The Definitive Guide toAI Data Centers
Ask the Guide

Numbers Provenance Register

Every volatile headline figure — date-stamped, sourced, and flagged where contested. Plus 1,357 inline figures across the chapters.

Market & capex

MetricValueAs ofSource
Global data center capex by 2030~$6.7T total (~$5.2T AI-capable + ~$1.5T traditional)2025McKinsey, 'The cost of compute'
Global data center capacity demand by 2030~219 GW total, ~70% from AI2025McKinsey, 'The cost of compute'
Top-4 US hyperscaler data center capex 2026~$600B (about +36% YoY)2026CreditSights / Dell'Oro Group
Global data center capex 2026approaching ~$1T2026Dell'Oro Group
Data center build financing gap 2025-2028~$2.9T total spend, ~$1.5T financing gap2025Morgan Stanley Research
Global data center electricity demand 2025 to 2030 · contested~485 TWh (2025) to ~950 TWh (2030)2026IEA, Electricity 2026 / Energy and AI
Data centers as share of US electricity by 20309% to 17% (up from ~4-5% today)2026EPRI, Powering Intelligence 2026
Global data center power demand growth to 2030+~165% vs 2023 (+~50% by 2027)2025Goldman Sachs Research

Power & energy

MetricValueAs ofSource
Rack power by GPU generationH100 ~40kW (2023), GB200 NVL72 ~120-132kW (2024), GB300 ~142kW (2025), Rubin VR200 ~190-230kW (2026), Rubin Ultra Kyber ~600kW (2027)2026SemiAnalysis / NVIDIA roadmap
GPU TDP trajectoryA100 300W, H100 700W, GB200 ~1.0-1.2kW, Rubin ~1.8kW, Rubin Ultra ~2.3kW2026NVIDIA / SemiAnalysis
Solid-state transformer (SST) efficiency~98% at 400kW (13.2kVAC to 800VDC); ~99% targeted2025SemiAnalysis / ETH Zurich
End-to-end electrical chain efficiencylegacy AC ~61-87.5%; 800VDC/DC chain over 92%2025SemiAnalysis, Datacenter Anatomy Pt 1
HV/substation power transformer lead time~128 weeks standard; ~144 weeks GSU; up to ~60 months in constrained markets2025-Q2Wood Mackenzie / pv magazine
Grid interconnection lead time (large load)about 3-7+ years end-to-end; up to ~10 yr in worst queues2025ERCOT / PJM filings synthesis
Behind-the-meter (BTM) gas announced~82-101 GW cumulatively by 2026 (~50 GW announced in 2025)2026Cleanview / SemiAnalysis
Aeroderivative gas turbine lead time18-36 mo+ (refurb under 12 mo); some cores quoted to ~243 weeks2025Data Center Frontier / Grid Capacity Intelligence
NERC large-load instantaneous loss events~1,500 MW lost on a single 230kV fault; 1.5 GW dropped in 82 s (VA, 2024)2026NERC Level 3 Alert / Utility Dive

Cooling & water

MetricValueAs ofSource
WUE (water usage effectiveness)industry avg ~1.8-1.9 L/kWh; best-in-class 0.3-0.7 L/kWh; closed-loop ~02025Vertiv / NREL synthesis
Microsoft FY2025 fleet WUE~0.30 L/kWh (next-gen closed-loop designs ~0)2025Microsoft
PUE by cooling typelegacy air 1.4-1.6; D2C liquid 1.05-1.15; two-phase immersion 1.01-1.102025SemiAnalysis / Uptime Institute
DLC flow rate rule of thumb~1.2-2.0 L/min per kW; GB200 ~80 L/min/rack2025Dober / NVIDIA OCP
GB200 NVL72 DLC operating specinlet 20-25C, ~80 L/min flow, ~200 L coolant, ~2.4 MW cooling capacity2025NVIDIA OCP / Introl
Single-phase D2C share of cooling market 2026~55%2026DCD / IDTechEx
DTC vs immersion cooling capexDTC ~$300-500/kW vs immersion ~$1,000+/kW2026Gottog / Introl synthesis

Compute & memory

MetricValueAs ofSource
GB200 NVL72 rack performance~1.44 ExaFLOPS FP4 (sparse), 13.4 TB unified memory, 130 TB/s NVLink2025NVIDIA
GB200 NVL72 physical~120 kW draw, ~1.36 t (3,000 lb), 18 compute + 9 NVSwitch trays2025NVIDIA OCP / Introl
NVIDIA per-GPU HBM capacity trajectoryH100 80GB, H200 141GB, B200 192GB, B300 288GB, Rubin Ultra 1TB (HBM4e)2026NVIDIA Developer
NVLink per-GPU bandwidthGen4 900 GB/s, Gen5 1.8 TB/s (Blackwell), Gen6 3.6 TB/s (Rubin)2026NVIDIA
HBM per-stack pricing · contestedHBM3 ~$200, HBM3E ~$300, HBM4 ~$500 (est.)2026Momoview / TrendForce synthesis
HBM supply allocation2026 HBM3E fully sold out; ~30% supply gap; +15-20%/quarter price rises2026SemiAnalysis / TrendForce
GPU useful life (book vs economic) · contested5-6 yr accounting vs 2-3 yr frontier-economic2026CNBC / Stanley Laman / SemiAnalysis

Networking & optics

MetricValueAs ofSource
Per-lane SerDes / optics speed ladder50G-100G-200G-400G per lane; 800G = 8x100G or 4x200G; 1.6T = 8x200G2025SemiAnalysis (AI networks)
800G optics powerDSP pluggable ~14-17W; LPO ~7-8.5W; CPO under 6W/800G2025SemiAnalysis / Broadcom
Oversubscription ratio by workloadtraining 1:1 non-blocking; inference 2:1-3:1; Meta ran 7:1 on 24k H1002025SemiAnalysis / Meta
NVLink scale-up vs scale-out bandwidthNVLink5 1.8 TB/s/GPU (NVL72 = 130 TB/s rack) vs ~400G NIC (about 18x)2025NVIDIA / SemiAnalysis
Copper reach at 800G/1.6Tpassive DAC ~1-2m; active copper (AEC) ~3-7m; optics beyond2025SemiAnalysis (GB200 architecture)
InfiniBand vs RoCEv2 latency & effective BWIB ~1-2 us; tuned RoCEv2 ~1.5-2.5 us; Spectrum-X ~95% effective throughput2025SemiAnalysis / NVIDIA

Economics

MetricValueAs ofSource
TCO of a 1 GW AI data center~$38B upfront capex, ~$8.5B/yr all-in (~$0.9B/yr opex)2026Epoch AI
1 GW TCO sensitivity to IT useful life · contested3-yr life ~$12B/yr; 5-yr ~$8.5B/yr; 7-yr ~$7B/yr2025Epoch AI / AM Compute synthesis
TCO $/GPU-hr (self-operated)~$0.74 at 2048-GPU scale @ 90% util; ~$1.03 small clusters2025SemiAnalysis
H100 rental rates 2026~$1.03/hr spot floor; neocloud median ~$2.29-3.50/hr; AWS ~$6.88/hr; Azure ~$12.29/hr2026SemiAnalysis H100 Index / AM Compute
Inference cost per million tokens (self-hosted)~$1.90/M tokens (8x H100 @ ~$19.20/hr, Llama-70B FP16)2025Introl / NVIDIA synthesis
Breakeven utilization (debt-financed cluster)~70%2025AM Compute / McKinsey
Hyperscaler depreciation policy changes · contestedMeta 4 to 5.5 yr (+$2.9B income); Amazon 6 to 5 yr (-$700M); MSFT/Google 4 to 6 yr2025HBS case / company filings
GPU residual value · contestedH100 retained ~60-83% at 18 months; ~20-40% residual after 3 yr2025Hashrate Index / CNBC synthesis
Wholesale colocation pricing 2025global avg ~$217/kW-month; Ashburn ~$215; range ~$120 (Atlanta) to ~$450 (Singapore)2025JLL / CBRE synthesis
Revenue per GW of AI capacity~$10-12B/GW/yr2025SemiAnalysis (onsite gas)

Reliability

MetricValueAs ofSource
Llama 3 405B training interruptions419 unplanned interruptions over 54 days on 16,384 H100s (~1 every 3 hr)2024Meta (Llama 3 paper) / Tom's Hardware
Llama 3 failure breakdownFaulty GPU 30.1%, HBM3 17.2%, SRAM 4.5%, processor 4.1%, network 8.4%, software 12.9%2024Meta (Llama 3 paper)
Top-tier H100 operator MTBF~7 days per 512 GPUs2025SemiAnalysis (100k H100 clusters)
Large LLM job failure rate~43.4% failure rate; ~37% hardware-attributed; ~73% recoverable via restart2024Alibaba (Unicron)
Goodput (effective training time)industry avg ~90%; best-in-class ~96%2025SemiAnalysis ClusterMAX / CoreWeave
Uptime Tier availabilityTier III 99.982% (~1.6 hr/yr down); Tier IV 99.995% (~26 min/yr)2025Uptime Institute

Regulatory watchlist

JurisdictionRuleThresholdStatusNext
EUEED Art. 26 (Dir. 2023/1791) — waste-heat recovery cost-benefit analysis required>1 MW total rated inputIn forceongoing
EUEED Art. 12 / Delegated Reg. (EU) 2024/1364 — mandatory DC energy/water/PUE reporting to EU database≥500 kW installed ITIn force (annual reporting)annual cycle
GermanyEnEfG — Energy Reuse Factor (ERF) floor for new data centers; PUE cap for new buildsERF ≥10% (Jul-2026) → ≥20% (Jul-2028); new-build PUE ≤1.2EnactedJul-2026
FranceLaw 2025-391 — heat valorization required for new facilities>1 MWEnactedphase-in
EUCSRD 'Omnibus' simplification — sustainability-reporting scope & timelines being amendedlarge-undertaking thresholds under revisionPENDING — verify final scope/datesTBD
US (Federal)SEC climate-disclosure rule — 2024 rule subject to 2026 rescission proposalregistrants (phased)PROPOSED RESCISSION — enforcement uncertainTBD
US (California)SB 253 (GHG disclosure) + SB 261 (climate-financial-risk disclosure)rev. >$1B (253) / >$500M (261)Enacted, phasing infirst reports 2026
GlobalGHG Protocol Scope 2 revision — market-based vs 24/7 / locational accounting under reviewaffects clean-power / CFE claimsUNDER REVISION (draft)TBD
US (Texas)ERCOT large-load interconnection — process & curtailable/large-load terms evolving (SB-driven)large loads (e.g. >75 MW)Evolvingongoing rulemaking
US / multiBIS advanced-compute export controls — periodic tightening affecting accelerator availability & data residencyadvanced GPUs/HBMEvolvingperiodic updates