Numbers Provenance Register
Every volatile headline figure — date-stamped, sourced, and flagged where contested. Plus 1,357 inline figures across the chapters.
Market & capex
| Metric | Value | As of | Source |
|---|---|---|---|
| Global data center capex by 2030 | ~$6.7T total (~$5.2T AI-capable + ~$1.5T traditional) | 2025 | McKinsey, 'The cost of compute' |
| Global data center capacity demand by 2030 | ~219 GW total, ~70% from AI | 2025 | McKinsey, 'The cost of compute' |
| Top-4 US hyperscaler data center capex 2026 | ~$600B (about +36% YoY) | 2026 | CreditSights / Dell'Oro Group |
| Global data center capex 2026 | approaching ~$1T | 2026 | Dell'Oro Group |
| Data center build financing gap 2025-2028 | ~$2.9T total spend, ~$1.5T financing gap | 2025 | Morgan Stanley Research |
| Global data center electricity demand 2025 to 2030 · contested | ~485 TWh (2025) to ~950 TWh (2030) | 2026 | IEA, Electricity 2026 / Energy and AI |
| Data centers as share of US electricity by 2030 | 9% to 17% (up from ~4-5% today) | 2026 | EPRI, Powering Intelligence 2026 |
| Global data center power demand growth to 2030 | +~165% vs 2023 (+~50% by 2027) | 2025 | Goldman Sachs Research |
Power & energy
| Metric | Value | As of | Source |
|---|---|---|---|
| Rack power by GPU generation | H100 ~40kW (2023), GB200 NVL72 ~120-132kW (2024), GB300 ~142kW (2025), Rubin VR200 ~190-230kW (2026), Rubin Ultra Kyber ~600kW (2027) | 2026 | SemiAnalysis / NVIDIA roadmap |
| GPU TDP trajectory | A100 300W, H100 700W, GB200 ~1.0-1.2kW, Rubin ~1.8kW, Rubin Ultra ~2.3kW | 2026 | NVIDIA / SemiAnalysis |
| Solid-state transformer (SST) efficiency | ~98% at 400kW (13.2kVAC to 800VDC); ~99% targeted | 2025 | SemiAnalysis / ETH Zurich |
| End-to-end electrical chain efficiency | legacy AC ~61-87.5%; 800VDC/DC chain over 92% | 2025 | SemiAnalysis, Datacenter Anatomy Pt 1 |
| HV/substation power transformer lead time | ~128 weeks standard; ~144 weeks GSU; up to ~60 months in constrained markets | 2025-Q2 | Wood Mackenzie / pv magazine |
| Grid interconnection lead time (large load) | about 3-7+ years end-to-end; up to ~10 yr in worst queues | 2025 | ERCOT / PJM filings synthesis |
| Behind-the-meter (BTM) gas announced | ~82-101 GW cumulatively by 2026 (~50 GW announced in 2025) | 2026 | Cleanview / SemiAnalysis |
| Aeroderivative gas turbine lead time | 18-36 mo+ (refurb under 12 mo); some cores quoted to ~243 weeks | 2025 | Data 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) | 2026 | NERC Level 3 Alert / Utility Dive |
Cooling & water
| Metric | Value | As of | Source |
|---|---|---|---|
| WUE (water usage effectiveness) | industry avg ~1.8-1.9 L/kWh; best-in-class 0.3-0.7 L/kWh; closed-loop ~0 | 2025 | Vertiv / NREL synthesis |
| Microsoft FY2025 fleet WUE | ~0.30 L/kWh (next-gen closed-loop designs ~0) | 2025 | Microsoft |
| PUE by cooling type | legacy air 1.4-1.6; D2C liquid 1.05-1.15; two-phase immersion 1.01-1.10 | 2025 | SemiAnalysis / Uptime Institute |
| DLC flow rate rule of thumb | ~1.2-2.0 L/min per kW; GB200 ~80 L/min/rack | 2025 | Dober / NVIDIA OCP |
| GB200 NVL72 DLC operating spec | inlet 20-25C, ~80 L/min flow, ~200 L coolant, ~2.4 MW cooling capacity | 2025 | NVIDIA OCP / Introl |
| Single-phase D2C share of cooling market 2026 | ~55% | 2026 | DCD / IDTechEx |
| DTC vs immersion cooling capex | DTC ~$300-500/kW vs immersion ~$1,000+/kW | 2026 | Gottog / Introl synthesis |
Compute & memory
| Metric | Value | As of | Source |
|---|---|---|---|
| GB200 NVL72 rack performance | ~1.44 ExaFLOPS FP4 (sparse), 13.4 TB unified memory, 130 TB/s NVLink | 2025 | NVIDIA |
| GB200 NVL72 physical | ~120 kW draw, ~1.36 t (3,000 lb), 18 compute + 9 NVSwitch trays | 2025 | NVIDIA OCP / Introl |
| NVIDIA per-GPU HBM capacity trajectory | H100 80GB, H200 141GB, B200 192GB, B300 288GB, Rubin Ultra 1TB (HBM4e) | 2026 | NVIDIA Developer |
| NVLink per-GPU bandwidth | Gen4 900 GB/s, Gen5 1.8 TB/s (Blackwell), Gen6 3.6 TB/s (Rubin) | 2026 | NVIDIA |
| HBM per-stack pricing · contested | HBM3 ~$200, HBM3E ~$300, HBM4 ~$500 (est.) | 2026 | Momoview / TrendForce synthesis |
| HBM supply allocation | 2026 HBM3E fully sold out; ~30% supply gap; +15-20%/quarter price rises | 2026 | SemiAnalysis / TrendForce |
| GPU useful life (book vs economic) · contested | 5-6 yr accounting vs 2-3 yr frontier-economic | 2026 | CNBC / Stanley Laman / SemiAnalysis |
Networking & optics
| Metric | Value | As of | Source |
|---|---|---|---|
| Per-lane SerDes / optics speed ladder | 50G-100G-200G-400G per lane; 800G = 8x100G or 4x200G; 1.6T = 8x200G | 2025 | SemiAnalysis (AI networks) |
| 800G optics power | DSP pluggable ~14-17W; LPO ~7-8.5W; CPO under 6W/800G | 2025 | SemiAnalysis / Broadcom |
| Oversubscription ratio by workload | training 1:1 non-blocking; inference 2:1-3:1; Meta ran 7:1 on 24k H100 | 2025 | SemiAnalysis / Meta |
| NVLink scale-up vs scale-out bandwidth | NVLink5 1.8 TB/s/GPU (NVL72 = 130 TB/s rack) vs ~400G NIC (about 18x) | 2025 | NVIDIA / SemiAnalysis |
| Copper reach at 800G/1.6T | passive DAC ~1-2m; active copper (AEC) ~3-7m; optics beyond | 2025 | SemiAnalysis (GB200 architecture) |
| InfiniBand vs RoCEv2 latency & effective BW | IB ~1-2 us; tuned RoCEv2 ~1.5-2.5 us; Spectrum-X ~95% effective throughput | 2025 | SemiAnalysis / NVIDIA |
Economics
| Metric | Value | As of | Source |
|---|---|---|---|
| TCO of a 1 GW AI data center | ~$38B upfront capex, ~$8.5B/yr all-in (~$0.9B/yr opex) | 2026 | Epoch AI |
| 1 GW TCO sensitivity to IT useful life · contested | 3-yr life ~$12B/yr; 5-yr ~$8.5B/yr; 7-yr ~$7B/yr | 2025 | Epoch AI / AM Compute synthesis |
| TCO $/GPU-hr (self-operated) | ~$0.74 at 2048-GPU scale @ 90% util; ~$1.03 small clusters | 2025 | SemiAnalysis |
| H100 rental rates 2026 | ~$1.03/hr spot floor; neocloud median ~$2.29-3.50/hr; AWS ~$6.88/hr; Azure ~$12.29/hr | 2026 | SemiAnalysis H100 Index / AM Compute |
| Inference cost per million tokens (self-hosted) | ~$1.90/M tokens (8x H100 @ ~$19.20/hr, Llama-70B FP16) | 2025 | Introl / NVIDIA synthesis |
| Breakeven utilization (debt-financed cluster) | ~70% | 2025 | AM Compute / McKinsey |
| Hyperscaler depreciation policy changes · contested | Meta 4 to 5.5 yr (+$2.9B income); Amazon 6 to 5 yr (-$700M); MSFT/Google 4 to 6 yr | 2025 | HBS case / company filings |
| GPU residual value · contested | H100 retained ~60-83% at 18 months; ~20-40% residual after 3 yr | 2025 | Hashrate Index / CNBC synthesis |
| Wholesale colocation pricing 2025 | global avg ~$217/kW-month; Ashburn ~$215; range ~$120 (Atlanta) to ~$450 (Singapore) | 2025 | JLL / CBRE synthesis |
| Revenue per GW of AI capacity | ~$10-12B/GW/yr | 2025 | SemiAnalysis (onsite gas) |
Reliability
| Metric | Value | As of | Source |
|---|---|---|---|
| Llama 3 405B training interruptions | 419 unplanned interruptions over 54 days on 16,384 H100s (~1 every 3 hr) | 2024 | Meta (Llama 3 paper) / Tom's Hardware |
| Llama 3 failure breakdown | Faulty GPU 30.1%, HBM3 17.2%, SRAM 4.5%, processor 4.1%, network 8.4%, software 12.9% | 2024 | Meta (Llama 3 paper) |
| Top-tier H100 operator MTBF | ~7 days per 512 GPUs | 2025 | SemiAnalysis (100k H100 clusters) |
| Large LLM job failure rate | ~43.4% failure rate; ~37% hardware-attributed; ~73% recoverable via restart | 2024 | Alibaba (Unicron) |
| Goodput (effective training time) | industry avg ~90%; best-in-class ~96% | 2025 | SemiAnalysis ClusterMAX / CoreWeave |
| Uptime Tier availability | Tier III 99.982% (~1.6 hr/yr down); Tier IV 99.995% (~26 min/yr) | 2025 | Uptime Institute |
Regulatory watchlist
| Jurisdiction | Rule | Threshold | Status | Next |
|---|---|---|---|---|
| EU | EED Art. 26 (Dir. 2023/1791) — waste-heat recovery cost-benefit analysis required | >1 MW total rated input | In force | ongoing |
| EU | EED Art. 12 / Delegated Reg. (EU) 2024/1364 — mandatory DC energy/water/PUE reporting to EU database | ≥500 kW installed IT | In force (annual reporting) | annual cycle |
| Germany | EnEfG — Energy Reuse Factor (ERF) floor for new data centers; PUE cap for new builds | ERF ≥10% (Jul-2026) → ≥20% (Jul-2028); new-build PUE ≤1.2 | Enacted | Jul-2026 |
| France | Law 2025-391 — heat valorization required for new facilities | >1 MW | Enacted | phase-in |
| EU | CSRD 'Omnibus' simplification — sustainability-reporting scope & timelines being amended | large-undertaking thresholds under revision | PENDING — verify final scope/dates | TBD |
| US (Federal) | SEC climate-disclosure rule — 2024 rule subject to 2026 rescission proposal | registrants (phased) | PROPOSED RESCISSION — enforcement uncertain | TBD |
| US (California) | SB 253 (GHG disclosure) + SB 261 (climate-financial-risk disclosure) | rev. >$1B (253) / >$500M (261) | Enacted, phasing in | first reports 2026 |
| Global | GHG Protocol Scope 2 revision — market-based vs 24/7 / locational accounting under review | affects clean-power / CFE claims | UNDER REVISION (draft) | TBD |
| US (Texas) | ERCOT large-load interconnection — process & curtailable/large-load terms evolving (SB-driven) | large loads (e.g. >75 MW) | Evolving | ongoing rulemaking |
| US / multi | BIS advanced-compute export controls — periodic tightening affecting accelerator availability & data residency | advanced GPUs/HBM | Evolving | periodic updates |