the edge market (global 2025) / of US edge by 2026~35% / ~54%20261.5Grand View Research; Coherent Market Insights (US)
inference share of AI compute in 2026 (½ in 2025); the growth pool the edge competes for~2/320261.5Deloitte TMT Predictions 2026; McKinsey
edge-site deploy time and install-time reduction under zero-touch provisioning (Vapor IO; ZTP fleet tooling)~1 hr / 90%+20261.5Vapor IO; Scale Computing / VMware VCF Edge
practical power envelope per edge micro-site (vs ~132 kW for a centralized NVL72 rack)a few kW – ~30 kW2026-06-29: 30 → 2026-07-07: 3020261.5practitioner ranges
time-to-power: greenfield self-build vs wholesale colo (live 50k+ GPU cluster) vs neocloud24–36 mo / 6–12 mo / days–weeks2026-06-29: 30 → 2026-07-10: 3020261.6SemiAnalysis; JLL 2026 Outlook
brownfield retrofit cost: cooling-only vs full AI retrofit; ~2/3 of pre-2015 DCs unsuitable for frontier density$2–3M / $5–10M per MW2026-06-29: 2.5 → 2026-07-10: 2.520251.6Tetra Tech / Schneider synthesis
global wholesale colo average 2025 (record); ~$120 Atlanta to ~$450 Singapore; ~1% vacancy~$217/kW-month20251.6JLL / CBRE synthesis
self-build TCO at 2,048-GPU scale, 90% utilization (~$1.03 small clusters) vs neocloud median ~$2.3–3.5/hr (contested — single-source) · contested~$0.74/GPU-hr20251.6SemiAnalysis, 'How much do GPU clusters really cost'
neocloud GPU rental vs hyperscaler pricing (8-GPU node ~$34/hr neocloud vs ~$98/hr hyperscaler)40–85% below20261.6SemiAnalysis H100 Index / AM Compute
rise in the 1-year H100 rental contract index, Oct 2025 to Mar 2026, as capacity tightened; on-demand largely sold out~+40%20261.6SemiAnalysis H100 Rental Index
breakeven utilization for a debt-financed cluster; swings -$330k to +$340k/mo (55% vs 85%) on a 1,024-GPU H100 build (contested — single-source) · contested~70%20251.6AM Compute / McKinsey
US large-load grid interconnection lead time end-to-end; up to ~10 yr in worst queues — the gate behind self-build~3–7+ yr20261.6LBNL Queued Up; ERCOT / PJM filings
practical air-cooling ceiling per rack; RDHx ~50–100 kW; DLC 100–200 kW+~41 kW20251.7ASHRAE TC 9.9; SemiAnalysis Datacenter Anatomy
per GB200 NVL72 rack (~115 kW liquid + ~17 kW air); GB300 ~142 kW; Rubin Ultra Kyber ~600 kW~120–132 kW2026-06-29: 126 → 2026-07-02: 12620251.7115/17 split: HPE QuickSpecs (OEM); 120 kW cooling: NVIDIA; ~132 kW all-in: Vertiv/451
GB200 NVL72 DLC coolant envelope (rack accepts up to 45 °C supply / 65 °C return, ASHRAE W45-class) and secondary-loop flow≤45 °C / ~170–235 L/min20251.7NVIDIA OCP contribution / Vertiv & vendor NVL72 reference designs
training non-blocking vs inference oversubscribed; 2:1 cuts back-end cost ~31% (contested — single-source); Meta ran 7:1 on 24k H100 · contested1:1 vs 2:1–3:120251.7SemiAnalysis Neocloud Playbook; Juniper AI-cluster design; Meta
GPU:CPU ratio shifting from training-era norm toward agentic-inference host demand~8:1 → ~2:1 and below20261.7TrendForce Insights
full AI liquid retrofit cost crossing the cooling cliff; still strands capacity~$5–10M/MW2026-06-29: 7.5 → 2026-07-10: 7.520261.7Tetra Tech / Schneider synthesis
capital premium for Tier IV (fault-tolerant) over Tier III (concurrently maintainable)~20–40%20251.7Uptime Institute Tier classification; industry cost syntheses
goodput (effective training time): industry avg vs best-in-class; reliability tax ~6–21% of paid GPU-hours, by provider~90% / ~96%20251.7SemiAnalysis ClusterMAX 2.0 / CoreWeave
1 GW AI data center: total-program capex (core stack ~$27.9/W plus land, build-out, financing) and all-in annual TCO (~$8.5M/MW-yr)~$38B / ~$8.5B/yr20261.8Epoch AI, AI datacenter cost breakdown
1 GW annual TCO at 3-yr / 5-yr / 7-yr IT useful life — the dominant lever$12B / $8.5B / $7B20251.8Epoch AI / AM Compute synthesis
self-operated TCO at 2048-GPU scale, 90% util; ~$1.03 small clusters (contested — single-source) · contested~$0.74/GPU-hr20251.8SemiAnalysis, 'How much do GPU clusters really cost'
breakeven utilization (debt-financed); 1,024-GPU cluster swings -$330k to +$340k/mo (contested — single-source) · contested~70%20251.8AM Compute / McKinsey
LLMflation: inference cost decline at fixed quality (Epoch Mar-2025: ~50x/yr median; ~200x/yr post-2024 models)~10x/yr2024-20251.8a16z (Nov 2024); Epoch AI (Mar 2025)
AI-app gross margin vs 70-90% for mature SaaS~41% to ~52%20261.8ICONIQ State of AI 2026; Bessemer
wholesale colo global avg 2025; BTS/CTL ~$150-220/kW-mo over 15 yr~$217/kW-mo20251.8JLL / CBRE synthesis
estimated understated AI D&A 2026-2028 (CONTESTED); industry AI D&A ~$400B/yr · contested~$176B20261.8Burry / secondary analyses; filings
large power transformer lead time (~144 wk GSU); up to ~5 yr in constrained markets — the schedule-dominating long pole~128 wk20252.1Wood Mackenzie Q2 2025 survey / pv magazine
large-load grid interconnection, application to energization; up to ~10 yr in the worst queues~3–7+ yr20262.1LBNL Queued Up; ERCOT / PJM filings
AI data-center shell-to-MEP-complete construction — fast vs the power track, so rarely the binding constraint12–18 mo20262.1Data Center Knowledge; Mastt
of the ~12 GW US capacity targeted for 2026 actively under construction by early 2026; the rest exposed to slippage~1/320262.1Industry construction tracking
Level-5 integrated systems testing for a liquid-cooled AI hall (vs 4–6 wk air-cooled) — the un-compressible commissioning tail10–14 wk20262.1Construct & Commission / 2026 outlook synthesis
best-in-class fleet failure rate after burn-in; new clusters fail far more for the first 3–4 weeks — the bring-up tail~1 failure / 512 GPUs / week20252.1SemiAnalysis (100k H100 clusters)
annual revenue per GW of AI capacity — so ~200 MW landing 6 months early is worth ~$1–1.2B; the schedule's dollar value (contested — single-source) · contested~$10–12B20252.1SemiAnalysis (onsite gas economics)
PJM interconnection deposits committed before the design is frozen: a fixed study deposit (hundreds of thousands, non-refundable once the request completes) plus a $15k/MW readiness deposit -- not a percentage$500k study + $15k/MW20262.1PJM large-load Expedited Interconnection Track filing / Inside Lines
build-cost threshold below which GMP procurement dominates; larger programs lean EPCM/progressive<$500M20262.2King & Spalding, Building Data Centers – Key Considerations
global average data-center construction cost (base building); tenant AI fit-out adds up to ~$25M/MW$11.3M/MW20262.2JLL 2026 Global Data Center Outlook
mainstream hyperscale / build-to-suit construction benchmark before buyer-specific scope$10–12M/MW20262.2JLL / industry synthesis
HV/substation transformer lead time (≈144 wk GSU; up to ~60 mo constrained) — OFE-by-default, pre-GC~128 wk20252.2Wood Mackenzie Q2 2025 survey / pv magazine
H100/H200 GPU lead times through 2025–2026 — bimodal in 2026: spot H100s improving to weeks while H200 and full enterprise systems run 6–12 mo; why GPUs are owner-furnished and on allocation~3–12 mo20262.2Industry synthesis (Buildermuse / vendors); Network World on H200 allocation (2026)
revenue per GW of AI capacity per year — the prize the delivery model protects by saving months (contested — single-source) · contested~$10–12B20252.2SemiAnalysis (onsite gas economics)
downtime the commissioning agent's IST exists to prove (vs Tier III ~1.6 hr/yr)Tier IV ~26 min/yr20252.2Uptime Institute
data-center load dropped in the six-fault, 82-second July-2024 Northern Virginia disturbance — the failure independent Cx and OE exist to prevent~1.5 GW in 82 s2024-07 (event); 2026 (alert)2.2NERC incident review / Level 3 Alert (2026); Utility Dive
GSU transformer lead time (standard power transformer ~128 wk); up to ~60 mo in constrained markets~144 weeks2026-06-29: 144 → 2026-07-12: 1442025-Q22.3Wood Mackenzie T&D survey / pv magazine; VAWN index concurs at 144 wk (2026-06)
HV switchgear & circuit-breaker lead time — HV breakers ~100–155 wk with little improvement reported since the ~151-wk late-2023 peak; HV GIS and full substations 3–5 years; MV eased separately~100–155 wk2025-112.3Duke Nicholas Institute / Wood Mackenzie via The Conversation (2025-11); Build.inc (2026)
GE Vernova gas-turbine equipment backlog + slot reservations (44 GW firm + 56 GW reservations); ≥110 GW targeted by end-2026100 GW2026-Q12.3GE Vernova Q1-2026 results (22 Apr 2026)
diesel genset lead time (1.25–3.25 MW class); largest engines to ~78 wk52–70 weeks2026-07-01: 61 → 2026-07-12: 612026-052.3GlobalPwr (2026-05): 52–70 wk for 1,250–3,250 kW; JLL 2026 Outlook: >50 wk Americas, >75 wk EMEA; CPM Pros
through 2026; ~30% supply gap; +15–20%/quarter price risesHBM3E sold out20262.3SemiAnalysis / TrendForce
TSMC CoWoS 2026 demand ~fully booked; NVIDIA holds ~50–60% of allocation~1M wafers20262.3SemiAnalysis / Silicon Analysts
US Section 232 duty on semi-finished copper (Aug 2025); electrical grid equipment got a 15% transitional rate through 202750% tariff2025–20262.3White House / CBP / Congress.gov
JLL global / US average data-center equipment lead time across packages~33–42 weeks20252.3JLL equipment lead-time index
Blackwell (B200/GB200) allocation backlog for new orders — the last cleanly reported order-to-delivery lead time; NVIDIA has since communicated allocation status, not weeks~12 mo2024-102.3Morgan Stanley via Data Center Dynamics (2024-10-11)
NVIDIA cloud GPU availability, all shipping generations — 'Blackwell sales are off the charts, and cloud GPUs are sold out'sold out2025-112.3NVIDIA Q3 FY2026 results (2025-11-19)
NVIDIA advanced-architecture supply-tightness horizon — purchase commitments secured into CY2027, 'further out in time than usual'into CY20272026-022.3NVIDIA Q4 FY2026 earnings call (2026-02-25)
Vera Rubin (VR200 NVL144) availability — first customer samples shipped Feb 2026; HBM4 ramp is the schedule risk (TrendForce trimmed Rubin's 2026 shipment share 29%→22%)production 2H 20262026-042.3NVIDIA Q4 FY2026 earnings call (2026-02-25); TrendForce via SDxCentral (2026-04-09)
AMD Helios / MI400-series rack availability — MI455X UALoE72 low volume 2H 2026, mass ramp Q2 2027; AMD Q1 2026 call: on track, with anchor deals (OpenAI 6 GW) absorbing early supply2H 20262026-07-12: 2026 → 2026-07-12: 20262026-052.3The Next Platform (2026-02-23); AMD Q1 2026 earnings call (2026-05); SemiAnalysis on volume timing
MV switchgear lead time (5–38 kV, data-center class) — eased to ~26–32 wk in early 2025, re-tightened on AI demand; some quotes 40–80 wk~44–65 wk2026-07-12: 54.5 → 2026-07-12: 54.520262.3CPM Pros (T&T 2025–26 context); VAWN index 44 wk avg (2026-06); Build.inc 40–65 wk (2026); DEI Power (2026-05); GEP (2025-01)
large LV switchboard lead time (1,200 A+) — commodity gear back to 6–10 wk; corroborated by VAWN (switchboards 32–41 wk, LV switchgear 54 wk, 2026-06); engineered draw-out quoted to 70–80 wk~35–62 wk2026-052.3DEI Power (2026-05); VAWN electrical lead-time index (2026-06)
large 3-phase UPS lead time (data-center class) — spread is wide by vendor and configuration; direct procurement cut one 32-wk quote to 10~40–72 wk20262.3CPM Pros; DPR Construction Q2 2025; Core Scientific FY2025 10-K (critical electrical/cooling 12–18+ mo)
large chiller lead time (data-center class) — standard 20–30 wk, 500+ ton centrifugal/screw 40–60 wk; the R454B/R1234ze refrigerant transition is the structural risk~20–60 wk2026-032.3The Cooling Report supply-chain guide 2026; CPM Pros
cooling tower / dry cooler lead time — flagged as a single-vendor point-of-failure risk; dual-source heat rejection~20–30 wk2026-032.3The Cooling Report supply-chain guide 2026
CDU lead time (direct-to-chip liquid cooling) — 16–24 wk standard, 26–52 wk in the current queue; DLC market +156% YoY in 2Q25 (Dell'Oro)~16–52 wk2026-032.3The Cooling Report supply-chain guide 2026; CPM Pros; Dell'Oro
busway / busduct lead time — standard runs average ~27 wk (VAWN 2026-06), custom engineered runs at the top of the range; fabrication can't start until final dimensioned drawings~27–52 wk20262.3CPM Pros; VAWN electrical lead-time index (2026-06)
distribution / pad-mount transformer lead time (sub-HV class) — distribution average ~30 wk (vs 100+ wk 2023 peak); pad-mount three-phase delays growing, dry-type 20–40 wk~20–40 wk20252.3Wood Mackenzie Q2 2025 supply-chain survey via EEPower; POWER Magazine (2026-01)
structural steel fabrication lead time (data-center shells) — max quote moved 6 → 14 wk with mills at ~80% capacity; the shell is not the long pole~6–14 wk2025-092.3DPR Construction Q3 2025 Market Conditions Report, citing Schuff Steel
800G optical transceiver lead time — volume orders ~36–50 wk (floor raised after NVIDIA's 2026-03 $4B EML pre-buy), non-priority buyers quoted past 2027; EML lasers are the constraint, 800G supply 40–60% below demand through 2027~36–50 wk2026-052.3McKinsey via TechTimes (2026-05); Supply Chain Connect (2026); TrendForce (2026-04)
MV cable lead time — standard copper MV often stocked domestically, LSZH builds ~10–12 wk; rarely the long pole, but a hyperscale campus consumes 20–50 miles and copper price is the real exposure (distributor-reported, single source)≤12 wk20262.3DistributorWire (2026, distributor-reported)
GB300 (Blackwell Ultra) availability — the volume platform of 2026; AI servers +28% YoY, GPUs 69.7% of shipments, ASIC-based servers 27.8%drives most 2026 shipments2026-012.3TrendForce (2026-01-20)
Blackwell vs Rubin 2026 shipment mix — TrendForce raised Blackwell to ~71% and cut Rubin to 22% on HBM4 validation, the ConnectX-8→9 migration, power and cooling stepsBlackwell ~71% of 20262026-042.3TrendForce via The Register (2026-04-08)
AMD MI355X availability — ramping through 1H26; Oracle first hyperscaler with public instances, building a 27,000+ node MI355X clusterramping through 1H262025-112.3AMD Q3 2025 earnings call (2025-11)
AWS Trainium3 availability — shipping since start of 2026 and nearly fully subscribed; Trainium2 largely sold out, much of Trainium4 reserved ~18 months aheadnearly fully subscribed2026-042.3Amazon Q1 2026 earnings, Andy Jassy (2026-04-30)
SK hynix memory supply — entire planned 2026 supply of key memory products (HBM, DRAM, NAND) pre-sold; ~56% HBM share, NVIDIA's lead suppliersold out through 20262025-102.3SK hynix Q3 2025 earnings via CNBC (2025-10-29)
Micron HBM supply — calendar-2026 HBM fully booked including HBM4, meeting only ~50–66% of core-customer demand; FY26 capex raised to $20BCY2026 fully booked2025-122.3Micron FQ1 2026 earnings via TrendForce (2025-12-18)
HBM4 ramp status — Samsung first to volume (>$1B revenue within ~4 months of Feb 2026 mass production); SK hynix meaningful volume shifts to 3Q26volume ramping 20262026-062.3TrendForce (2026-06-23)
high-speed networking supply — lead times for advanced substrates, optics, and high-speed networking components lengthening into 2026; no public weeks figure for AI switches themselveslead times lengthening2025-122.3Dell'Oro Group 2026 predictions (2025-12-16)
Broadcom Tomahawk 6 (102.4 Tb/s) availability — production volume from March 2026, sample-to-production in under three quarters on 'unprecedented' demand; Arista TH6-based 1.6T switches ship from Q4 2026volume ramp Mar 20262026-032.3Broadcom press release (2026-03); Arista PR (2026-06-09)
ConnectX-9 1.6 Tb/s NIC availability — the ConnectX-8→9 migration is a named reason TrendForce cut Rubin's 2026 shipment share from 29% to 22%Rubin ramp risk2026-042.3TrendForce via The Register (2026-04-08)
typical performance-LD cap as a share of EPC contract price (range 5-25%); ~10-20% overall liability cap10-15%20252.4Womble Bond Dickinson; EPC-market practice
HV / GSU power-transformer lead time setting equipment-delivery risk in supply agreements; up to ~60 mo constrained~128-144 wk2025-Q22.4Wood Mackenzie / pv magazine
large-load grid interconnection lead time the LGIA/ISA milestones must bind around; up to ~10 yr worst queues~3-7+ yr20252.4ERCOT / PJM filings synthesis
security deposit demanded under a US large-load tariff, paired with a 15-yr contract and minimum-demand charges$2.7M / MW20252.4Davis Graham / utility large-load tariff filings
take-or-pay floor and minimum term in Ohio's large-load tariff (pay 85% of contracted capacity, 12-yr minimum)85% / 12 yr20252.4Utility Dive / Ohio PUCO tariff
transmission-service options FERC ordered PJM to offer co-located load (firm / non-firm contract-demand, interim non-firm, NITS)4 options20252.4FERC PJM co-located load order (Dec 18, 2025)
rack-level uptime SLA already offered by leading neoclouds (CoreWeave, Oracle); service-credit-backed99%20252.4SemiAnalysis ClusterMAX 2.0
US states with at least one approved large-load tariff governing data-center grid contracts (7 more pending)23 states2026-06-29: 23 → 2026-07-02: 232026-052.4Columbia Climate Law (Sabin Center) / state tariff tracking
CoreWeave DDTL 4.0 (Mar 2026): first investment-grade-rated GPU-backed financing; SOFR+2.25% floating / ~5.9% fixed, matures 2032$8.5B / A3 / A(low)20262.5CoreWeave investor release; Moody's / DBRS; Bloomberg
data-center build 2025-2028 and the financing gap beyond hyperscaler cash flow~$2.9T / ~$1.5T20252.5Morgan Stanley Research, Bridging the Data Center Gap
data-center ABS issuance 2025 → projected 2028 (first refinancing wave); 2026-27 ABS/CMBS ~$30-40B/yr in other syntheses~$8B → ~$25B20262.5Morgan Stanley; RBCCM / SFA synthesis
breakeven utilization for a debt-financed cluster; 1,024-GPU H100 swings -$330k to +$340k/mo (the DSCR driver) (contested — single-source) · contested~70%20252.5AM Compute / McKinsey
GPU residual after 3 yr (CONTESTED); rental rates -64-75% from peak — the assumption under every ABS/residual structure · contested~20-40%20252.5Hashrate Index / CNBC synthesis
CoreWeave backlog vs revenue — single-anchor concentration; FY25 rev $5.13B, interest ~40% of adj. EBITDA~$66.8B (~13x)2026-06-29: 66800000000 → 2026-07-02: 668000000002025-20262.5CoreWeave Q4/FY2025 earnings press (SEC EDGAR)
energy: largest controllable opex line in the 1 GW model — why the PPA is a financing precondition~$0.6B/yr20262.5Epoch AI / SemiAnalysis synthesis
CoreWeave DDTL cost-of-capital compression, 2023 (DDTL 1.0 floating) → 2026 (DDTL 4.0 fixed) as the sponsor de-risked~15% → ~5.9%20262.5Bloomberg / DCD; company releases
projected global data-center-linked insurance premium by 2030 (from ~$10.6B), fastest-growing property line~$24.2B20262.6Swiss Re Institute (sigma)
data-center construction-related coverage premium projected for 2026 — ~2x the global aviation market~$10B2026-06-29: 10000000000 → 2026-07-07: 1000000000020262.6S&P Global Ratings
share of total data-center loss costs now liquid/water-related (near-nonexistent in air-cooled halls)~24%20262.6Swiss Re
fire's share of loss cost vs share of loss events — rare but catastrophic (15-yr study)42% / 11%20252.6FM (FM Global)
single-program lifecycle capacity (CAR + DSU + operational PD/BI) assembled because single carriers can't cover a hyperscale campus$3.5B20262.6Aon / Carrier Management
US data-center capacity in 3+ large-hail-day zones / significant tornado zones — the nat-cat accumulation problem~25% / ~40%20262.6Swiss Re
typical cost of GPU residual-value insurance, netted from the loan's interest yield; covers shortfall to ~80% of the warrantied floor~150 bps/yr20252.6American Compute / USD.AI
annual data-center cyber-insurance premium increases, 2022–202525–40%/yr20262.6Hotaling Insurance Services
calibration agreement of mature data-center CFD vs measured readings, given accurate boundary conditions1–5%20252.7Cadence Reality / Future Facilities 6SigmaDCX; independent validation (Compass, Binghamton)
GB200 NVL72 heat split — removed by liquid vs by air; the CFD must validate both loops~115 / ~17 kW2026-06-29: 115 → 2026-07-07: 11520252.7HPE QuickSpecs (OEM 115/17 split); 132 kW all-in Vertiv/NVIDIA + 451 Research
GPU throttle on a coolant inlet/flow deviation outside the DLC envelope — the cost of a wrong thermal modelup to ~50%20252.7NVIDIA GB200 NVL72 DLC spec
share of site demand from cooling (VFD-driven motor load) that injects harmonics into power-system studies~40%20252.7PSC Consulting / data-center load modeling
back-end fabric cost avoided by oversubscribing 1:1 → ~3:1 where the workload allows — the fork fabric sim retires (contested — single-source) · contested~31%20252.7SemiAnalysis Datacenter Anatomy
industry-average vs best-in-class training goodput — the spread a discrete-event model lets you size against~90% / ~96%20252.7SemiAnalysis ClusterMAX 2.0 / CoreWeave
best-in-class H100 fleet failure rate feeding the goodput/availability model (new clusters far worse)~1 / 512 GPUs / wk20252.7SemiAnalysis (100k H100 clusters)
first NVIDIA Omniverse DSX reference designs validated as a high-fidelity digital twin before construction beginsgigawatt-scale20252.7NVIDIA Omniverse DSX / Vera Rubin DSX blueprint
US generator interconnection queue, active end-2025 (gen + storage; down from ~2,290 GW end-2024)>2,060 GWend-20253.1LBNL, US interconnection queue data through 2025 (May 2026)
ERCOT large-load interconnection queue, mid-2026 (~89% data centers; 63→233→438+ GW arc)438+ GW2026-063.1ERCOT news release (18 Jun 2026)
PJM application-to-commercial-operation timeline (vs <2 yr in 2008)>8 yr20253.1RMI / PJM filings
AI data center construction time — the grid, not the build, is the long pole12–18 mo20263.1Data Center Knowledge; Mastt
revenue per GW of AI capacity per year; ~$1–1.2B for 200 MW online 6 mo early (contested — single-source) · contested~$10–12B20253.1SemiAnalysis (onsite gas economics)
behind-the-meter gas announced cumulatively by 2026 (~82 GW announced in 2025; ~2 GW operating today)~90 GW2026-073.1Cleanview, 'Bypassing the Grid' (mid-2026)
practical sub-50 ms inference reach (50 ms RTT ⇒ ~5,000 km propagation-only; switching/queuing/compute overhead and the ~1.3–1.5× fiber route factor cut it to practical reach)~1,000–1,500 km2026-07-07: 1250 → 2026-07-10: 125020253.1derived (guide analysis; fiber/latency, route factor → 3.6)
fully-built cost of a gigawatt-scale AI campus incl. IT (all-in ~$30/W; ~$38B total-program with land, build-out, financing)~$30–40B20263.1cross-ref: Ch 6.3 (McKinsey ~$30/W) / Ch 1.8 (Epoch AI ~$38B)
US interconnection queue end-2025 (end-2024: ~2,290 GW = ~1,400 GW gen + ~890 GW storage, ~10,300 projects)>2,060 GWend-20253.2LBNL, US interconnection queue data through 2025 (May 2026)
median interconnection request-to-COD for projects built 2018–2024 (doubled from <2 yr in 2000–2007)>4 years20253.2LBNL, Queued Up: 2025 Edition
large-load interconnection end-to-end; PJM application-to-COD has exceeded 8 yr; up to ~10 yr in worst queues~3–7+ yr20263.2LBNL Queued Up; ERCOT / PJM filings
large-load requests ERCOT is tracking (~89% data centers; 63 GW end-2024 → 233 GW end-2025 → 438 GW Jun 2026)438+ GW2026-063.2ERCOT news release (18 Jun 2026)
PJM 2018–2020 cohort that withdrew before an interconnection agreement; ~24% reached operation65–80%2026-07-07: 72.5 → 2026-07-10: 72.520253.2PJM queue analysis
new US load integratable at just 0.5% annual curtailment (avg event ~2 hr) — the flexibility headroom~98–100 GW20253.2Duke Nicholas Institute / Latitude Media
HV power-transformer lead time (~128 wk standard, ~144 wk GSU, up to ~60 mo constrained); the co-equal gate~128–160+ wk2025–20263.2Wood Mackenzie / pv magazine
AI-capacity revenue value of speed; energizing 200 MW six months early ≈ $1–1.2B incremental revenue (contested — single-source) · contested~$10–12B/GW/yr20253.2SemiAnalysis (onsite gas)
power as a share of AI data center operating cost; the largest single opex line at scale25-60%2026-06-29: 42.5 → 2026-07-10: 42.520263.3Epoch AI 1 GW TCO model
energy cost in the canonical 1 GW model — the single largest opex line~$0.6B/yr20263.3Epoch AI / SemiAnalysis synthesis
average commercial rate in the top-10 US data center states (≈ identical to other states)~14.4-14.5¢/kWh2025-263.3Electric Choice / RealClearEnergy synthesis
Texas/ERCOT commercial rate — among the lowest in the US on the deregulated market~8.7¢/kWh20263.3Electric Choice
new Virginia data-center electricity consumption tax (effective Jul 1 2026, sunsets Jul 1 2028) — an added cost line$0.011/kWh20263.3Virginia HB30 (2026 budget, Item 3-5.24); Data Center Knowledge
revenue per GW of AI capacity — the prize that power cost is netted against (contested — single-source) · contested~$10-12B/GW/yr20253.3SemiAnalysis (onsite gas)
large-load interconnection lead time; up to ~10 yr in worst queues — the availability gate~3-7+ yr20253.3ERCOT / PJM filings synthesis
US grid headroom integratable at 0.5% annual curtailment — the basis for flexible/non-firm deals~98-100 GW20253.3Nicholas Institute / Duke 'Rethinking Load Growth' synthesis
behind-the-meter data-center generation announced cumulatively by mid-2026 (~82 GW of it since Jan 2025); ~2 GW online → 2.8-3.2 GW by year-end (→ 4.9)~90 GW2026-073.4Cleanview, 'Bypassing the Grid' (mid-2026)
grid interconnection lead time (large load); PJM application-to-COD over 8 yr; up to ~10 yr worst queues~3–7+ yr20263.4LBNL Queued Up; ERCOT / PJM filings
AI revenue per GW; energizing 200 MW six months early worth ~$1-1.2B (contested — single-source) · contested~$10-12B/GW-yr2026-06-30: 11 → 2026-07-10: 1120253.4SemiAnalysis
Google global 24/7 CFE score 2024; 9 of 20 regions above 80%, Asia-Pacific ~12%~66%20253.4Google 2025 Environmental Report / Moving toward 24x7 CFE
nuclear contracted by hyperscalers: TMI/Crane 835 MW (~2028), Amazon-Talen ~2 GW, Kairos SMR ~500 MW (~2030-35)>10 GW2025-20263.4DataCenterDynamics; SMR Intel; filings
Q1 2026 projected VPPA settlement value, wind vs solar — solar cannibalized to ~zero~€15 vs ~€0/MWh20263.4NZCB VPPA Opportunity Index
industrial gas-fired LCOE for on-site/BTM power, above projected ERCOT grid prices>~$120/MWh2026-06-29: 120 → 2026-07-10: 12020263.4ISO/EPRI
PJM contract-demand transmission services (firm / non-firm / interim non-firm) created by FERC co-location order, compliance from early 2026, ~2028 transition3 new20253.4FERC order Dec 18, 2025
revenue per GW of AI capacity; ~$1–1.2B from getting 200 MW online 6 months early (contested — single-source) · contested~$10–12B/GW/yr20253.5SemiAnalysis (onsite gas)
planned US behind-the-meter capacity tracked; ~90% announced in 2025; ~2 GW online mid-2026, 2.8–3.2 GW by year-end~90 GW / 59 projects2026-06-29: 56 → 2026-07-03: 902026-07-033.5Cleanview (Bypassing the Grid)
reciprocating-engine simple-cycle efficiency (highest of any combustion tech); ~2-min start, 10% min load, 12–18 mo lead45–48%2026-06-29: 46.5 → 2026-07-10: 46.520263.5POWER Magazine; Wärtsilä
H/HA-class combined-cycle efficiency vs 37–48% simple cycle; 3–5+ yr lead, 24–36 mo install>60–64%20253.5Mitsubishi Power; Grid Capacity Intelligence
Bloom–Oracle SOFC framework (1.2 GW initially contracted); near-zero criteria pollutants; ~$3,000–4,000/kWup to 2.8 GW20263.5Bloom Energy IR / press
aeroderivative turbine lead time (refurb under 12 mo; some cores quoted to ~243 weeks) — new-build order slots quoted for 2028–2030 as of 202618–36 mo+2026-06-29: 27 → 2026-07-12: 2720253.52025 synthesis; Data Center Frontier / power-eng aeroderivative coverage (2026)
solar/wind overbuild for firm off-grid 24/7; firmed ideal-site solar ~$113/MWh (2025) → ~$77/MWh (2030) (contested — single-source) · contested~7x20253.5Princeton ZeroLab 24/7 CFE firming modeling
SMR first-of-a-kind capex; restarts ~2027–28, SMR scale 2032–2035; ~10% of DC demand growth by 2035$6,400–12,700/kW2024–20263.5SMR Intel / Deloitte synthesis
one-way propagation delay in single-mode fiber (index ~1.47); ~5 us/km rule of thumb, ~10 us/km round-trip~4.9 us/km20253.6M2 Optics; MapYourTech (5-microsecond rule)
round-trip propagation floor added by 1,000 km of fiber, before any switching or queuing~10 ms20253.6M2 Optics; fiber-latency physics
per-hop fabric latency: InfiniBand ~1–2 us, tuned RoCEv2 ~1.5–2.5 us (microseconds, vs milliseconds for WAN)~1–2.5 us20253.6SemiAnalysis / NVIDIA
less inter-site data exchanged by DiLoCo-class intermittent sync vs full synchronous DP — what makes cross-region training viable100–500x20253.6Google DeepMind DiLoCo; Epoch AI
end-to-end target for latency-sensitive inference; ~100 ms is the human 'instantaneous' thresholdsub-50 ms20263.6HumAI; practitioner edge-inference benchmarks
of candidate sites eliminated at the fiber-diligence stage in many markets — for route diversity, not power~50%20253.6Global Data Center Hub (fiber-as-bottleneck)
construction cost for dual diverse new-build fiber routes, before the permitting and make-ready tail~$150k/mile20253.6Global Data Center Hub; Netrality
fiber pairs per path, on physically diverse routes, below which AI-grade sites are viewed as functionally stranded24–48 pairs20253.6Global Data Center Hub
indirect (Scope 2, grid) water vs direct (Scope 1, on-site) water for US data centers~12x20233.7LBNL 2024 US Data Center Energy Usage Report
US data-center direct water consumption, 2023 actual to 2028 projected range17 → 38-73 bn gal2026-06-29: 17 → 2026-07-07: 172023-20283.7LBNL; EESI synthesis
peak-summer withdrawal for a ~100 MW evaporatively-cooled campus (WUE-derived; Google Council Bluffs, a larger campus, consumed ~2.8M gal/day, 2024)~1.5-3 M gal/day2024-20253.7Google water disclosures; Apstech Advisors
Microsoft FY2025 fleet WUE; next-gen closed-loop designs approach ~0~0.27 L/kWh2026-06-29: 0.3 → 2026-07-02: 0.2720253.7Microsoft FY2025 Environmental Sustainability Report / Microsoft Datacenters efficiency disclosure
industry-average WUE (evaporative); best-in-class 0.3-0.7; closed-loop ~0~1.8-1.9 L/kWh20253.7Vertiv / NREL synthesis
Texas average data-center water intensity; ~49 bn gal statewide by end-2025~793 gal/MWh2026-06-29: 793 → 2026-07-02: 79320253.7Houston Advanced Research Center (HARC), 'Thirsty Data and the Lone Star State'
annual free-cooling hours in a Nordic climate; PUE as low as ~1.09~7,000-8,000+ hr20253.7w.media; Nordic operator disclosures
US state data-center water bills proposed per session; ~2-4 enacted 2025-202630+ bills / 13-16 states2025-20263.7MultiState legislative tracking
acres per MW of IT on GW-scale AI campuses (full campus incl. substation/cooling yard)0.7–0.92025–20263.8Reuters / operator filings (Hyperion, Stargate)
GB200 NVL72 wet rack weight (≈3,000 lb) — sets the structural floor-loading basis~1.36 t20253.8NVIDIA OCP / OEM datasheets (HPE, Lenovo, Supermicro)
uniform live-load basis for liquid-cooled AI halls (vs ~150 psf legacy); column loads >1,000 kips250–500+ psf20253.8Geopier; StructureMag; TechTarget
flood hazard area for Risk Category II+ under ASCE 7-22 Supplement 2; ASCE 24 freeboard on critical facilities500-yr (0.2%)2024–20253.8ASCE/SEI 7-22 Supplement 2; ASCE 24
code-assigned Risk Category for a typical data center is II; mission-critical owners often elect RC III/IV seismic performance (importance factor, operability) as an owner requirement, not a code mandateRC II code / III–IV elective20253.8ASCE 7 / IBC; Langan
Phase I ESA cost per site (ASTM E1527-21); ~2–3 weeks to report; higher for industrial brownfields$2,200–4,00020263.8A3E; industry practitioner ranges
share of gross campus acreage that is non-buildable on a typical large parcel (setbacks, buffers, floodplain, infra yards)~50%20253.8Practitioner developable-ratio heuristic
distinct approvals/permits/sign-offs across federal, state, grid, and local layers for a large campus50+20263.9CRS R48762; Dgtl Infra / practitioner build-lifecycle guides
EO 14318 Qualifying-Project thresholds for accelerated federal permitting & new categorical exclusions$500M / 100 MW20253.9EO 14318 (White House); Federal Register 90 FR 35385
Seven County v. Eagle County ruling narrowing NEPA scope; explicitly names data centers8-020253.9U.S. Supreme Court (23-975), May 29 2025
EU EED data-center reporting threshold (~31 data points incl. PUE/WUE/ERF/REF; due 15 May annually)≥500 kW2024 (in force)3.9European Commission; Delegated Reg. (EU) 2024/1364
large-load grid interconnection lead time; up to ~10 yr in worst queues — the usual long pole~3-7+ yr20253.9ERCOT / PJM filings synthesis
FERC RM26-4 proposed large-load federal-interconnection threshold (DOE Sec. 403); on 18 Jun 2026 FERC opened §206 show-cause proceedings against all six RTOs/ISOs, docket left open>20 MW2026-063.9FERC RM26-4; §206 show-cause orders (18 Jun 2026)
air-permit lead time: minor-source 6-12 mo vs PSD/nonattainment NSR 18-36+ mo for combustion bridge power6-36+ mo20263.9EPA CAA resources for data centers; Nixon Peabody
typical shell construction time — shorter than interconnection, so 'built' precedes 'energized'12-24 mo20263.9Data Center Knowledge; practitioner build-lifecycle guides
Texas sales-tax revenue forgone to the data-center exemption (~$1B+/yr)~$3.2B / 2 yr20263.10Texas Comptroller; Texas Tribune
Georgia FY26 estimated exemption cost — a 664% jump over the prior $327M estimate~$2.5B2026-06-29: 2500000000 → 2026-07-02: 250000000020263.10GA Governor's Office of Planning & Budget, via AJC / Good Jobs First; Stateline
Virginia data-center sales-tax exemption cost; Senate proposed phase-out to Jan 2027~$1.6B/yr2026-06-29: 1600000000 → 2026-07-10: 160000000020263.10Virginia JLARC; Stateline
US states moving to pause, cap, condition, or repeal data-center tax breaks~2 dozen states20263.10Stateline; MultiState
Texas property-tax abatement tiers over 10 yr; Virginia local reductions reach ~80%50% or 75%2026-06-29: 50 → 2026-07-10: 5020253.10Texas Comptroller
take-or-pay floor in the Oregon POWER Act / Schedule 96 large-load tariff template~90% min-demand20263.10Oregon PUC; Utility Dive; OPB
share of distribution/transmission upgrade borne by the load under the large-load tariff template100% upgrade cost20263.10Oregon PUC; Environment+Energy Leader (23 states)
indicative individual GBER aid ceiling per undertaking in the draft 2026 EU state-aid regime€30M / project20263.10European Commission GBER consultation; Hogan Lovells
data-center projects blocked or delayed in Q1 2026 (75+ projects) — matching all of 2025 in one quarter~$130BQ1 20263.11Data Center Watch / 10a Labs; NBC News, Tom's Hardware
active local opposition groups across 49 states (more than doubled YoY)833Q1 20263.11Data Center Watch / 10a Labs
data-center share of PJM's $16.4B capacity-auction cost — the cost-shift driver~$6.5B (40%)20263.11PJM market monitor; Utility Dive
typical residential bill increase attributed in part to data-center demand (Dominion; PJM hubs $16-18/mo)~$16/mo20263.11Inside Climate News; IEEFA
with at least one approved large-load tariff (7 more pending) ring-fencing DC costs23 states2026-06-29: 23 → 2026-07-02: 232026-053.11Columbia Climate Law (Sabin Center) / state tariff tracking
Americans opposing a local data center / worried it raises their energy bill~70% / 78%20263.11Consumer Reports; survey synthesis
state data-center bills filed in 30+ states in the first six weeks of 2026300+ bills20263.11MultiState; ArentFox Schiff
representative property-tax abatement / PILOT term (Arkansas IDB-financed example)up to 65% / 30 yr20263.11NW Arkansas Democrat-Gazette; NCSL
US AI-chip export policy changes in 12 months (Diffusion Rule published → rescinded → H20 gated → un-gated w/ tax → China case-by-case)5 regime shiftsJan 2025–Jan 20263.12US BIS notices; SemiAnalysis AI Diffusion analysis
of China-market revenue Nvidia (H20) and AMD (MI308) remit to the US government for export licenses — unprecedented15%20253.12US Commerce / Tom's Hardware / Fortune reporting
Nvidia H20 inventory/charge taken when the chip was license-gated mid-2025$4.5BQ1 FY20263.12Nvidia Q1 FY26 8-K (SEC)
advanced Nvidia accelerators authorized for the Gulf (G42 + HUMAIN, ~35k each) under the RTE compact~70,000 chipsNov 20253.12US Commerce statement; CNBC; DCD
target capacity of Stargate UAE (G42, OpenAI, Oracle, Nvidia, Cisco, SoftBank)1 GW2026-06-29: 1 → 2026-07-10: 120263.12Middle East AI News
of non-US data-center capacity (by investment value) operated by US-headquartered firms — the operator-nationality gap~48%20253.12arXiv 2508.00932, How Sovereign Is Sovereign Compute?
EU Data Act applies — obliges providers to block unlawful non-EU government access to EU-stored data and challenge conflicting requestsSept 202520253.12EU Data Act / Kiteworks GDPR–CLOUD analysis
BIS moves named advanced-compute exports to China/Macau from presumption-of-denial to case-by-case reviewJan 15 202620263.12US BIS final rule; Baker McKenzie sanctions analysis
Northern Virginia market size 2026 (operational + pipeline); ~96% of 2026 scheduled supply already committed, vacancy near zero~20 GW20263.13CBRE / Mordor Intelligence; Loudoun County
Texas/ERCOT data center demand by 2028 (~1/3 of projected US demand); large-load queue grew ~300% YoY~40 GW2026-06-29: 40 → 2026-07-10: 4020263.13ERCOT filings
share of Ireland's national electricity used by data centers in 2024 (up from ~5% in 2015); ~80% renewable + self-supply mandate to reconnect~22%20253.13CRU / EirGrid; KPMG Ireland
Nordic data center capacity 2025 (to ~1.98 GW by 2031); PUE as low as ~1.09; up to ~8,000 free-air-cooling hours/yr~1.32 GW20253.13Mordor Intelligence; Data Center Knowledge
planned capacity of Stargate UAE (Abu Dhabi); 1 GW first cluster, ~200 MW first phase targeted 2026 — gated by US chip-export licenses~5 GW20263.13G42 / DCD; Stargate UAE announcements
Saudi HUMAIN sovereign-AI buildout target by end of decade; chip access via G2G security assurances~1.9 GW2026-06-29: 1.9 → 2026-07-10: 1.920263.13HUMAIN/MIS contract reporting
Singapore wholesale colocation pricing — highest globally; ≥300 MW unlocked via Green DC Roadmap / DC-CFA2 (apps close Mar 2026)~$403/kW-mo2026-06-29: 450 → 2026-07-02: 40320263.13CBRE Global Data Center Trends 2026; IMDA/EDB
revenue per GW of AI capacity per year — the arithmetic that makes speed-to-power the dominant scoring weight across every cluster (contested — single-source) · contested~$10–12B20253.13SemiAnalysis (onsite gas economics)
more power through the same conductor (or less copper for equal power) moving from 415 VAC to 800 VDC distribution~85% / ~45%2026-06-29: 85 → 2026-07-07: 8520254.1NVIDIA 800 VDC Architecture blog (May 2025)
facility conversion stages collapsed by the 800 VDC chain; end-to-end efficiency ~82% → ~87%4 → 220264.1NVIDIA / SemiAnalysis 800VDC Revolution
utility-to-VRM efficiency: legacy AC span (typical ~82%) vs the SST/800 VDC chain~61–87.5% vs ~87%20254.1SemiAnalysis, Datacenter Anatomy Pt 1 / Inside the 800VDC Revolution
solid-state transformer efficiency at 400 kW (13.2 kVAC → 800 VDC); ~99% targeted~98%20254.1SemiAnalysis / ETH Zurich INTELEC 2025
per-rack power roadmap: H100 ~40 kW (2023), GB200 NVL72 ~120–132 kW (2024), Kyber ~600 kW (2027)~40 → 600 kW20264.1SemiAnalysis / NVIDIA roadmap
HV/substation power transformer lead time (up to ~60 months in constrained markets) — the schedule-dominating long pole~128–144 wk2025-Q24.1Wood Mackenzie / pv magazine
AI load lost in the six-fault, 82-second July 2024 NoVA sequence — why ride-through and topology resilience are now first-order design inputs~1,500 MW20244.1NERC Incident Review (large load loss); NERC Level 3 Alert (May 2026)
EU F-gas ban on putting new F-gas-insulated switchgear into operation: <=24 kV (24-52 kV from 2030; GWP>=1 HV bands from 2028/2032; Art. 13 derogations)1 Jan 202620264.2EU F-Gas Regulation (EU) 2024/573; European Commission
SF6 global-warming potential vs CO2 — why the F-gas phase-out is forced~24,000x20264.2IPCC / EU F-Gas Regulation
fault current from inverter-based sources (vs 5-10x for synchronous machines) — breaks conventional overcurrent grading~1.1-1.2x20254.2NERC / IEEE PES-PSRC; NREL IBR protection studies
typical HV/MV transformer unit, N+1, to serve ~150 MW peak; ~10% MVA cushion over MW~80 MVA x32026-06-29: 80 → 2026-07-10: 8020264.2SemiAnalysis Datacenter Anatomy Pt 1
ERCOT 'large load' threshold forcing full interconnection + protection study (25 MW = modeling)75 MW20264.2ERCOT Large Load Integration; SB6
MV switchgear lead time; ~128 wk standard HV transformer, ~144 wk GSU-class100+ wk20254.2Wood Mackenzie; pv magazine; supply-chain synthesis
instantaneous data-center load lost in the six-fault, 82-second 230 kV sequence (July 10 2024 NoVA event) — informed NERC's May 2026 Level 3 alert; ride-through recommended, not yet an enforceable standard~1,500 MW20244.2NERC Incident Review (large load loss); NERC Level 3 Alert (May 2026)
GIS footprint vs equivalent AIS lineup — the compactness premium for constrained/hostile sites~1/320264.2Siemens Energy / ABB MV switchgear synthesis
proposed CLE bright-line: ≥1 MW IT load within ≥20 MW aggregate at a single POI ≥60 kV≥1 MW + ≥20 MW20264.3NERC draft registry criteria (Appendix 5B); Steptoe analysis
data-center load tripped in the six-fault, 82-second July 2024 230 kV sequence — the synchronized-loss problem~1,500 MW20244.3NERC Incident Review (large load loss); NERC Level 3 Alert (May 2026)
load dropped in a single Virginia disturbance window; basis for NERC's ride-through recommendations (load standards drafting under Project 2026-02)1.5 GW / 82 s20244.3NERC / Utility Dive
NERC's highest (Essential Actions) alert tier — first used for load, on computational loadsLevel 320264.3NERC / Davis Wright Tremaine
CIP-015 (INSM) effective date; phased compliance Sep 2, 2028 / 2030Sep 2, 202520254.3FERC Order No. 907; Federal Register
NERC standards-development project defining the CLE functional class and obligationsProject 2026-0220264.3NERC / Climate Solutions Legal Digest
power-transformer / GSU lead time — the schedule lever the ownership fork controls~128–144 wk2025-Q24.3Wood Mackenzie / pv magazine
large-load grid interconnection lead time end-to-end; up to ~10 yr in worst queues~3–7+ yr20264.3LBNL Queued Up; ERCOT / PJM filings
current distortion of a modern AFE/active-PFC AI server PSU by design (vs 30%+ for 6-pulse)<5% THD2026-06-29: 5 → 2026-07-07: 520254.4IEEE 519-2022; AFE rectifier specs
K-rating warranted for AI-hall distribution transformers serving legacy non-linear loadsK-13 to K-202026-06-29: 13 → 2026-07-10: 1320254.4SemiAnalysis; QTE; Eaton
eddy-current heating multiplier of the 5th / 7th / 13th harmonic vs fundamental (∝ order²)≈25x / 49x / 169x20254.4NRETEC; CalcPanel K-factor guides
IEEE 519 current-distortion (TDD) limit at the PCC, sliding with the Isc/IL short-circuit ratio; voltage THD capped at 5% (1–69 kV) / 8% (<1 kV)5–20% TDD20224.4IEEE 519-2022
solid-state transformer efficiency, 13.2 kVAC → 800 VDC (ETH Zurich INTELEC benchmark); ~99% targeted~98% @ 400 kW20254.4SemiAnalysis / ETH Zurich
end-to-end utility-to-VRM efficiency of the SST/800 VDC chain vs a typical ~82% AC path (best modern AC ~87.5%)~87%20254.4SemiAnalysis, Datacenter Anatomy Pt 1 / Inside the 800VDC Revolution
HV/substation power-transformer lead time (≈144 wk GSU; to ~60 mo in constrained markets)~128 weeks2025-Q24.4Wood Mackenzie / pv magazine
projected SST TAM by 2030 / disaggregated power-rack (sidecar) TAM at its 2028 peak~$32B / ~$11B20254.4SemiAnalysis, Inside the 800VDC Revolution Pt 1
energy storage integrated in GB300 NVL72 power shelves; ~half the PSU volume is capacitance65 J/GPU20254.5NVIDIA Developer Blog / LITEON
peak grid-demand reduction demonstrated while training Megatron-LLM with energy-enhanced power shelves30%20254.5NVIDIA Developer (GB300 steady power)
reported on-rack energy storage on Vera Rubin (~6× GB300), per BESS-for-AI guidance~400 J/GPU2026-06-29: 400 → 2026-07-07: 4002026 (roadmap)4.5NVIDIA (production-ready BESS for AI factories)
ORV3 BBU ramp to full power on 48.5 V busbar droop; busbar held ≥46 V; PSU ≥97.5% (30–100% load)<2 ms2026-06-29: 2 → 2026-07-07: 220254.5OCP Open Rack V3 / ORV3 BBU spec
EDPp (electrical design power, peak) as the real sizing basis vs nameplate TDP~1.5× TDP20264.5Uptime Institute Journal; OCP/Diablo 400
double-conversion loss vs eco-mode/advanced-standby efficiency3–5% vs >99%2026-06-29: 4 → 2026-07-07: 420264.5Vertiv; ScienceDirect (battery/UPS systems)
battery-capacity cut and double-conversion-loss cut from distributed BBU + supercaps vs central 2N~50% / 2–3%20264.5SemiAnalysis; Eaton 800 VDC reference architecture
instantaneous data-center load lost in the six-fault, 82-second July 2024 NoVA sequence — the macro symptom of the transient problem~1,500 MW20244.5NERC Incident Review (large load loss); NERC Level 3 Alert (May 2026)
GB200 NVL72 rack draw (≈132 kW TDP class, ~120 kW continuous); rack busbar ~2,800 A at 48 V~132 kW2026-06-29: 132 → 2026-07-07: 13220254.6115/17 split: HPE QuickSpecs (OEM); ~132 kW all-in: Vertiv/451
Rubin Ultra Kyber-class rack on 800 VDC; shelves disaggregate to a sidecar~600 kWH2 2027 (announced)4.6NVIDIA GTC (Jensen Huang); DCD, Tom's Hardware
OCP ORV3 HPR power shelf output (6×~5.5 kW rectifiers; ~27.5 kW at N+1) onto the 48V busbar33 kW20244.6Advanced Energy/Artesyn ORV3 HPR datasheet; Delta 33 kW ORV3 (OCP)
overhead busway ampacity AI halls now spec (vs legacy 400A); 415/480V is majority of new deploys800–1600A20264.6Eaton PowerWave; busway market synthesis
rack power per ampere from 208V→415V three-phase: halves current, quarters I²R loss~2×20254.6Server Technology 415V PDU guidance
SELV touch-safety ceiling that fixes the OCP rack bus at 48V (no DC-arc service procedures)<60 V20254.6OCP ORV3 base spec; IEC 60364
in-rack copper NVLink cables in an NVL72 — the copper-mass reality that drives the busbar-cooling fork5,18420254.6SemiAnalysis GB200 architecture
SST efficiency at 400 kW (13.2 kVAC→800 VDC), the single-stage path feeding the in-rack rail~98%20254.6SemiAnalysis / ETH Zurich INTELEC 2025
end-to-end utility-to-VRM efficiency on the 800VDC/SST chain (vs ~61–87.5% legacy AC, typical ~82%)~87%20254.7SemiAnalysis, Datacenter Anatomy Pt 1 / Inside the 800VDC Revolution
SST efficiency, 13.2 kVAC → 800 VDC at 400 kW (prototype)~98%20254.7SemiAnalysis / ETH Zurich INTELEC 2025
more power through the same wire gauge, 3-wire DC (POS/RTN/PE) vs 4-conductor 415 VAC — NVIDIA's Oct 2025 framing; its May 2025 basis gave ~85%. Topology-specific vendor comparison, not a constant+157%2025-264.7NVIDIA, Building the 800 VDC Ecosystem (Oct 2025)
Mt. Diablo / Diablo 400 bipolar rail, chosen to ride the EV supply chain±400 VDC20254.7OCP Diablo 400 spec; Google (OCP EMEA)
Rubin Ultra Kyber-class rack the 800 VDC architecture targets~600 kWH2 2027 (announced)4.7NVIDIA GTC (Jensen Huang); DCD, Tom's Hardware
targeted UL listing for the first SST in NVIDIA's MGX reference (DG Matrix)~Q2 202620264.7SemiAnalysis, Inside the 800VDC Revolution
±400 VDC HVDC cables feeding an IT rack from the sidecar (Diablo later rev)16 × 50 kW2026-06-29: 16 → 2026-07-07: 1620264.7OCP Diablo 400 v0.7.0
first commercial 800 VDC power products (Vertiv, Schneider, Eaton, Delta), aligned to KyberH2 202620264.7DataCenterDynamics; vendor releases
ABB orders to VoltaGrid (27 booked through 2025 + 35 in Mar-2026 extension), with flywheel + eHouse, for islanded data-center stability62 sync condensers20264.8ABB / DataCenterDynamics
Tesla Megapacks deployed at xAI Colossus from Nov-2025 for power smoothing, ride-through, voltage/frequency support of a ~2 GW cluster>$375M2025-264.8DataCenterDynamics / SemiAnalysis
VoltaGrid generation for Vantage's ~1.4 GW IT load (~64% overbuild) — the PUE x redundancy x derate stack made concrete~2.3 GW2026-06-29: 2.3 → 2026-07-10: 2.320264.8Power Engineering
demonstrated grid-forming BESS black-start capability — proof an inverter-led island can stand up from dead200 MW / 275 kV2026-06-29: 200 → 2026-07-10: 20020254.8grid-forming BESS demonstrations (industry literature)
on-board capacitance shaping the sub-ms edge: ~65 J/GPU (GB300) to ~400 J/GPU (Vera Rubin) before the bus ever sees the pulse~65–400 J/GPU20264.8NVIDIA / SemiAnalysis
real-power injection window of BESS/flywheels — the bridge from the ms GPU step to the genset's seconds-to-minutes ramp5–30 s2026-06-29: 17.5 → 2026-07-10: 17.520254.8BESS/flywheel vendor data
reciprocating-engine ramp and time-to-full-load — fast for combustion, still geological vs a ms GPU step>100%/min · ~2 min20254.8Grid Capacity Intelligence / Data Center Frontier
instantaneous data-center load lost in the six-fault, 82-second July 2024 NoVA sequence — the event that made ride-through and islanding first-order planning inputs~1,500 MW20244.8NERC Incident Review (large load loss); NERC Level 3 Alert (May 2026)
behind-the-meter gas announced cumulatively by mid-2026 (~60 GW with equipment orders, ~two-thirds of projects; only ~1 GW under construction)~90 GW2026-074.9Cleanview, 'Bypassing the Grid' (mid-2026)
AI-driven incremental US gas demand projected by 2030; US gas demand to ~150 Bcf/d by 2031~6 Bcf/d2030-20314.9SemiAnalysis / RAND; INGAA
additional US pipeline capacity INGAA projects is needed by 2052 to serve load growth+39%20254.9Natural Gas Intelligence / INGAA
Modified Wobbe Index band for gas-turbine fuel; allowable swing ~±5% outer, ±1–3% practical40–55 MJ/m³20254.9Gas-turbine fuel-gas supply guidance
energy per LNG truck vs CNG truck; LNG extends economics to ~300 mi, CNG best within ~60 mi826 vs 450–550 Dth20254.9Natural Gas Intelligence (LNG/CNG for five-nines)
fuel-cost saving from interruptible-service rates vs firm transport (the synthetic-firm incentive)30–40%+20254.9Natural Gas Intelligence; standby-fuel-system analyses
fuel burn, simple-cycle turbine vs combined-cycle/fuel-cell — the conditioning/efficiency stakes~8–9 vs ~6 MMBtu/MWh2026-06-29: 8.5 → 2026-07-10: 8.520254.9Grid Capacity Intelligence
reciprocating-engine lead time (needs no boost compression) vs 18–36 mo+ for aeroderivative turbines12–18 mo20254.9Data Center Frontier / Grid Capacity Intelligence
data-center load lost simultaneously during the six-fault, normally-cleared 230 kV sequence (Eastern Interconnection, July 2024)~1,500 MW20244.10NERC Incident Review (large load loss); NERC Level 3 Alert (May 2026)
successive system faults from the lightning-arrestor failure; voltage sagged to 0.25-0.40 pu for 42-66 ms per fault6 faults / 82 s20254.10NERC Incident Review (Jan 2025)
post-event over-voltage as ~1.5 GW of load dropped and the grid over-swung high1.07 pu2026-06-29: 1.07 → 2026-07-01: 1.0720254.10NERC Incident Review (Jan 2025)
NERC Level 3 "Essential Actions" Alert issued; 7 mandated actions, responses due Aug 3, 2026May 4, 202620264.10NERC / Morgan Lewis; Utility Dive
ERCOT Large Load site threshold; "Large Electronic Load" = ≥50% power-electronic computational demand (NOGRR 282 / NPRR 1308)≥75 MW / ≥50%2026-06-29: 75 → 2026-07-02: 7520254.10ERCOT NOGRR 282 / NPRR 1308 (filed Nov 2025)
SPP demand-facility ride-through floor (0.90-1.10 pu continuous; short excursions to ~0.5 pu)~0.50 pu, ~0.15 s20254.10SPP demand-facility VRT curve
typical power-factor obligation at the POI for large transmission-connected loadsunity-0.95 lag20264.10ISO/utility large-load tariffs (synthesis)
large-load grid interconnection lead time, ISO-dependent; ride-through now a study deliverable~3-7+ yr20254.10ERCOT / PJM filings synthesis
millisecond idle→peak step of GB300-class racks; phase-coherent, common-mode-injecting load the reference plane must absorb>150 kW/rack2026-06-29: 150 → 2026-07-07: 15020254.11SemiAnalysis Datacenter Anatomy; NVIDIA
DC bus voltages replacing the grounded 415/480 VAC neutral at high rack power (Mt Diablo / Diablo 400 is bipolar ±400 V about a grounded midpoint; NVIDIA / OCP 800 VDC designs run ungrounded/HRG or midpoint-referenced — see Ch 4.7)±400 / 800 VDC20264.11SemiAnalysis 800VDC Revolution; OCP
NVIDIA monopolar 800 VDC power-sidecar reference design (air-cooled samples/production mid-2026; liquid-cooled VR Ultra variant late-2026) — the power rack, not the ~600 kW (2H 2027) Kyber compute rack~660 kW20264.11SemiAnalysis, Inside the 800VDC Revolution Pt 1 (May 2026)
Type 1 / Class I SPD test waveform (direct-lightning current); Type 2 / Class II uses 8/20 µs10/350 µs20244.11IEC 61643-11 / IEC 61643-01:2024
current North American SPD listing basis, ANSI-approved 2025; NEC mandates UL 1449 listingUL 1449 5th Ed.20254.11UL Standards; NEMA Surge
international lightning-protection standard (risk, LPS, SPM, services); NFPA 780 is the NA counterpartIEC 6230520244.11IEC; NFPA
substation grounding standard bounding step/touch voltage against body-current limits and soil resistivityIEEE 802013 (current ed.)4.11IEEE Std 80
required grounding-conductor resistance vs ungrounded conductor at 400 / 800 VDC to meet protective grounding class — sizing tightens sharply with voltage60% / 23%20254.11Siemens DC-data-center protections paper
capacitive energy storage in GB300 NVL72 power shelf for ramp smoothing; ~half the PSU volume65 J/GPU20254.12NVIDIA Developer Blog / LITEON
peak grid-demand reduction from GB300 power-shelf storage + burn + ramp controlup to 30%20254.12NVIDIA Developer Blog; ServeTheHome
targeted on-package + facility capacitance for Vera Rubin closed-loop smoothing~400 J/GPU2026-06-29: 400 → 2026-07-07: 4002026 (roadmap)4.12NVIDIA (production-ready BESS for AI factories)
IEEE 519-2022 voltage-distortion limit at the PCC for 1-69 kV; 3% per individual harmonic; 8% below 1 kV5% THD-V20224.12IEEE 519-2022 (via TCI/Elspec)
IEEE 519 current-distortion limit, sliding by Isc/IL short-circuit ratio5-20% TDD20224.12IEEE 519-2022 (via NFM Consulting)
power-budget headroom oversubscription, training vs inference fleets~3% vs ~21%20244.12Patel et al., POLCA (Microsoft Research), ASPLOS 2024 (arXiv 2308.12908)
electrical commissioning level taxonomy (factory test -> integrated systems test)L1-L520254.12Construct & Commission; BMP MEP
utility revenue-meter accuracy class at the point of interconnection0.2S / 0.5S20244.12IEC 62053-22 metering practice
practical air-cooling ceiling per rack; RDHx bridges ~50–100 kW; DLC clears 200+ kW~41 kW20255.1ASHRAE TC 9.9; SemiAnalysis Datacenter Anatomy
volumetric heat capacity of water vs air — the reason liquid operates in a different cooling regime~3,500&times;20255.1Thermodynamic reference; ASHRAE TC 9.9
per GB200 NVL72 rack (~115 kW removed by liquid, ~17 kW by air); ~3&times; over the air wall~120–132 kW2026-06-29: 126 → 2026-07-02: 12620255.1115/17 split: HPE QuickSpecs (OEM); 120 kW cooling: NVIDIA; ~132 kW all-in: Vertiv/451
GB200 NVL72 coolant acceptance envelope (ASHRAE W45-class; ~25 &deg;C was NVIDIA's launch operating example, not a floor)up to ~45 &deg;C supply / ~65 &deg;C return; ~170–235 L/min20255.1NVIDIA GB200 NVL72 launch / Vertiv–451 Research reference architecture
per GB300 NVL72 rack (up to ~155 kW peak); CPUs/GPUs/NVSwitch liquid, optics/storage air~135–142 kW20255.1Schneider Electric / HPE / Lenovo datasheets
per Rubin Ultra Kyber rack (NVL144) on 800 VDC (roadmap, 2H2027)~600 kWH2 2027 (announced)5.1NVIDIA GTC (Jensen Huang); DCD, Tom's Hardware
projected Rubin Ultra per-GPU TDP (A100 was ~0.3 kW) — a ~7&times; climb in seven years~2.3 kW2026 (pre-ship)5.1NVIDIA / SemiAnalysis roadmap
cost to retrofit an air-cooled hall across the cliff to AI liquid cooling; still strands capacity$5–10M/MW2026-06-29: 7.5 → 2026-07-10: 7.520265.1SemiAnalysis
practical air-cooling ceiling per rack (conventional contained hall); ~50 kW with close-coupled in-row/fan-wall~41 kW20255.2ASHRAE TC 9.9; SemiAnalysis Datacenter Anatomy
ASHRAE recommended inlet band; allowable A1 15–32, A2 10–35, A3 5–40, A4 5–45 °C18–27 °C20245.2ASHRAE TC 9.9 Thermal Guidelines (5th ed.)
active in-row cooler capacity per rack (contained); fan-wall close-coupled similar~30–50 kW20255.2Trane / Vertiv product data; Uptime Institute
GB200 NVL72 rack draw — ~90 kW past the air ceiling; ~115 kW liquid + ~17 kW air~132 kW2026-06-29: 132 → 2026-07-07: 13220255.2115/17 split: HPE QuickSpecs (OEM); ~132 kW all-in: Vertiv/451
legacy air-cooled PUE band; good containment + warm supply pushes toward low-1.2s1.4–1.620255.2SemiAnalysis / Uptime Institute
common 2026 dense-hall split: liquid for GPU racks, air for networking/storage/CPU residual~70 / 3020255.2Tom's Hardware cooling state of play 2025; SemiAnalysis
wet weight of a loaded NVL72-class rack — exceeds common raised-floor point-load ratings~3,000 lb2026-06-29: 3000 → 2026-07-10: 300020255.2NVIDIA OCP
share of pre-2015 data centers deemed unsuitable for dense-AI liquid retrofit (air-era stock)68%2026-06-29: 68 → 2026-07-10: 6820255.2DCD retrofit analysis
practical air-cooling ceiling per rack — the floor of the bridge band; RDHx extends to ~50–100 kW, DLC to 200+ kW~41 kW20255.3ASHRAE TC 9.9; SemiAnalysis Datacenter Anatomy

Refine your search or pick a Part to narrow the 1,434 matches.