Numbers Provenance Register
1,041 canonical active claims across 1,133 chapter occurrences. Each claim appears once by default; expand it to see every place the guide quotes it. The all-rows view lists those occurrences alongside claims not yet placed in a chapter and the figures on the watchlist.
1,041 active claims1,133 chapter occurrences725 with revision history22 flagged contested— review due
Recently revised
How to use this register
Before carrying a figure into a design or model, match its value kind and scope to your decision, check its as-of and review dates, and open the cited source. A contested figure has credible sources that materially disagree, so test the decision against the competing values. Sparklines and the recently revised list show recorded changes; review health follows each figure's review-by date and fast, standard, or slow cadence.
- derived · 103 Derived values are guide calculations; their result is only as fixed as the stated inputs, scope, and method.
- estimate · 102 Interpret this value according to its displayed kind, scope, source, and as-of date.
- forecast · 84 Forecasts are time-bound outlooks, not specifications; test decisions across plausible scenarios.
- modeled · 80 Interpret this value according to its displayed kind, scope, source, and as-of date.
- observed · 646 Reported measurements, counts, and specifications keep the precision and scope stated by their source; an exact specification is not treated as a range.
- proposed · 1 Proposed values are not yet final requirements or specifications; track the decision or rulemaking that can change them.
- recommendation · 25 Recommendations are design guidance, not measured facts; validate them against the project boundary conditions.
1,041 matches · showing 1–100
| Canonical claim | Value | Kind / scope | Trend | As of | Review | Updated | Occurrences | Source |
|---|---|---|---|---|---|---|---|---|
| 1.6T switches were only sampling in 2Q 2026; Dell'Oro forecasts a faster ramp than 800G, over 5M ports within 1-2 years of shipments | 1.6T switches: sampling 2Q 2026, volume ramp expected 2H 2026 | forecastDell'Oro forecast (2025-12-17): 2026 first volume year, ramp faster than 800G, >5M ports within 1-2 years of shipments. In 2Q26, 800G was the vast majority of AI back-end switch shipments and 1.6T was sampling. 1.6T transceivers (a separate population) were already shipping in volume in 2026 per LightCounting. | — | 2026-09 | checking… | — | Dell'Oro Group, 2026 predictions (2025-12-17) and 2Q26 AI back-end switch release (2026-09-03) | |
| 1 GW AI data center: up-front capex ~$37.9B (servers ~$21.2/W with GPUs inside that line, facility ~$11.4/W, network ~$4.9/W, land + utility works ~$0.3/W) and all-in annual TCO (~$8.5M/MW-yr) | ~$38B / ~$8.5B/yr | modeledEpoch’s modeled US hyperscaler, 1 GW IT, GB200 NVL72. Default IT life about 5.3 years; annualized economic cost includes capital opportunity cost, not a literal debt-service bill or exact five-year sensitivity. | — | 2026-05 | checking… | — | Epoch AI, Amelia Michael and Ben Cottier, May 2026 model and linked workbook. May 19 is the article’s up-front CapEx CSV revision date; annualized-cost CSV is marked May 14. This record’s asof_date identifies the CapEx data revision, not article publication or verification. | |
| Guide sensitivity: annualized cost at exactly 3, 5 and 7 IT years using Epoch’s May 2026 inputs | $12B / $8.8B / $7.3B | derivedGuide sensitivity with fixed Epoch May 2026 capital, annual operating cost and WACC. Only economic earning life changes; book depreciation is separate. The workbook default is about 5.3 years, not exactly five. | — | 2026-05 | checking… | — | Epoch AI, Amelia Michael and Ben Cottier, May 2026 workbook; Guide fixed-life calculation. | |
| liquid-cooling retrofit headline capex vs greenfield liquid-cooled build, per MW · contested | ~$2M/MW vs upwards of $11M/MW | estimate | — | 2026 | checking… | — | STL Partners, 'The retrofitting roadmap: An evolution of liquid cooling' (press summary, 28 May 2026). Headline capex only; STL cautions that customer migration, downtime, lost tenancy revenue and legacy constraints can change the investment case. The release contains no 20-rack case and no performance-ratio figure. | |
| HBM3E is deployed with NVIDIA GB300; HBM4 is in high-volume production for NVIDIA Vera Rubin | HBM3E → GB300; HBM4 → Vera Rubin | observed | — | 2026 | checking… | — | SK hynix GTC 2026 account; Micron HBM4 production release | |
| US public-water-system PFAS settlement payments, 2024-2036 (final approval March 2024); 3M recorded a $10.3B pretax charge in Q2 2023 | $10.5-12.5B nominal (2024-2036) | observedThe range is the nominal total 3M will pay under the settlement. The $10.3 billion pretax charge recorded in 2023 is a separate, present-value accounting figure. | — | 2024-2036 | checking… | — | 3M 2025 Form 10-K, Note 17 (filed 2026-02-03); the Q2 2026 Form 10-Q repeats the figures | |
| Meta observed failure-rate change for jobs using at least 512 GPUs following its lemon-node mitigation | 14% → 4% | observedNamed RSC deployment and event definition; a useful external comparison, not a required calibration target for other fleets. | — | 2024-10 | checking… | — | Meta paper. | |
| 800G/1.6T optical transceiver lead time for volume orders: 40+ weeks in distributor reports, set by EML laser, PAM4 DSP and packaging supply. NVIDIA's March 2026 agreements ($2B equity in each of Lumentum and Coherent, plus multibillion-dollar purchase commitments and capacity rights) tied up EML output; McKinsey expects 800G supply to run 40–60% short of demand through 2027 and 1.6T 30–40% short through 2029. · contested | 40 wk or more (distributor-reported); EML capacity allocated past 2027 for buyers outside NVIDIA's agreements | estimate | — | 2026-07 | checking… | — | not placed | Utmel transceiver shortage outlook (distributor, 2026-07-09) |
| an abandoned factory shell became a 100k-GPU training cluster (xAI Colossus, Memphis); xAI claims 200k after 92 more days | 122 days | observed | 2024-10 (122 days, NVIDIA); 2025-02 (92-day doubling, xAI) | checking… | — | NVIDIA Newsroom (Spectrum-X / xAI Colossus), 2024-10-28, for the 122 days; xAI, Colossus page (x.ai/colossus), for the 92-day doubling (company claim) | ||
| ABB synchronous-condenser orders announced for VoltaGrid: 27 booked in Q1-Q3 2025 plus 35 additional orders announced 26 March 2026; ordered, not installed | 62 sync condensers | observed | — | 2026-03-26 | checking… | — | ABB | |
| accelerated GPU economic life — contested: bear case 2–3 yr on obsolescence; Goldman Sachs estimates 4–6 yr useful life vs 5–6 yr book · contested | 2–3 yr (bear case) vs 4–6 yr (GS) | forecastGoldman Sachs characterizes four to six years as the estimated useful life of AI accelerators; it does not substantiate a general two-to-three-year economic life or a universal 20–40% three-year residual value. | — | 2026 | checking… | — | Goldman Sachs Global Institute, Tracking Trillions (April 2026), for the 4–6 yr useful life; Michael Burry (November 2025, reported by CNBC 2025-11-11) for the 2–3 yr bear case | |
| accelerator-utilization (AU) threshold for a valid MLPerf Storage v3.0 training result, by workload | Unet3D ≥90%, RetinaNet ≥85% | observedv3.0 emulates NVIDIA B200 or AMD MI355 accelerators. The v2.0 round (August 2025, H100 emulation) used 90% for 3D U-Net and ResNet-50 and 70% for CosmoFlow; thresholds and emulated accelerators are not comparable across versions. | — | 2026-09 | checking… | — | MLCommons, MLPerf Storage benchmark definitions (v3.0); v3.0 results published 2026-09-01 | |
| gross acres per MW of planned IT capacity at Vantage's Frontier campus (project example) | ~0.86 | derivedFrontier's announced acreage and planned IT capacity. Another project's capacity comes from its own site layout. | — | 2025 announcement | checking… | — | Vantage Data Centers | |
| active generation + storage in US interconnection queues (end-2025; >1.5x US installed capacity; down from 2,290 GW end-2024) | >2,060 GW | observedGeneration and storage seeking transmission interconnection — supply-side context, not a large-load service queue. LBNL attributes the year-on-year decline to both withdrawals and fewer new requests and cautions that reform effects are not yet measurable. | end-2025 | checking… | — | LBNL, Queued Up: 2026 Edition (data through end-2025) | ||
| Vertiv Liebert CRV chilled-water in-row units: net sensible cooling capacity per unit | 27.1 / 46.3 / 53.2 kW per CW unit | observedCR032, CR040 and CR050 respectively, at 95 F return dry bulb and 23% RH; CR032 and CR040 at 45/55 F entering/leaving water, CR050 at 55/68 F. The CR035 DX and water/glycol variants have separate schedules. Converting a per-unit duty into kW per rack depends on the number of units serving the row, airflow distribution, containment and the redundancy case. It is not a rating for a close-coupled fan wall. | — | 2025-05 | checking… | — | Vertiv | |
| local opposition groups active in 49 states as of September 2026; there were 396 at end-2025 | 843 | observed | — | 2026-09-21 | checking… | — | Data Center Watch, reported by Bloomberg via Insurance Journal, September 21, 2026 | |
| active RDHx capacity per rack — Motivair ChilledDoor is rated up to 75 kW with 100% heat removal | up to 75 kW — Motivair ChilledDoor | observed | — | 2026-09 | checking… | — | Motivair ChilledDoor product page; HPE QuickSpecs, Motivair Liquid Cooled Doors (V4, 2022-04-04) for the rating condition | |
| Growth in North American (US and Canada) natural-gas pipeline capacity needed by 2052 over 2022 levels, as projected by the INGAA Foundation | +39% | forecastCovers the US and Canada from 2022 to 2052; the 39% is the increase over 2022 capacity. | — | 2026-03 | checking… | — | INGAA Foundation, North American Midstream Infrastructure fact sheet | |
| Commerce announced buyer-specific approvals for G42 and HUMAIN, each equivalent to up to 35,000 NVIDIA Blackwell GB300 chips | Up to 35,000 GB300-equivalents per buyer | observedNovember 19, 2025 announced authorization ceiling for each of two named companies. Not a count of physically shipped chips and not a destination-wide authorization. | — | 2025-11-19 | checking… | — | U.S. Department of Commerce — November 19, 2025 announcement; buyer-specific ceiling confirmed in the official press-release index. Full licenses were not obtained; current shipment permission is unverified. | |
| Delivery lead time for a new aeroderivative gas turbine, class-level market estimate (Enverus, reported by Reuters, September 2026) | ~40 mo | estimateA market-wide estimate for new units, not an OEM quote. GE's aero-products supply agreement (SEC filing, December 2024) put LM slot capacity as full for 2025 and 2026, and sets a 24-month lead-time goal only for orders placed after 1 January 2027. Units already in rental fleets are a separate channel: APR Energy advertises 30–90 days to deploy its own turbines. | — | 2026-09 | checking… | — | Reuters (republished by Pipeline & Gas Journal, September 2026), citing Enverus | |
| aggregate NVLink bandwidth in one GB200 NVL72 domain (72 GPUs, 13.4 TB HBM3E) | ~130 TB/s | observedNamed NVIDIA generation/platform, bidirectional aggregate convention; not delivered collective bandwidth, cache coherence or arbitrary mixed-platform performance. | 2026 | checking… | — | NVIDIA GB200 NVL72 product page (undated; checked 2026-10-16) | ||
| ICONIQ respondent expectation for AI-application gross margin in 2026 | ~52% expected | forecastExpected 2026 AI gross margin in the ICONIQ survey population; not a measured whole-industry result or a directly comparable mature-SaaS margin series. | — | 2026 | checking… | — | ICONIQ 2026 State of AI Bi-Annual Snapshot; publication day not stated in the accessible page, year recorded. | |
| US data-center natural-gas consumption forecast for 2030, RBC Capital Markets | ~6 Bcf/d | forecastRBC's forecast covers US data centers, and its likely range is 6–7 Bcf/d. | — | 2026-05 | checking… | — | RBC Capital Markets, Natural gas powers the data center boom | |
| AI accelerator refresh interval in an aggressive first-owner/economic case (vs 5–6-year published book lives) | 18-36 mo | estimateA stress scenario for the fastest economic refresh, not an observed fleet average. Published server book lives are 5–6 years (Amazon 5–6, Meta 5.5, CoreWeave 6; see 14.9). | — | 2026 | checking… | — | Guide scenario (Chapters 1.8, 14.9); ITAD-vendor framing (STS Electronic Recycling, Apr 2026) | |
| AI/HPC scheduler share: Slurm / Kubernetes / in-house (rule of thumb) | ~70% / ~20% / ~10% | estimateTrade-press rule of thumb with no stated sampled population or method; not a fleet census or a scheduler selection criterion. | — | 2026-05-15 | checking… | — | HPCwire, Slurm vs Kubernetes in the Age of AI, 15 May 2026; denominator not specified. | |
| AI inference capacity to 2030 (~35% CAGR) vs training 23.1 → 62.2 GW (~22%) | 20.9 → 93.3 GW | forecastMcKinsey base-case scenario for AI workload capacity growth to 2030, not observed capacity. Compare it only against other capacity forecasts on the same denominator. | — | 2026 | checking… | — | McKinsey, 'The next big shifts in AI workloads' | |
| Goldman Sachs Global Institute's May 2026 estimate of construction cost per MW for next-generation AI data centers | ~$15-20M/MW | estimateGoldman budgets compute, data centers and power separately but does not publish a full list of what its construction figure includes or the MW denominator it uses. Reconcile those boundaries before combining budgets. | — | 2026-05 | checking… | — | Goldman Sachs Global Institute, 'Tracking Trillions: the assumptions shaping the scale of the AI build-out' (2026-05-01) | |
| EPA Region 8 indicates about 60-90 days to review a minor-source permit with no comments or refined modeling, and 6-12 months for a synthetic-minor or PSD permit with comments or refined modeling | ~60 days-12 months (Region 8 indicative estimates) | estimateOne region's examples, not a national maximum and not the whole project schedule. A complete application, PSD pre-application modeling and any required ambient monitoring all come before the review clock starts. | — | 2026-08-24 | checking… | — | U.S. Environmental Protection Agency, Region 8 | |
| Micron completed price-and-volume agreements for its entire calendar-2026 HBM supply, including HBM4 | Micron CY2026 HBM supply contracted | observedMicron's own disclosure of its calendar-2026 supply (December 17, 2025). SK hynix and Samsung made separate statements that their 2026 HBM was committed. Not an allocation to a specific buyer. | — | 2025-12-17 | checking… | — | Micron — Micron FQ1 2026 earnings presentation, December 17, 2025. | |
| AMD's Instinct Kubernetes reference architecture lists AGFHC burn-in recipes of about 12 hours (production qualification) and about 24 hours (maximum stress validation) | ~12–24 h (two recipes: ~12 h, ~24 h) | recommendationAMD reference-design recipe names: about 12 h production qualification, about 24 h maximum stress validation. AMD's health-check design gates each new node on a 24 h burn-in. AGFHC is not publicly available; access requires AMD authorization. A vendor-published test recipe for AMD Instinct nodes, not a universal acceptance bound or an NVIDIA-platform procedure. | — | 2026-10 | checking… | — | not placed | AMD Instinct Kubernetes Reference Architecture, GPU Server Intake & Validation (AGFHC Recipe Options table); undated page, read 2026-10-16 |
| AMD EPYC 9965 (Turin, 5th Gen) host CPU, DDR5-6400; Intel Xeon 6 up to 128 P-cores with MRDIMM-8800. Successor 6th Gen EPYC (Venice) announced July 2026 with up to 256 cores. | 192 cores / 500 W | observedTurin is the shipping-volume reference. Venice SKU power and general availability are not yet published. | — | 2025 | checking… | — | AMD (EPYC 9965 product page; AMD release 2026-07-23); Intel; Phoronix | |
| AMD Helios (72× MI455X) rack-scale FP4 (1.4 EF FP8), 31 TB HBM4, ORW double-wide | ~2.9 EF FP4 | observedAMD's advertised rack spec (Helios product page). AMD's page does not say whether the 2.9 EF FP4 figure is dense or sparse (checked 2026-10-16). Helios is in production; initial systems are installed at OpenAI, Meta, Microsoft and Oracle (AMD, 2026-10-12). | — | 2026-10 | checking… | — | AMD, Helios product page, Rackscale Solution Specs and footnote 2. | |
| AMD Helios rack-scale systems (72 x MI455X, EPYC Venice): in production; initial systems installed and running at OpenAI, Meta, Microsoft and Oracle, production ramping | In production; initial systems installed (Oct 2026) | observedVendor statement (AMD OCP 2026 keynote recap, 2026-10-12). No shipment start month or volumes published. Helios is a reference design for OEM/ODM products; procurement requires that supplier's allocation and support commitment. | — | 2026-10-12 | checking… | — | not placed | AMD newsroom, OCP 2026 Lisa Su keynote takeaways (2026-10-12); AMD Advancing AI launch (2026-07-22/23) |
| AMD MI350-series (MI350X/MI355X) 8-GPU node: switchless full mesh, seven Infinity Fabric links per GPU | 1,075.2 GB/s per GPU (7 × 153.6 GB/s) | observedBidirectional per-link rate in a fully connected 8-GPU node; a GPU pair gets one link by default, and collectives on fewer than eight GPUs use a fraction of the aggregate. | — | 2026-09 | checking… | — | not placed | AMD ROCm documentation, workload optimization (checked 2026-09-29). |
| AMD MI355X FP4 (no sparsity) and HBM3E capacity; 1,400 W peak board power | ~10.1 PFLOPS / 288 GB | observed | — | 2025 | checking… | — | AMD Instinct MI355X GPU datasheet (LE-92703-00, 10/25) | |
| AMD MI355X MXFP4 matrix peak, CDNA4 | ~10.1 PFLOPS | observed | 2026-10 | checking… | — | AMD Instinct MI355X GPU brochure, peak theoretical performance table | ||
| AMD–OpenAI multi-year, multi-generation Instinct agreement; first 1 GW (MI450) planned from H2 2026; warrant for up to 160M AMD shares | 6 GW | observedAnnounced agreement totals, not installed capacity. AMD's 2026-07-23 release expected OpenAI to bring Helios online from Q4 2026, accelerating through 2027. AMD's 2026-10-12 update reports initial systems installed and running at OpenAI, without volumes. | 2025-10 | checking… | — | AMD / TechCrunch / DCD | ||
| AMD ROCm 10.1.0 release, 5 October 2026 | ROCm 10.1.0 (release snapshot) | observedVersioned software release; hardware, OS, model operators and communication features require the selected compatibility matrix. Known issues: RCCL all-reduce with the Low Latency (LL) protocol might produce incorrect results on MI350 Series GPUs (workaround NCCL_PROTO=LL128,Simple); Hugging Face training of BERT, RoBERTa-large, DistilBERT-base and GPT-2 on MI350X might take 6–22% longer than expected (first reported in 10.0.0; stated in the latest release notes). Not an assertion of equal performance or complete API coverage. | — | 2026-10-05 | checking… | — | AMD ROCm release notes, version 10.1.0, 5 October 2026 (known issues). | |
| Americans who oppose a local AI data center / who worry new data centers will raise their energy bills | ~70% / 78% | observedGallup fieldwork March 2-18, 2026, published May 13; Consumer Reports survey of 2,146 US adults in November 2025, released March 11, 2026 | — | 2026 | checking… | — | Gallup, local AI data-center construction survey (March 2026); Consumer Reports, energy survey (November 2025) | |
| Bain’s September 2025 modeled annual monetization gap for anticipated 2030 AI demand | ~$800B | modeledBain’s stated scaling and monetization assumptions, including reinvested savings. Not an observed shortfall or the probability of the guide’s bubble scenario. | — | 2025-09 | checking… | — | Bain & Company, Global Technology Report press release, 23 September 2025. | |
| Water-side economizing energy saving, heat-source system (chillers, cooling towers, pumps), one Suwon laboratory building | 52.8% modeled for Jan–Mar and Nov–Dec; 42.0% measured in a separate 42-day trial | modeledOne laboratory building in Suwon, Korea. The 52.8% is a simulation of the five economizer months (January to March, November and December), averaged over five weather years: 574.1 MWh without economizing against 270.8 MWh with it. The 42.0% is a measurement over 28 December 2023 to 8 February 2024 against the same period a year earlier. The boundary is the heat-source system: chillers, cooling towers and pumps. Neither number is a full-year saving. | — | 2024-12 | checking… | — | Kim, Lee & Kim, J. Korean Inst. Archit. Sustain. Environ. Build. Syst. (2024-12-30) | |
| annual EED reporting deadline (prior calendar year); FY2023 first cycle was 15 Sep 2024 | 15 May | observedAnnual operator submission to the European database for the preceding calendar year | 2026-08-24 | checking… | — | Commission Delegated Regulation (EU) 2024/1364 | ||
| ANSI/NETA Acceptance Testing Specifications — the current as-installed bar for switchgear, breakers, relays and primary injection | ATS-2025 | observed | — | 2025 | checking… | — | ANSI/NETA ATS-2025; NETA World Journal | |
| ANSI/TIA-942-C publication date; Rated terminology predates Revision C | May 2024 | observedTIA announced Revision C on May 9, 2024. Rated terminology predates this edition; TIA/Uptime announced the Tier-name separation in March 2014. Licensed clause-level conformance and local adoption require separate records. | — | 2024-05-09 | checking… | — | TIA, Revision C announcement, 9 May 2024; Rated terminology predates this revision. | |
| Anthropic run-rate revenue: company-announced figure and later press-reported figure | >$47B (company, May 2026); ~$65B reported end-Jul 2026 | observedRun-rate, gross of cloud-partner sales. The $47B figure is Anthropic's own (May 2026 announcement); the ~$65B figure was reported by Bloomberg and Reuters on 2026-08-17 citing a person familiar and was not published by the company. Not audited or recognized revenue. | — | 2026-08 | checking… | — | not placed | Anthropic Series H announcement (2026-05-28); Reuters and Bloomberg (2026-08-17) |
| Data-center securitization issuance (ABS plus CMBS), H1 2026, by region (Apollo) | US $18B / EU $0.9B / UK $1.1B | estimateApollo's tally of issued deals, data as of August 2026; includes ABS and CMBS. value_central is the US figure. Not the Morgan Stanley ABS-only forecast. | — | H1 2026 | checking… | — | not placed | Apollo Daily Spark (Huw van Steenis), Europe AI build-out is missing a financing market, 2026-09-17 |
| NVIDIA architecture cadence: Blackwell 2024, Blackwell Ultra 2025, Rubin 2026 (shipping since June 2026), Rubin Ultra (NVIDIA targets H2 2027; SemiAnalysis reported 2028 in July 2026), Feynman 2028 (roadmap) · contested | annual | forecastA published annual platform roadmap. The Rubin Ultra and Feynman entries are targets, not observations. | — | 2026-10 | checking… | — | NVIDIA Developer Blog, Vera Rubin POD (2026-03-16); The Next Platform (2025-03-19) | |
| ASHRAE's published commissioning references and the status of its data-center commissioning guideline | G0-2019 / G1.1-2025 / Std 202-2024; G1.6P in first public review (Oct 2026) | observedPublished editions per ASHRAE's register; 1.6P is a public-review draft, not a published guideline. Its draft scope reaches 'up to the server racks and IT equipment'. Adopted editions determine project applicability; Guideline 1.1 covers new HVAC&R commissioning. | — | 2026-10 | checking… | — | ASHRAE, Titles, Purposes and Scopes register. | |
| ASHRAE liquid classes by max supply water temp (17/27/32/40/45 °C, plus W+); all share a 2 °C lower limit | W17 / W27 / W32 / W40 / W45 / W+ | observed | — | 2024 | checking… | — | ASHRAE TC 9.9 Thermal Guidelines, 5th ed. (revised reference card, Table 3.1) | |
| ASHRAE–PNNL–NEMA AI Data Center Energy Performance Framework release | 10 June 2026; guidance, not mandatory requirements | observedLifecycle design, commissioning, operation and retrofit guidance; does not supersede codes or standards. | — | 2026-06-10 | checking… | — | ASHRAE release announcement, published 2026-06-10; release and framework scope. https://www.ashrae.org/about/news/2026/ashrae-nema-and-pnnl-release-ai-data-center-energy-performance-framework-to-guide-next-generation-design-and-operation | |
| ASHRAE A1–A4 recommended inlet band; allowable A1 15–32, A2 10–35, A3 5–40, A4 5–45 °C; H1 narrower at 18–22 recommended / 15–25 allowable | 18–27 °C | observedClasses A1–A4. The high-density H1 class has a narrower 18–22 °C recommended and 15–25 °C allowable range, and the equipment OEM determines which class applies. Allowable exposure is qualified operation, not the reliability of recommended conditions, and the associated humidity and altitude limits apply alongside the dry-bulb band. | — | 2024 | checking… | — | ASHRAE TC 9.9 Thermal Guidelines (5th ed.) | |
| ASHRAE fifth-edition equipment air and liquid classes; separate 2024 liquid-cooling resiliency bulletin | A1–A4 + H1 / W17, W27, W32, W40, W45, W+ | observedEquipment capability classes, not chosen operating points. Do not infer a rack TCS temperature, FWS setpoint, return temperature, or PUE from a class alone. | — | 2021 | checking… | — | ASHRAE AI Data Center Energy Performance Framework, Figure 2, confirms W17/W27/W32/W40/W45/W+. H1 is recorded in ASHRAE’s fifth-edition Thermal Guidelines; the 2024 resiliency publication is a separate Technical Bulletin. H1 primary reference: https://www.ashrae.org/file%20library/technical%20resources/bookstore/emergence-and-expansion-of-liquid-cooling-in-mainstream-data-centers_wp.pdf (2021), fifth-edition discussion. | |
| ASHRAE TC 9.9 Thermal Guidelines edition that renamed the liquid W-classes (W17/W27/W32/W40/W45/W+) and added W40 and the H1 air class | 5th ed. (2021) | observed | — | 2021 | checking… | — | ASHRAE, Emergence and Expansion of Liquid Cooling in Mainstream Data Centers (2021), p. 4: fifth-edition class changes. | |
| RAND assesses eight of 38 weight-theft vectors as likely infeasible for OC1–OC3 but likely feasible for OC4 or OC5 under its defined capabilities | 8 of 38 | estimateRAND 2024 expert assessment under its defined adversary capabilities, not a categorical impossibility statement. RAND's main text counts eight (printed p.18); its Table 5.2 scores nine vectors this way, and its summary says about a dozen are infeasible for non-state actors. | — | 2024-05-30 | checking… | — | RAND, Securing AI Model Weights, 30 May 2024, printed page 18 (PDF page 28). | |
| Point-source far-field geometrical spreading per distance doubling; setback economics depend on land and receptor geometry | ~6 dB | derivedAbout 6 dB per distance doubling for unchanged far-field point-source directivity and propagation terms. Setback adds no attenuation hardware or its maintenance, but consumes land; compare its land and layout cost with source treatment. An extended line-source regime approaches 3 dB per doubling. | 2017 | checking… | — | ASHRAE Handbook, Sound and Vibration, logarithmic distance attenuation. | ||
| total Sweden-Finland transfer capability in each direction after the 380 km Aurora Line was commissioned on 13 Nov 2025 (+~900 MW FI→SE / +~800 MW SE→FI) — which also raised permitted withdrawal around Luleå/Boden from 345 to 745 MW | ~1,900 MW | observedNext planned link, Fenno-Skan 3 (~800 MW), is not expected until ~2038. | — | 2025-11 | checking… | — | not placed | Fingrid; Svenska kraftnät |
| Time for a rated GPU cloud to replace a failed HGX-class node after SemiAnalysis injected a fault (ClusterMAX 3.0). SemiAnalysis's bar is under one hour end to end, with detection in about two minutes. | 24–35 min automatic; 43–57 min with manual steps | observedHGX-class nodes with hot spares. Fully automatic replacements took 24 min (AWS EKS Auto Mode) and 35 min (Together AI). TensorWave (~43 min) needed a click to approve, and Hyperstack (52 min) and Moonlite (57 min) relied on manual SRE steps. A failed NVL72 tray cannot be hot-swapped: Nebius returned a GB300 node to service automatically in 8 h 40 min. ClusterMAX 1.0 (Mar 2025) said top clouds spin up new nodes within 90 seconds, an unattributed observation rather than a measurement. | — | 2026-09 | checking… | — | SemiAnalysis ClusterMAX 3.0 | |
| average annualized insider-risk cost in the cited report: $19.5M across 354 organizations (2025 data), up from $17.4M in 2024 | $19.5M | observedSurvey/benchmark population includes negligent, malicious and credential-theft incidents; not a data-center-specific loss estimate. | — | 2026 | checking… | — | Ponemon Institute, 2026 Cost of Insider Risks Global Report; DTEX sponsor report page. Study observations concern 2025. | |
| average breach cost in IBM's 2026 Cost of a Data Breach report: $4.99M global and $11.5M US | $4.99M global / $11.5M US | observedOne IBM/Ponemon report vintage (2026). The global average rose 12% from $4.44M in the 2025 report; AI-enabled malicious breaches averaged $6M. Do not mix 2025 and 2026 values. | — | 2026 | checking… | — | IBM/Ponemon Cost of a Data Breach 2026 (IBM release 2026-07-29; IBM Think data-breach topic page) | |
| average eCrime breakout time (initial access to first lateral movement) in 2025, down from 48 min in 2024; fastest 27 s | 29 min | observed | — | 2025 | checking… | — | CrowdStrike 2026 Global Threat Report | |
| Average time to contain an insider incident: 67 days in 2025, down from 81 days in 2024; far longer than a checkpoint copy takes | 67 days | observedTime to contain in the study population, not a model-theft detection budget or time to copy a checkpoint. | 2026 | checking… | — | Ponemon Institute, 2026 Cost of Insider Risks Global Report (2025 data) | ||
| AWS facilities hit by drones in the UAE, plus one blast-damaged in Bahrain, Mar 2026; likely the first publicly confirmed military attack on a US hyperscaler's data centers | 2 hit + 1 damaged | observedBy September 2026 AWS said the Bahrain region and resources held only in UAE zone mec1-az2 could not be restored; the damage exceeded what its multi-AZ services are designed to withstand. The IRGC attribution and intent rest on IRGC-linked media. | — | 2026-03 | checking… | — | DefenseScoop, 2026-03-03; AWS status update reported by Reuters, 2026-09-16 | |
| AWS Trainium3 availability — shipping since start of 2026 and nearly fully subscribed; Trainium2 largely sold out; Trainium4 planned for 2027; AWS in-house chips (Trainium+Graviton+Nitro, blended) ~$25B annualized with ~$225B multi-year commitments | nearly fully subscribed | observed | — | 2026-08 | checking… | — | not placed | Amazon Q1 2026 earnings, Andy Jassy (2026-04-30): Trainium3 nearly fully subscribed, Trainium2 largely sold out, Trainium4 ~18 months from broad availability, >$225B Trainium commitments; Amazon Q2 2026 earnings call (2026-07-30): chips business run rate over $25B |
| Guide footprint-average screening range for 3,000–3,300 lb over approximately 0.64 m² | ~435-480 psf | derivedThe class-mass endpoints are screening assumptions, not HPE minimum and maximum ratings. HPE's GB200 QuickSpecs give a separate 3,245 lb filled-rack figure. Neither result is a support, rolling or uniform floor rating. | — | 2026-09 | checking… | — | HPE GB200 QuickSpecs V6, September 8, 2026: system dimensions and loaded-with-fluid mass; the guide range uses its own class-mass screening assumptions. | |
| Modeled back-end fabric cost saving from virtual-modular-switch layout plus 2:1 oversubscription versus NVIDIA's non-blocking 1,024-H100 reference architecture (SemiAnalysis, October 2024; single-source) · contested | ~31% | modeledSemiAnalysis estimates a 31.6% total saving for the combined design versus the NVIDIA reference architecture; the non-blocking virtual-modular-switch layout alone saves 24.9%. These are modeled design-specific costs, not observed fleet savings. | 2024-10 | checking… | — | SemiAnalysis, AI Neocloud Playbook and Anatomy | ||
| back-end NICs per training node = 3.2 Tb/s/node; 1:1 non-blocking 8-rail fat-tree | 8× 400 Gb/s | observedThe GB200/HGX case: one 8-rail training reference design, not a workload-name mandate. The source describes an 8-GPU HGX node (8×400 Gb/s per node); a GB200 NVL72 compute tray holds 4 GPUs (4×400G per tray). Appendix B's counts use one 400 Gb/s back-end attachment per GPU. GB300 NVL72 uses an 800G ConnectX-8 per GPU (two 400GbE planes) and Vera Rubin 1.6 Tb/s per GPU, so never size those fabrics at 400G per GPU. Validate the port rate, rail count and blocking ratio against the target traffic matrix, topology, placement, failure headroom and step-time target. | — | 2025 | checking… | — | SemiAnalysis AI Neocloud Playbook | |
| Back-end NICs per GPU: one 400G/800G on GB200/B300, two 800G ConnectX-9 on Vera Rubin NVL72; plus a front-end/storage NIC | 1–2 per GPU | observedOne per GPU through Blackwell (GB200 SuperPOD reference architecture). Two 800G ConnectX-9 per GPU on Vera Rubin NVL72 (NVIDIA Rubin platform blog: 1.6 Tb/s per GPU). | — | 2025 | checking… | — | NVIDIA GB200 SuperPOD reference architecture; NVIDIA Rubin platform blog | |
| Bain's September 2026 AI data-center buildout scenario: total buildout to 2030 and the capacity it adds | $5–6.5T buildout; about 150 GW or more by 2030 | forecastBain's own buildout scenario, including the power, chips, skilled labor and permitting constraints it names. Not the same metric as McKinsey's $3.7T–$7.9T AI-capable capex range: do not net or average the two. Bain also counts at least 75 projects worth $130B blocked or delayed by local opposition in Q1 2026. | — | 2026-09 | checking… | — | not placed | Bain & Company, Technology Report 2026: AI data-center boom: can we build it if they come, 29 September 2026 |
| Bain's September 2026 model: new annual AI revenue still needed in 2031 to reach the ~$6T AI market that would sustain ~$1.5T/yr of AI infrastructure spend | ~$4.2T | modeledBain sizes the consumer and enterprise AI market at $1.2–1.8T and computes the gap to ~$6T from the top of that range, assuming capex near 25% of revenue. A model requirement on a 2031 horizon, not an observed shortfall; not comparable with Bain's 2025 ~$800B gap for 2030, which used a different model. | — | 2026-09 | checking… | — | Bain & Company, Technology Report 2026: New innovation is required to fund AI's $6 trillion buildout, 2026-09-29 | |
| Bartz v. Anthropic: $1.5 billion settlement fund approved in the final judgment dated 20 July 2026 | $1.5B settlement fund | observedSettlement fund approved in Bartz v. Anthropic, not proof of completed distribution, a per-work net payout, appellate finality or a general permission to train. | — | 2026-07-20 | checking… | — | US District Court, final approval and judgment, Document 680, 20 July 2026; official settlement-administrator court-document copy. | |
| Estimated cut in total battery capacity, and the double-conversion loss avoided, when one distributed battery set replaces duplicated A- and B-side central UPS battery sets | ~50% / 2–3% | estimateSemiAnalysis analyst estimate, not a field measurement. Dropping the separate A-side and B-side central UPS batteries yields the roughly 50% capacity reduction, and removing the central AC-DC-AC pair avoids its 2-3% loss. The installed result depends on the runtime and redundancy you need and on any central conversion you keep. | — | 2026-05 | checking… | — | SemiAnalysis, Inside the 800VDC Revolution, Part 1 (2026-05-26) | |
| Cleanview’s mid-2026 US behind-the-meter generation announcement stock | ~90 GW announced | estimateTracker estimate of announced US behind-the-meter generation at 59 data centers, largely gas turbines and engines by equipment type (Cleanview publishes no gas share); not contracted output, operating supply or a forecast that all announcements will commission. Public summary reports about 2 GW operating and 1.2% under construction. | mid-2026 (publisher’s observation window) | checking… | — | Cleanview, Michael Thomas, Bypassing the Grid, public report; explicitly describes mid-2026 operating capacity and gives no publication day. Verification on September 8, 2026 does not create a source date. | ||
| Cleanview: announced US data-center behind-the-meter generation since January 2025, largely gas turbines and engines (about 2 GW operating mid-2026) | ~82 GW | estimateTracker count of announced US behind-the-meter generation capacity since January 2025 — not contracted output, permitted plant or operating supply. The same tracker's cumulative announcement stock is a separate claim. | 2026-07 | checking… | — | Cleanview | ||
| Meta Llama 3 405B pre-training: 41% BF16 MFU on 16,384 H100 GPUs at 8,192-token sequence length | ~41% | observedNamed 405B pre-training configuration, BF16 MFU against the H100's dense BF16 peak; other Table 4 layouts report 38% or 43%. No Hopper fleet-wide MFU band or universal acceptance threshold follows. | 2024-07 | checking… | — | Meta, The Llama 3 Herd of Models, Table 4, July 2024 report. | ||
| BF16 MFU gain on H100 (GPT-3 175B) from software and kernel maturation over ~12 months, Jan–Dec 2024, on unchanged silicon | 34% → 54% | observedMeasured on a 128-H100 cluster, January-December 2024, as BF16 MFU against the H100's dense BF16 peak. The endpoints are rounded MFU percentages, so the ratio of the two (54/34) is approximate. | — | 2024-12 | checking… | — | SemiAnalysis, H100 vs GB200 NVL72 Training Benchmarks, 2025-08-20 | |
| BIS moves qualifying direct US advanced-computing exports to China/Macau to case-by-case review | Jan 15 2026 | observedItems must stay under set performance thresholds and the applicant must make the rule's certifications. Reexports and in-country transfers keep a presumption of denial. | — | 2026-01-15 | checking… | — | US BIS final rule, Federal Register, January 15, 2026 | |
| 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 mo | observedOctober-2024 launch-period backlog for new B200/GB200 orders. A historical allocation datapoint, not a current quote: schedule a 2026 procurement decision from a product-specific, clock-defined entry on the live lead-time register. | — | 2024-10 | checking… | — | Morgan Stanley via Data Center Dynamics (2024-10-11) | |
| Blackwell (B200) confidential-computing overhead for a tuned LLM serving stack; stock stacks and eight-GPU training cost more | ~1–3% tuned serving | observedOne physical host with Intel TDX and NVIDIA Blackwell CC; ~1–3% tuned serving, 30–40% stock serving and ~10–13% eight-GPU training in the evaluated configurations. Not a multi-node or all-workload bound. | — | 2026-09 | checking… | — | arXiv, independent paired-run B200 study arXiv:2608.26575v2; submitted 2026-08-27, revised 2026-09-01 | |
| 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 steps | Blackwell ~71% of 2026 | forecast | — | 2026-04 | checking… | — | not placed | TrendForce via The Register (2026-04-08) |
| Bloom–Oracle master agreement supports procurement up to 2.8 GW; initial 1.2 GW contracted, with deployment underway and continuing into 2027. | up to 2.8 GW | observed2.8 GW is the agreement's ceiling, not contracted or operating capacity. | — | 2026-04-13 | checking… | — | Bloom Energy IR / press | |
| BlueField-4 Arm Neoverse V2 complex (64B transistors, ~6× BF-3 compute); 128 GB LPDDR5; PCIe Gen6 | 64-core | observed | — | 2026 | checking… | — | NVIDIA / WWT / HPCwire | |
| NVIDIA advertises up to 800 Gb/s connectivity for the BlueField-4 DPU; the Vera-based BlueField-4 STX storage processor is a separate product profile | 800 Gb/s | observedAdvertised DPU interface capability announced October 2025, not an end-to-end application benchmark or a guarantee that all scale-out traffic traverses this DPU. A vendor-announced specification, not a measured storage-path result. The Vera-based STX storage processor is a separate profile. | 2025-10 | checking… | — | NVIDIA BlueField platform page; BlueField-4 DPU announced at GTC Washington D.C. (2025-10-28) | ||
| BMC runs on standby power and boots before the host; a rooted BMC is an OS-invisible, persistent foothold under the CPU | always-on | observed | — | 2025 | checking… | — | Eclypsium / OCP DC-SCM | |
| BloombergNEF forecast of installed US data-center capacity in 2030 and 2035 (July 2026 outlook) | 118 GW (2030) / 194 GW (2035) | forecastInstalled US data-center capacity as BNEF defines it; the public article does not say whether the GW are IT load or facility power. A forecast, not committed builds or a load-interconnection queue. | — | 2026-07 | checking… | — | BloombergNEF, US data-center capacity outlook (forecast published July 21, 2026) | |
| Ohm's-law screening current at an assumed 1,000 ohm body resistance: about 800 mA at 800 VDC versus about 48 mA at 48 VDC; injury depends on path, duration and actual contact impedance | ≈800 mA | derivedIllustrative Ohm’s-law arithmetic at an assumed 1,000 Ω; potentially fatal exposure, not a prediction of current through a person, an injury threshold, a DC arc model or permission for exposed work. The source’s protective-grounding context does not become a rack-maintenance procedure. | 2026-08-24 | checking… | — | U.S. Occupational Safety and Health Administration, 29 CFR 1926 Subpart V Appendix C, 2025 CFR edition, III.D.1, printed p. 509; GovInfo final text read September 8, 2026. | ||
| SemiAnalysis recommendation for factory high-temperature burn-in before deployment in its October 2024 operator playbook; not an observed settling time or universal acceptance duration | ≥3–4 weeks | recommendationSemiAnalysis recommends at least 3–4 weeks of factory high-temperature burn-in before deployment for the playbook it describes, plus a separate 2–3-day on-site burn-in. The signed project/OEM/contract plan sets the actual campaign and stopping rule. | — | 2024-10 | checking… | — | SemiAnalysis, AI Neocloud Playbook and Anatomy | |
| Broadcom Jericho4 deep-buffer router: 3 nm, HBM packet buffer (capacity not published), RoCE over more than 100 km via 3.2 Tbps HyperPorts (Broadcom claim) | 51.2 Tbps | observedHBM capacity is not published. Broadcom's product copy gives a buffering multiple against on-chip memory without a baseline, so no buffer size can be derived. Buffer bandwidth is 2x Jericho3-AI (Next Platform briefing). | — | 2025 | checking… | — | Broadcom / The Next Platform | |
| 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 2026 | volume ramp Mar 2026 | observed | — | 2026-03 | checking… | — | not placed | Broadcom press release (2026-03); Arista PR (2026-06-09) |
| Tomahawk 6 advertised switching capacity; family shipping in production volume | 102.4 Tbps | observed102.4 Tb/s chip family; 512x200G or 1,024x100G electrical options. Confirm each system's port count and optical variant with its vendor, including support and delivery. Third-generation TH6-Davisson co-packaged optics were shown as partner demos at OCP 2026 (October), not as shipping products. | 2026-03-12 | checking… | — | Broadcom production-volume announcement, 2026-03-12. | ||
| Enabled Energy estimate for an existing-hall upgrade supporting 20–40 kW racks through cooling, electrical resiliency and airflow/containment changes · contested | ~$5–6M per MW | estimateNot a liquid-cooled, NVL72-class conversion; the $2M/MW cooling-only figure and the two-thirds-unsuitable statistic are dropped (not in the source). | — | 2025-07 | checking… | — | Enabled Energy, originally published as DCD opinion, 2025-07-08 | |
| King & Spalding's April 2025 experience: commercial-cloud projects below $500M tended to use GMP; $5–10B AI/hyperscale projects tended to use target-price painshare/gainshare | <$500M | observedLaw-firm experience with compensation mechanisms; not a surveyed market threshold or a rule selecting EPCM | — | 2025-04 | checking… | — | King & Spalding, Building Data Centers – Key Considerations | |
| busway / busduct lead time — CPM Pros puts busway at 30–52 wk, custom engineered runs at the top of the range; fabrication can't start until final dimensioned drawings | ~30–52 wk | observed | — | Re-observed 2026-10-17 | checking… | — | not placed | CPM Pros, long-lead equipment guide (undated; re-observed 2026-10-17) |
| 6SigmaDCX calibration results in two published air-side case studies | 1% grille flow / 5% AHU load; generally <5% tile flow | observed | — | 2016 | checking… | — | Future Facilities 6SigmaDCX case studies (Compass; Binghamton/ES2) | |
| California SB 253 Scope 1/2 first-report date (CARB initial regulation, OAL approved 12 Oct 2026); SB 261 enforcement enjoined | 10 Nov 2026 | observedOAL approved CARB's initial regulation on 12 Oct 2026, and it took effect at once. The first-year reports cover Scope 1 and 2. The initial regulation requires no assurance; CARB will accept 2026 reports with or without it. A separate CARB rulemaking is setting Scope 3 and assurance from 2027. SB 261 enforcement stays enjoined while the Ninth Circuit appeal runs. | — | 2026-10-16 | checking… | — | California Air Resources Board | |
| Caliptra open silicon RoT co-developed by Microsoft, Google, AMD and NVIDIA (2022); AMD has committed to integrate it in its 2026+ products | 4 contributors | observedAMD's commitment is a plan. No public source names a shipping data-center part with Caliptra, so check per part whether the silicon carries Caliptra or a proprietary root. | 2026-10 | checking… | — | CHIPS Alliance Caliptra 2.1 announcement (2025-10-15); AMD blog (October 2024) | ||
| capital premium for Tier IV (fault-tolerant) over Tier III (concurrently maintainable) — construction + fit-out | ~25–40% facility | estimateSecondary-source whole-facility planning estimate quoted by Savills in May 2024; no disclosed estimating population or method. No universal multiplier follows from Uptime Tier criteria. | — | 2024-05 | checking… | — | Savills Research, European Data Centres, May 2024; publisher report. Quoted secondary estimate; not measured project evidence. | |
| Capacity under construction in CBRE's eight primary North American data-center markets, H1 2026, and its preleased share | 7,481 MW; >80% preleased | observedMore than 80% preleased (74.3% a year earlier); less than 1,500 MW left available for preleasing. Asking rents for 3–10 MW requirements rose 8.3% in H1 2026. Eight primary markets only. | — | H1 2026 | checking… | — | CBRE, North American data center demand continues to outpace supply despite record construction, 2026-09-08 | |
| CBRE average asking rent for 10+ MW wholesale colocation requirements, Northern Virginia | $160–185/kW-mo | observedAsking rates for 10-plus-MW requirements in CBRE's H1 2026 Northern Virginia market report; about $1.9–2.2M/MW-year in rent. Asking, not contracted, rates. | — | H1 2026 | checking… | — | CBRE, North America Data Center Trends H1 2026 (Northern Virginia) | |
| ProphetStor-reported CDU energy saving in its Proven Benefits discussion | 22–28% | observedThe reported range belongs to the supplier’s Proven Benefits discussion; it is not established as the result of the separate twenty-job, single-server integration test. | — | 2026 | checking… | — | ProphetStor, Predictive Liquid Cooling for AI Data Centers, Proven Benefits discussion (undated supplier page). | |
| CDU heat-exchanger approach (FWS→TCS) as a design allowance; the OCP Project Deschutes 1.0.0 specification targets a 3 °C approach at 2,000 kW | ~3–5 °C | recommendationThe ~3–5 °C range is the guide's L2L design allowance. Project Deschutes 1.0.0 targets 3 °C at 2,000 kW and 500 GPM on each side (26 February 2026, §5.3). Vertiv's row-based CoolChip CDU 350 is L2A: its March 2025 rating is 350 kW at 15 °C ATD against 27 °C room air. Compare L2L quotes at the same FWS/TCS duty; compare L2A quotes against their room-air condition. | — | 2026-10 | checking… | — | OCP, Project Deschutes: Data Center Facilities Specification 1.0.0 (Google), effective 2026-02-26 |