The Definitive Guide toAI Data Centers
Ask the GuideAboutAccount

Numbers Provenance Register

1,202 canonical active claims across 1,277 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.

1,202 active claims1,277 chapter occurrences555 with revision history20 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 · 96 Derived values are guide calculations; their result is only as fixed as the stated inputs, scope, and method.
  • estimate · 71 Interpret this value according to its displayed kind, scope, source, and as-of date.
  • forecast · 89 Forecasts are time-bound outlooks, not specifications; test decisions across plausible scenarios.
  • modeled · 94 Interpret this value according to its displayed kind, scope, source, and as-of date.
  • observed · 825 Reported measurements, counts, and specifications keep the precision and scope stated by their source; an exact specification is not treated as a range.
  • proposed · 2 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.
1,202 matches · showing 1–100
Canonical claimValueKind / scopeTrendAs ofReviewUpdatedOccurrencesSource
1.6T switch deployment; ramp faster than 800G, >5M ports within 1-2 years of shippingfirst volume yr 2026
forecast
—2026checking…—Semtech press release quoting Dell’Oro Group
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.
—2026-05checking…—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.
—2026-05checking…—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
—2026checking…—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 RubinHBM3E → GB300; HBM4 → Vera Rubin
observed
—2026checking…—SK hynix GTC 2026 account; Micron HBM4 production release
3M completed its exit from PFAS manufacturing at the end of 2025, ending its Novec 1230 production; FK-5-1-12 remains available from other manufacturers.Completed at end-2025
observedcompany: 3M
—2026-08checking…—not placed3M, 3M to Exit PFAS Manufacturing by the End of 2025
3M PFAS settlement with US public water systems (final approval Mar 2024); Novec exit by end-2025$12.5B
observed
—2024-2025checking…—DCD; The Cooling Report; 3M filings
3M's last Novec/PFAS fluorochemical orders; production wind-down through 2025 (drove the two-phase-immersion exit)Mar 2025
observed
—2025checking…—not placedServeTheHome / DCD / C&EN
Meta observed failure-rate change for jobs using at least 512 GPUs following its lemon-node mitigation14% → 4%
observedNamed RSC deployment and event definition; a useful external comparison, not a required calibration target for other fleets.
—2024checking…—Meta paper.
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 wk
observed
—2026-05checking…—not placedMcKinsey via TechTimes (2026-05); Supply Chain Connect (2026); TrendForce (2026-04)
abandoned factory shell to 100k-GPU training cluster (xAI Colossus, Memphis); doubled to 200k in a further 92 days122 days
observed
—2024–2025checking…—NVIDIA Newsroom (Spectrum-X / xAI Colossus)
ABB synchronous-condenser orders announced for VoltaGrid: 27 booked in Q1-Q3 2025 plus 35 additional orders announced 26 March 2026; ordered, not installed62 sync condensers
observed
—2026-03-26checking…—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 · contested2–3 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.
—2026checking…—Goldman Sachs
accelerator-utilization threshold for a passing MLPerf Storage result (70% for Cosmoflow); holding this line is the loader's core objective≥90%
observed
2026-06-29: 90 → 2026-07-07: 902025-08checking…—MLCommons, MLPerf Storage v2.0
accelerators served per storage system in MLPerf Storage v2.0 vs v1.0 — storage scaling tracking compute scaling~2x accelerators
observed
—2025checking…—not placedMLCommons (MLPerf Storage v2.0)
acres per MW of IT on GW-scale AI campuses (full campus incl. substation/cooling yard)0.7–0.9
observed
—2025–2026checking…—Reuters / operator filings (Hyperion, Stargate)
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.
2026-07-01: 2060 → 2026-09-04: 2060end-2025checking…—LBNL Queued Up program data through end-2025 (emp.lbl.gov/queues). The 2025-edition PDF (queued_up_2025_edition_12.15.2025.pdf) is the end-2024 report — 10,303 active projects, 2,290 GW — and is not the source for this end-2025 figure.
in-row cooler capacity per cooling unit (Liebert CRV CR032–CR050); kW per rack depends on unit-to-rack ratio, containment and air conditions~30–50 kW
observedVertiv publishes these as per-unit cooling capacities (CR032/CR035/CR040/CR050). Converting them into kW per rack requires the number of units serving the row, airflow distribution, containment, entering-air conditions and the redundancy case. The figure does not establish an equivalent close-coupled fan-wall rating.
—2025checking…—Vertiv
active local opposition groups across 49 states (more than doubled YoY)833
observed
—Q1 2026checking…—Data Center Watch / 10a Labs
active RDHx capacity per rack — Motivair ChilledDoor is rated up to 75 kW with 100% heat removalup to 75 kW — Motivair ChilledDoor
observed
—2026-07checking…—Motivair ChilledDoor product page
additional US pipeline capacity INGAA projects is needed by 2052 to serve load growth+39%
forecast
—2025checking…—Natural Gas Intelligence / INGAA
Commerce announced buyer-specific approvals for G42 and HUMAIN, each equivalent to up to 35,000 NVIDIA Blackwell GB300 chipsUp 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.
—2025-11-19checking…—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.
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+
observed
—2025checking…—2025 synthesis; Data Center Frontier / power-eng aeroderivative coverage (2026)
vendor-projected shift toward more host CPU capacity for some agentic workloads; ratios require a named accelerator, CPU-package/socket and host denominatordirectional; profile host throughput per workload
observedAn eight-accelerator dual-socket host is 8 accelerators per host but 4 accelerators per CPU package. Do not compare ratios with different denominators.
—2026checking…—not placedTrendForce; Intel (via TrendForce)
aggregate NVLink bandwidth in one GB200 NVL72 domain (72 GPUs, ~13.4 TB unified memory)~130 TB/s
observedNamed NVIDIA generation/platform, bidirectional aggregate convention; not delivered collective bandwidth, cache coherence or arbitrary mixed-platform performance.
2026-06-29: 130 → 2026-09-16: 1302026checking…—NVIDIA NVLink platform page (undated); no publication date invented.
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.
—2026checking…—ICONIQ 2026 State of AI Bi-Annual Snapshot; publication day not stated in the accessible page, year recorded.
AI-driven incremental US gas demand projected by 2030; US gas demand to ~150 Bcf/d by 2031~6 Bcf/d
forecast
—2030-2031checking…—SemiAnalysis / RAND; INGAA
AI hardware refresh interval compressing from a 5-7 yr norm — the engine of recurring embodied carbon18-36 mo
observedhardware refresh interval
—2026checking…—Goldman/CNBC
AI/high-performance storage market 2025, growing ~24–25% CAGR; vendor consolidation underway~$36B / ~24% CAGR
observed
—2025checking…—not placedMarket research (DataM, SNS Insider et al.)
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.
—2026-05-15checking…—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.
—2026checking…—McKinsey, 'The next big shifts in AI workloads'
AI-optimized build cost including tenant technology fit-out, excl. IT/servers — shell+core alone runs ~$11.3M/MW (JLL 2026); do not stack fit-out on top of an all-in facility figure~$15-25M/MW
estimateJLL's 2026 outlook supports ~$11.3M/MW shell and core (single-tenant, ~50 MW, excluding land and active IT) and a separate tenant technology fit-out reaching ~$25M/MW that includes IT such as GPUs and networking. The facility-only $15-25M/MW band therefore rests on Goldman's ~$15M/MW model plus market data, not on JLL's fit-out figure. State the MW denominator (IT MW) and the included infrastructure before adding any silicon budget.
—2026checking…—JLL 2026 Outlook; Goldman Sachs; iRecruit / Construct Elements market data
dense-hall slab capacity has no universal screening allowance; engineer the declared dead, live, concentrated, rolling, rigging, seismic and anchorage load states under the adopted codeProject-specific structural design
recommendationFinal capacity follows site soil, slab, reinforcement, joints, punching, bearing, deflection, vibration, and the declared equipment and movement load states.
—2026-08-24checking…—not placedASCE/SEI 7-22; ACI CODE-318-25
AI storage market (2025), ~24% CAGR; QLC + data reduction displacing HDD across hot and capacity tiers~$36B
observed
—2025checking…—not placedPrecedence Research; MarketsandMarkets (AI-Powered Storage)
AI storage market size and CAGR — the consolidation pressure behind certified-appliance lock-in~$36B / ~24%
observed
—2025checking…—not placedPrecedence Research, AI-Powered Storage Market (2025)
air-permit timing for combustion bridge power is jurisdiction- and case-specific; EPA Region 8 examples span roughly 60 days to a year or more depending on modeling, comments and permit path~60 days-1 year+; case-specific
observed
—2026-08-24checking…—U.S. Environmental Protection Agency, Region 8
SemiAnalysis October 2024 eight-GPU H100 compute-node BOM estimate, before and after host/DPU optimization$270k → $256.4k per node
estimateHistorical 1,024-H100 cluster with 128 eight-GPU nodes. The compute-node subtotal excludes rack/cluster fabric, networked storage, software and facilities. This is an analyst estimate for that configuration, not a 2026 quote or a complete installed-cost range.
—2024-10-03 historical examplechecking…—SemiAnalysis — AI Neocloud Playbook and Anatomy, October 3, 2024; compute-chassis BOM estimate.
Micron completed price-and-volume agreements for its entire calendar-2026 HBM supply, including HBM4Micron CY2026 HBM supply contracted
observedMicron supplier disclosure at 2025-12-17; not all suppliers and not an allocation to a specific buyer.
—2025-12-17checking…—Micron — Micron FQ1 2026 earnings presentation, December 17, 2025.
AMD EPYC 9965 (Turin) host CPU, DDR5-6000/6400; Intel Xeon 6 up to 128 P-cores w/ MRDIMM-8800192 cores / 500 W
observed
—2025checking…—AMD; Intel; Phoronix
AMD Helios (72× MI455X) rack-scale FP4 (1.4 EF FP8), 31 TB HBM4, ORW double-wide~2.9 EF FP4
modeledHelios OEM/ODM reference design; the appendix uses the advertised configuration and separately records the delivery forecast.
—2026-09checking…—AMD, Helios product page, Rackscale Solution Specs and footnote 2.
AMD Helios reference-design systems: AMD expects volume deployments in 2H 2026Volume deployments expected 2H 2026
forecastHelios is a reference design for OEM/ODM products. Procurement requires that supplier’s allocation and support commitment.
—2026checking…—not placedAMD, Helios product page, availability FAQ.
AMD MI355X availability — ramping through 1H26; Oracle first hyperscaler with public instances, building a 27,000+ node MI355X clusterramping through 1H26
observed
—2025-11checking…—not placedInvesting.com earnings-call transcript
AMD MI355X FP4 (no sparsity) and HBM3E capacity; 1,400 W peak board power~10.1 PFLOPS / 288 GB
observed
—2025checking…—AMD Instinct MI355X GPU datasheet (LE-92703-00, 10/25)
AMD MI355X FP4 (sparse), CDNA4, with native FP6 support~20 PFLOPS
observed
—2025checking…—AMD MI355X product page
AMD MI355X/MI325X 8-GPU system lead time — editorial proxy, no OEM-published SKU clock as of mid-2026. MI325X ~6–12 wk standard; MI355X ~8–16 wk; 12–24 wk for large or custom liquid-cooled deployments~8–20 wk
estimate8-GPU systems, accepted orders; proxy range
—2026-07checking…—not placedEditorial proxy triangulated from SLYD OEM lead-time benchmarks (2026-01) and Crusoe MI355X program ~3–6 mo announcement-to-service (2025-06-12); no public OEM SKU-level quote found as of 2026-07
AMD–OpenAI Instinct supply agreement over 5 yr; first 1 GW (MI450) H2 2026; up to 160M AMD shares to OpenAI6 GW
observed
—2025checking…—AMD / TechCrunch / DCD
AMD ROCm 10.0.0 release, 26 August 2026ROCm 10.0.0 (release snapshot)
observedVersioned software release; hardware, OS, model operators and communication features require the selected compatibility matrix. Not an assertion of equal performance or complete API coverage.
—2026-08-26checking…—AMD ROCm release notes, version 10.0.0, 26 August 2026.
Americans opposing a local data center / worried it raises their energy bill~70% / 78%
observed
—2026checking…—Gallup, Americans Oppose AI Data Centers in Their Area (fieldwork by Recon MR, 2-18 March 2026), for the ~70% opposing a local data center; Consumer Reports survey (2026) for the 78% worried about their energy bill
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.
—2025-09checking…—Bain & Company, Global Technology Report press release, 23 September 2025.
annual cooling-energy cut from water-side economizing vs mechanical chilling, temperate climates~40–65%
observed
—2023–2025checking…—Journal of Korean Institute of Architectural Sustainable Environment and Building Systems
annual EED reporting deadline (prior calendar year); FY2023 first cycle was 15 Sep 202415 May
observedAnnual operator submission to the European database for the preceding calendar year
2026-06-29: 15 → 2026-08-24: 152026-08-24checking…—Commission Delegated Regulation (EU) 2024/1364
annual free-cooling hours in a Nordic climate; PUE as low as ~1.09~7,000-8,000+ hr
observedRegional screening figure with the <10 °C average-annual-temperature condition, not a site design basis — Chapter 5.8 runs the site’s climate bins; operator disclosures (Google Hamina 1.09 TTM 2024) are the PUE evidence.
—2026-05-13checking…—w.media, Just chill: cooling innovations in Nordic data centers setting new sustainability standards (Conor McNevin, 13 May 2026): average annual temperatures below 10 °C providing up to 8,000 hours of free-air cooling each year, enabling facility-level PUE as low as 1.09; the article attributes the figures to no primary dataset.
ANSI/NETA Acceptance Testing Specifications — the current as-installed bar for switchgear, breakers, relays and primary injectionATS-2025
observed
—2025checking…—ANSI/NETA ATS-2025; NETA World Journal
ANSI/TIA-942-C publication date; Rated terminology predates Revision CMay 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.
—2024-05-09checking…—TIA, Revision C announcement, 9 May 2024; Rated terminology predates this revision.
architecture cadence — Blackwell 2024, Blackwell Ultra 2025, Rubin 2026, Rubin Ultra 2027, Feynman 2028annual
observed
—2026checking…—The Next Platform
ASHRAE register lists published commissioning references and proposed data-center guidelineG0-2019 / G1.1-2025 / Std 202-2024; G1.6P proposed
observedRegister observation, not a new publication date. Adopted editions determine project applicability; Guideline 1.1 covers new HVAC&R commissioning.
—2026checking…—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 limitW17–W45+
observed
—2025checking…—ASHRAE TC 9.9 Thermal Guidelines 5th ed.; Upsite
ASHRAE–PNNL–NEMA AI Data Center Energy Performance Framework release10 June 2026; guidance, not mandatory requirements
observedLifecycle design, commissioning, operation and retrofit guidance; does not supersede codes or standards.
—2026-06-10checking…—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 / 5–25 allowable18–27 °C
observedClasses A1–A4. The high-density H1 class has a narrower 18–22 °C recommended and 5–25 °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.
—2024checking…—ASHRAE TC 9.9 Thermal Guidelines (5th ed.)
ASHRAE fifth-edition equipment air and liquid classes; separate 2024 liquid-cooling resiliency bulletinA1–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.
—2021checking…—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 class5th ed. (2021)
observed
—2021checking…—ASHRAE, Emergence and Expansion of Liquid Cooling in Mainstream Data Centers (2021), p. 4: fifth-edition class changes.
RAND's 2024 read of frontier-program security posture against the OC4–OC5 adversaries that want the weights~SL2–SL3
estimateRAND RRA2849-1 (2024) proposes the benchmark and discusses capability gaps; it is not a named, dated assessment of named 2026 programs. A current grade requires an identified assessor, date, scope, and evidence.
—2024checking…—RAND, Securing AI Model Weights (RRA2849-1, 2024); IFP / IST SL5 Task Force for the SL5 gap
RAND assesses eight of 38 weight-theft vectors as likely infeasible for OC1–OC3 but likely feasible for OC4 or OC5 under its defined capabilities8 of 38
estimateRAND 2024 expert assessment under its defined adversary capabilities; eight vectors rated likely infeasible for OC1–OC3, not a categorical impossibility statement.
—2024-05-30checking…—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.
2026-06-29: 6 → 2026-09-16: 62017checking…—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.
—2025-11checking…—not placedFingrid; Svenska kraftnät
reported tenant handoff to newly spun-up nodes after a one- or two-node failure~90 sec
observed
—2026checking…—SemiAnalysis AI Neocloud Playbook / ClusterMAX
average annualized insider-risk cost in the cited report: $19.5M across 354 organizations$19.5M
observedSurvey/benchmark population includes negligent, malicious and credential-theft incidents; not a data-center-specific loss estimate.
—2026checking…—Ponemon Institute, 2026 Cost of Insider Risks Global Report; DTEX sponsor report page. Study observations concern 2025.
average breach cost in IBM's 2025 Cost of a Data Breach report: $4.44M global and $10.22M US$4.44M global / $10.22M US
observedOne IBM/Ponemon report vintage. Do not combine 2026 values or AI-breach findings with the 2025 value, as-of date and locator.
—2025checking…—IBM/Ponemon Cost of a Data Breach 2025
average commercial rate in the top-10 US data center states (≈ identical to other states)~14.4-14.5¢/kWh
observed
—2025-26checking…—not placedElectric Choice / RealClearEnergy synthesis
average cost of data-center downtime (≈$5,000/min)$300k/hr
observed
—2025checking…—not placedUptime Institute / industry downtime analyses
average eCrime breakout time (initial access to first lateral movement) in 2025, down from 48 min in 2024; fastest 27 s29 min
observed
—2025checking…—CrowdStrike 2026 Global Threat Report
average energy savings from retro-/re-commissioning of an existing facility (often more)10–20%
observed
—2023checking…—not placedU.S. DOE FEMP / LBNL retro-commissioning studies
Average time to contain an insider incident: 67 days, down from 86 in 2023; far longer than a checkpoint copy takes67 days
observedTime to contain in the study population, not a model-theft detection budget or time to copy a checkpoint.
2026checking…—Ponemon Institute, 2026 Cost of Insider Risks Global Report; DTEX sponsor report page. Study observations concern 2025.
AWS facilities directly hit by drones (UAE) + 1 blast-damaged (Bahrain), Mar 2026 — first confirmed combat strike on US-run hyperscale DC2 + blast
observed
—2026-03checking…—DefenseScoop / DCK / MWI (West Point)
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 commitmentsnearly fully subscribed
observed
—2026-08checking…—not placedAmazon Q1 2026 earnings, Andy Jassy (2026-04-30); Q2 2026 earnings + Bezos remarks (Aug 2026): Trainium4 2027, ~$25B annualized in-house chip revenue, $225B commitments
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 GB200 provides the separate named 3,245 lb profile. Neither result is a support, rolling or uniform floor rating.
—2026-07checking…—HPE GB200 QuickSpecs V5, July 20, 2026: system dimensions and loaded-with-fluid mass; the guide range uses the stated class-mass screening assumptions.
back-end fabric cost avoided with 2:1 oversubscription (virtual-modular-switch design) vs the non-oversubscribed reference — where the workload allows (contested — single-source) · contested~31%
observedSemiAnalysis models the 31.6% saving for 2:1 oversubscription and recommends 2:1 for small clusters; the cited passage does not model ~3:1 (Meta ran 7:1 between pods).
2026-06-29: 31 → 2026-07-14: 312024checking…—SemiAnalysis, AI Neocloud Playbook and Anatomy
back-end NICs per training node = 3.2 Tb/s/node; 1:1 non-blocking 8-rail fat-tree8× 400 Gb/s
observedOne named 8-rail training reference design, not a workload-name mandate. Validate the port rate, rail count, and blocking ratio against the target traffic matrix, topology, placement, failure headroom, and step-time objective.
—2025checking…—SemiAnalysis AI Neocloud Playbook
back-end SuperNIC provisioning (8×400G/800G per 8-GPU server = 3.2–6.4 Tb/s/node) + a separate front-end/storage NIC1 per GPU
observed
2026-06-29: 1 → 2026-07-10: 12025checking…—NVIDIA RA
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 the named litigation, not proof of completed distribution, a per-work net payout, appellate finality or a general permission to train.
—2026-07-20checking…—US District Court, final approval and judgment, Document 680, 20 July 2026; official settlement-administrator court-document copy.
ISO-Base platform displacement (standard ~6 in; XD 10.25 in) and supported equipment weight (~$15-20k/unit, practitioner estimate — vendor does not publish pricing)6–10 in / up to 10,000 lb
observed
—2026checking…—not placedWorkSafe Technologies (ISO-Base / ISO-Base XD)
battery-capacity cut and double-conversion-loss cut from distributed BBU + supercaps vs central 2N~50% / 2–3%
observed
—2026checking…—SemiAnalysis; Eaton 800 VDC reference architecture
Cleanview’s mid-2026 US behind-the-meter generation announcement stock~90 GW announced
estimateTracker estimate of announced US generation capacity across 59 projects; 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.
behind-the-meter (mostly gas) announced since Jan 2025; the speed-vs-carbon tradeoff in one number~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.
2026-06-29: 82 → 2026-08-24: 822026checking…—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; other Table 4 layouts report 38% or 43%. No Hopper fleet-wide MFU band or universal acceptance threshold follows.
2026-06-29: 41 → 2026-09-16: 412024-07checking…—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 silicon34% → 54%
observedThe endpoints are rounded MFU percentages, so a throughput ratio computed from them (54/34) is approximate; quote the gain only from the source's unrounded figures.
—2024checking…—SemiAnalysis (H100 vs GB200 NVL72 Training Benchmarks, Oct 2025)
BIS moves named advanced-compute exports to China/Macau from presumption-of-denial to case-by-case reviewJan 15 2026
observed
—2026checking…—US BIS final rule; Baker McKenzie sanctions analysis
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.
—2024-10checking…—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.
—2026-09checking…—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 stepsBlackwell ~71% of 2026
forecast
—2026-04checking…—not placedTrendForce via The Register (2026-04-08)
Bloom–Oracle SOFC framework (1.2 GW initially contracted); near-zero criteria pollutants; ~$3,000–4,000/kWup to 2.8 GW
observed
—2026checking…—Bloom Energy IR / press
BlueField-4 Arm Neoverse V2 complex (64B transistors, ~6× BF-3 compute); 128 GB LPDDR5; PCIe Gen664-core
observed
—2026checking…—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 profile800 Gb/s
observedAdvertised DPU interface capability, not an end-to-end application benchmark or a guarantee that all scale-out traffic traverses this DPU.
2026checking…—NVIDIA BlueField Platform, DPU portfolio and connectivity description.
BMC runs on standby power and boots before the host; a rooted BMC is an OS-invisible, persistent foothold under the CPUalways-on
observed
—2025checking…—Eclypsium / OCP DC-SCM
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.
2026-08-24checking…—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.
—2024-10checking…—SemiAnalysis, AI Neocloud Playbook and Anatomy
Broadcom Jericho4 deep-buffer router; 3nm; HBM packet memory (~160× on-chip); RoCE over >100 km via 3.2 Tbps HyperPorts51.2 Tbps
observed
—2025checking…—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 2026volume ramp Mar 2026
observed
—2026-03checking…—not placedBroadcom press release (2026-03); Arista PR (2026-06-09)
Tomahawk 6 advertised switching capacity; family shipping in production volume102.4 Tbps
observed102.4 Tb/s chip family; 512×200G or 1,024×100G electrical options. System ports and optical variants require separate support and delivery evidence.
2026-07-02: 102.4 → 2026-09-16: 102.42026-03-12checking…—Broadcom production-volume announcement, 2026-03-12.
brownfield retrofit cost: cooling-only vs full AI retrofit; ~2/3 of pre-2015 DCs unsuitable for frontier density · contested~$2M / ~$5–6M+ per MW
observed
—2025checking…—Enabled Energy