Corrections log
Every accepted correction, public and dated — caught by readers or by the guide's own editorial and adversarial review passes. Spotted an error? Use “Suggest an edit” at the end of any chapter; accepted fixes are credited.
independent adversarial re-review (codex GPT-5.6, xhigh) of the full P0/P1/P2 remediation
A six-agent independent code review of the shipped remediation surfaced defects that were each verified against the code and fixed: the account data-export/delete endpoint was unreachable for signed-in users (missing from the auth middleware); the project-finance engine mis-handled degenerate cash-flow streams, sign changes across zero years, and non-finite inputs, and the connected-scenario adapter rounded capex/revenue so the calculators disagreed with their own scenario; the RAG grounding check validated against text the model never saw and matched numbers by substring (30 inside 2030); the defense-in-depth HTML sanitizer could be defeated by nested-tag reassembly and entity-encoded URL schemes; and several data-controls, rate-limit, and accessibility gaps. Hardware figures were corrected (B200 180 GB, GB200 186 GB, GB200/GB300 2.5 dense PFLOPS). All fixes carry regression tests; the campus financial-model workbook was re-verified cell-for-cell against the tested engine.
Chapter 0.2 and the product positioning promise decision tools, but most artifacts remain prose/tables. A practitioner should be able to carry a controlled template into a project without reverse-engineering it from paragraphs
Five practitioner artifacts now ship as versioned downloads (v1.0, /templates), each declaring intended user, required inputs and units, reference design, revision and approval fields, and a complete worked example: a campus financial model whose live Excel formulas mirror the guide's unit-tested CFADS engine (verified cell-for-cell against it, with an engine-check column that exposes formula drift); a site diligence and kill-gate checklist with documentary evidence requirements and a fully scored reference site; a commissioning issue log with level/severity conventions and two-party close-out plus a MOP/SOP/EOP skeleton containing a complete worked maintenance-bypass MOP; a design-basis and single-line assumptions template pre-filled with the reference liquid-cooled campus; and an FMEA starter with pinned rating anchors and worked failure modes. Chapters 0.2, 0.5, 2.5, 3.13, 13.2, and Appendix F link to the current versions.
current tools are independent and can silently use incompatible generations or boundaries. A connected model would embody the guide's strongest cross-domain idea
The ten calculators now derive from one shared project scenario (P2-05): a pure, unit-tested model (workload, hardware generation, facility, location, capex rates, financing, schedule) runs the guide's cross-domain chain — GPU choice → rack density → cooling regime → IT/facility MW → capex and schedule → energy/water/carbon → $/GPU-hr and $/M tokens → CFADS finance — and one derivation feeds every calculator's inputs, so a hardware-generation change propagates to density, cooling, capex, unit economics, and IRR at once. Local edits detach a field explicitly and can be reset; hardware overrides and PUE outside the cooling regime's plausible band are flagged as mixed boundaries instead of silently combined. The scenario is versioned with the defaults' as-of vintage, diffs itself against the published defaults, shares via URL fragment (never sent to a server), exports to CSV with provenance per field, and saves to an account only on an explicit Save.
the prose reads as templated — repeated antithesis, stock scaffolds ("the master fork", "The decision:", "The consequence:"), 55-word theses, and quota-like uniformity across 173 chapters
Every chapter received a corpus-wide anti-template line edit (P1-13): stock scaffolding removed, repeated "X is not Y, it is Z" antithesis varied or cut where it corrected no real misconception, slogan conclusions grounded, and each thesis compressed to a single claim plus its boundary. "Master fork/variable" fell from 77 occurrences to 5, "The fork" from 258 to 126, and the mean thesis from 55 words to 31 (chapters over 40 words: 171 to 1) — the survivors are the legitimate uses where the phrase names a real engineered choice, not deleted by quota. Every edit passed a machine gate preserving all numbers, units, ranges, dates, cross-references, and HTML, and a second-editor sample audit of 21 chapters across all Parts found zero meaning, scope, hedge, or attribution drift.
continuous replication bandwidth (~200 Gbps) to hold a 1-hour RPO: ~$50k/mo
The ~$50k/month price attached to continuous 200 Gbps cross-region replication could not be supported by any public pricing (ordinary cloud egress rates imply far more). The claim now states only the defensible arithmetic — holding a 1-hour RPO on ~100 TB of training state requires ~200 Gbps sustained — and the cost figure is withdrawn.
87 canonical figures carried no source link, derivation, or bibliography entry
Closed out the unsourced tail of the claim library: 48 figures now carry an independently verified source URL (each fetched and figure-checked before applying); 31 honest guide computations are now labeled as derivations with their arithmetic or assumption basis stated on the record; and 8 figures nobody could back were removed outright — including two more traced to the previously-disavowed source with fabricated attributions, a claimed 45–60%→70–85% Kubernetes-DRA utilization jump, a 25–40%/yr cyber-insurance escalation, a 99.99% three-region availability attribution, a per-failed-GPU NVLink degradation figure, a CPO MTBF target, and a factory-integration install-time comparison. The publish pipeline now fails — not warns — if any active figure lacks a source or stated derivation.
duplicate share of multi-snapshot Common Crawl removed by dedup before quality filtering: ~50–67%
Re-scoped to what RefinedWeb reports: fuzzy+exact deduplication removes ~50% of the already-filtered remainder (~12% of original documents), and it runs after language/quality filtering, not before. The ~50–67% figure conflated combined filtering-plus-dedup removal with duplicate share; dedup-only rates vary 7.6–60.8% across corpora.
GB200 NVL72 secondary-loop flow ~170–235 L/min
Vendor GB200 NVL72 reference designs specify lower per-rack flow at a wider temperature rise: QCT's QoolRack runs up to 130 L/min and nVent's five-rack design ~138 L/min per rack (~12 °C rise). The guide's ~170–235 L/min figure is its own derivation assuming a tighter 7–10 °C rise and remains marked as such; the claim now leads with the vendor-published range.
streaming + compressed local-SGD traffic cut: 16x vs DiLoCo / ~3000x vs DDP
Re-derived from SparseLoCo's Table 2 (message sizes × synchronization counts): ~120x less total traffic than 8-bit DiLoCo and ~3,600x less than 16-bit AdamW DDP at 0.78% density, or ~30x/~900x at 3.12% — no configuration yields 16x/~3,000x together. The claim and Chapter 8.8 now carry the derived ranges.
heat-exchanger approach temperature for liquid-to-air coils: ~3–5 °C
Corrected to ~10–15 °C nominal for current liquid-to-air CDU coils (Vertiv CDU070 specifies 15 °C; Delta's L2A specifies 10 °C; ATD varies ~7–22 °C with conditions). The ~3–5 °C figure belongs to liquid-to-liquid plate heat exchangers and remains correct in that scope (Chapters 5.6 and 5.11).
Modified Wobbe Index band for gas-turbine fuel: 40–55 MJ/m³
The 40–55 MJ/m³ band is a conventional volumetric Wobbe-index range, not a Modified Wobbe Index specification — GE's MWI is temperature-corrected and controlled within ±5% of the turbine-specific target (±1–3% practical). The claim and Chapter 4.9 now separate the two units.
PJM 2018–2020 cohort: 65–80% withdrew before an interconnection agreement; ~24% reached operation
The 65–80% PJM-wide cohort figure could not be verified (the underlying 24%/13% statistics are a Northern Illinois progression history in PJM's 2020 RTEP). Replaced with LBNL's Queued Up national statistics: only ~19% of projects (14% of capacity) requesting interconnection 2000–2018 reached commercial operation by end-2023. Chapters 3.2 and 3.13 updated.
68% of enterprise data centers built before 2015 lack the power density and cooling for modern AI workloads
The 68% figure could not be verified and traces to a fabricated Uptime Institute attribution in a previously-disavowed source. Replaced with the verifiable statistic from the cited DCD analysis: over 60% of enterprise data centers are more than seven years old — air-era stock predating AI-class density. Chapters 5.2 and 5.10 updated.
per-GPU HBM capacity: H100 80 → B300 288 GB → Rubin Ultra ~1 TB
Scoped: Rubin Ultra's ~1 TB HBM4E is a package-level figure across four GPU compute chiplets, not memory attached to one GPU die — uniformly calling the sequence per-GPU obscured a material packaging change. The claim and Chapter 7.6 now state the package-level scope (roadmap specs remain in flux).
data-center demand-response capacity Google has contracted (~1 GW, 2025)
Google announced the 1 GW demand-response milestone on March 19, 2026 (initial utility agreements date from 2025). The claim's as-of date now reads March 2026 with the primary announcement linked.
CoWoS-L reticle multiple: 5.5× (>98% yield) in volume now; 9.5× on the 2027 roadmap
Per TSMC's 2025 annual report: 3.5×-reticle CoWoS-L has been in production since 2024; 5.5× was certified in 2025 with volume production starting in 2026; 9.5× is in development (~2027 target). The claim and Chapter 7.7's table and prose now carry the staged timeline instead of describing 5.5× as already in volume.
hot/warm standby ~1.4x (~60% lower DR premium)
Fixed DR-pattern cost multipliers are workload-specific estimates, not standards — AWS's Well-Architected guidance defines the tiers (active-active, warm standby, pilot light) without fixed ratios, and published taxonomies put warm standby at ~1.4–1.6x. Chapter 12.3 now presents the multipliers as illustrative ranges.
decommissioning bond sizing in the 2025 model ordinance: 110% of net cost
The 110%-of-cost-minus-salvage security requirement with five-year re-estimates appears in the April 2026 revision of the York County model data-center ordinance, not the original August 2025 edition, which ends without a decommissioning-security section. The claim and Chapter 14.10 now date the provision to the 2026 revision.
physical build time for a phased hyperscale campus: 12–18 mo
Corrected to 18–24 months for a hyperscale campus (the cited source assigns 12–18 months to colocation-scale builds). Shell-and-core and steel-only durations of 12–18 months in Chapters 2.1 and 3.2 are narrower scopes and stand. The grid-is-the-long-pole framing is unchanged — interconnection still runs 3–8+ years.
Texas/ERCOT commercial rate ~8.7¢/kWh — among the lowest in the US on the deregulated market
Re-scoped: ~8.4¢/kWh (EIA, April 2026) is the all-in statewide average commercial retail price — a different denominator from deregulated-market supply quotes (~6.8¢/kWh, Electric Choice June 2026). The claim now states the EIA statewide average and notes the scope difference.
Virginia data-center sales-tax exemption cost ~$1.6B/yr; Senate proposed phase-out to Jan 2027
Virginia's official biennial report (RD40, 2026) puts the FY2025 exemption benefit at $1.94B (up from $1.29B in FY2024), and the Senate's proposed January 2027 phase-out was not adopted — the final 2026 budget retained the exemption and ordered a further tax-policy study. Chapters 3.1 and 3.10 now carry the reported figure and the outcome.
back-end fabric cost avoided by oversubscribing 1:1 → ~3:1 (~31%)
The SemiAnalysis estimate the figure rests on models a 31.6% saving for 2:1 oversubscription with its virtual-modular-switch design — it recommends 2:1 for small clusters and does not model ~3:1. The claim and every chapter passage now attribute the ~31% saving to a 2:1 design (the claim remains marked contested/single-source).
the register merges two independent datasets; a figure in Part 4 and the same figure in the register are not one record
The curated provenance register has been fully merged into the claim library: 46 curated entries that duplicated chapter figures now display as the single shared record (ending their double-listing on the register), 8 genuinely register-only records migrated as first-class claims, and every register row's revision history follows its claim. A further 22 formatting-variant duplicates (dash and unit-notation twins, including the interconnection lead-time and revenue-per-GW families) were merged. Appendix D's one-record promise now describes the actual implementation. 90% of the guide's canonical figures now carry a direct source link or reproducible derivation.
Tier IV carries a ~20–40% capital premium over Tier III
The widely-quoted flat figure conflated two scopes. Published evidence puts the whole-facility Tier IV premium at roughly 10–25% (Schneider Electric cost-calculator outputs ~11.5%; McKinsey implies ~25%; the Uptime-derived cost model ~9%), while duplicating the electrical/UPS topology alone runs ~40–50% and one project-specific case reached ~41%. Every surface quoting the old figure (chapters 0.5 and 13.1, Appendix C, the quiz, and the Part 0 deck) now states the scoped range.
41 flagged value conflicts from the source-verification wave
Every figure flagged during source verification as disagreeing with its located source was independently re-verified and adjudicated: 38 figures corrected — among them the CMMC Phase II certification date (suspended by DoD on July 13, 2026), ERCOT data-center load (~28 GW is the 2028 waypoint; ~40 GW belongs to 2030), B200 usable HBM (180 GB), base-isolator design displacement (~8 in), China's hub green-electricity target (end-2025, not 2030), global data-center water consumption (~703B gallons for 2025), and a grid-forming battery black-start described as demonstrated that is in fact planned. One unsupported figure (a GPU-recovery time its cited study never reports) was removed outright.
no overwrite-based method satisfies the Purge threshold for SSD/NVMe
Media-sanitization guidance corrected: IEEE 2883-2022 and NIST SP 800-88 Rev. 2 accept verified device-level sanitize operations — cryptographic erase, block erase, and the drive's own Sanitize-overwrite command — as Purge. What fails the threshold is host-software overwrite, which cannot reach overprovisioned space. The guide previously said only cryptographic erase or destruction qualified.
a mission-critical data center is typically designed to Risk Category IV
Corrected the seismic and structural classification guidance: per NIST GCR 23-037, most data centers are code-assigned Risk Category II — Risk Category IV applies only when a facility is designated for critical functions, and the Ip = 1.5 component importance factor follows that designation (or life-safety scope), not data centers as a class. Chapters 6.3 and 6.7 now state the code reality and frame Category IV as the elective design choice mission-critical owners routinely make.
Crossing the cliff in a retrofit costs $5–10M/MW and still leaves stranded capacity
The flat $5–10M/MW figure for retrofitting an air-cooled hall to AI liquid cooling could not be traced to either of its stated sources and originated from a source the guide had already ruled out. It has been replaced everywhere it appeared (chapters 0.1, 1.1, 1.6, 1.7, 5.1, 6.1 and Appendix C) with the verified tiered picture: a cooling-only retrofit of a power-suitable hall runs roughly $2M/MW (STL Partners, 2026), while a full conversion to AI density runs roughly $5–6M/MW and approaches greenfield cost at 100 kW+ racks, where power distribution — not cooling — becomes the binding constraint (Enabled Energy; Uptime Institute). The still-strands-capacity caveat was well supported and stands.
labels are not locators; only 27 of 1,383 key numbers carried a direct source link; auxiliary surfaces copy strings
Half of the guide's canonical figures (524 of 1,041) now resolve to a direct source link or a reproducible derivation: 459 were linked to the curated bibliography, 31 more received individually web-verified source URLs, and figures on chapter pages now link their source directly. The comparison pages, regional guides, lead-time tracker, and its embed now render figures from the same claim records as the chapters — a publish-time check fails the release if any of them ever references a stale or superseded figure. Two mis-attributions found during verification (a liquid-cooling market-share figure credited to the wrong analyst house, and a mis-cited power-oversubscription paper) were corrected.
Recurring checkpoint, Llama-goodput, TEE and TCO figures are copied across Part 9–11; repeated key numbers already drift in meaning across canonical homes
Reviewed all 155 groups of figures that repeated the same value across chapters: 328 duplicate records were merged into 130 canonical claims (each chapter keeps its own framing but now renders the one shared record), so a repeated figure can no longer drift into different vintages or attributions — among them the training-goodput benchmark quoted in 27 chapters, the July 2024 Virginia load-loss event quoted in 18, and the H100 cluster MTBF quoted in 14. Same-number coincidences and genuinely different scopes were kept as separate figures. Two mis-citations found during the review were corrected by the merge.
the sentinel selects by an old four-digit year, ignores fast-changing same-year claims, excludes contested claims, checks at most five, and does not advance a last_checked state
The figure freshness checker now schedules reviews by how fast each figure actually moves — prices and lead times monthly, standards yearly — instead of only flagging figures with an old calendar year. Every check is recorded and advances the schedule, contested figures are checked like any other, and a verified change is drafted with its full source (including the URL) so the citation survives all the way to the published page.
labels are identities, repeated values drift, auxiliary surfaces copy strings, history midpoint-parses ranges, and sources/key numbers are disconnected
Every key-number figure in the guide (1,370) is now a stable claim record with a permanent id, structured value or range, vintage, and source fields; chapters render claims by reference, so one correction updates every occurrence, the register, decks, and exports in the same release. The public register and register.json now carry the claim id and source link, /claims.json serves the full claim library, and a figure's revision history follows its claim id — rewording a label no longer resets it. Range values are tracked as ranges; the register no longer charts invented midpoints.
datePublished: H2 2027 (announced)
Chapter pages were emitting claim vintages (strings like 'H2 2027 (announced)') as article publication dates in structured data. Chapters now carry real editorial dates — first publication and last content change, maintained automatically by the publish pipeline — used in the page metadata, markdown renditions, citations, and sitemap. The synthesized FAQ structured data was removed (the decision lists were never a visible FAQ, and Google retired FAQ rich results in mid-2026).
nothing you enter leaves your browser
Made the calculator privacy copy literally true: calculations run locally, share links keep inputs in the URL fragment (never sent to the server — share links were also moved from query strings to fragments), and only an explicit signed-in Save uploads a scenario.
conductivity floor
Corrected the flush acceptance terminology: ≤5 µS/cm is a ceiling, not a floor, and the 1.5× hydrostatic hold applies to the isolated pipe scope per the governing code — assembled-loop components (CDU, hoses, quick-disconnects) are proven to their own MAWP-derived limits.
compare it against the ~$8.5M/MW-year all-in TCO of a self-build
Fixed an incompatible-scope comparison: colo lease rates (~$1.8–2.6M/MW-yr) buy the powered shell only, while the self-build figure includes the IT fleet — the like-for-like comparison requires adding the tenant's own IT, power, and operations to the lease case. Appendix B's per-GPU ladder now names each scope ($43k rack hardware → $45–55k all-in server → $60–70k with cluster share).
typical seismic Risk Category for mission-critical data centers: RC III / IV
Clarified that code assigns a typical data center Risk Category II; III/IV seismic performance is an owner election, not a code mandate. Also noted that a retired power plant's interconnection rights do not automatically convey with the land.
your Phase I/II ESA exit cleanly severs CERCLA exposure
Corrected the CERCLA framing in 14.10 and 3.8: All Appropriate Inquiries is the threshold of the purchaser defenses (which also require continuing obligations after closing), and a state No Further Action letter is the strongest closeout available but not an absolute release — it does not automatically bind federal authorities and carries reopeners.
A facility that smooths its electrical transient also smooths the thermal one
Corrected the BESS thermal claim: storage changes where the electrons come from, not how much heat the silicon makes — only a power cap flattens the thermal waveform.
the die temperature tracks coolant supply temperature almost rigidly
Rebuilt the thermal-transient physics: the die rides a power-dependent offset above coolant temperature, and a heat step at maintained flow is buffered by the loop's fluid and metal (a 115 kW step into ~200 L warms the inventory only ~1.4 °C in ten seconds) — the transient with no inertia is loss of flow, a redundancy problem, not a controls one. 'No thermal inertia' claims in 5.11, 12.2, and Appendix F are now scoped to flow loss.
a faster, firmer path through the interconnection queue
Corrected the large-load tariff framing: a tariff sets service terms and cost allocation; interconnection studies and physical upgrades still set the timeline.
a clean, isolated AC output
Clarified that double-conversion (VFI) UPS decouples the output waveform but does not by itself provide galvanic isolation — that still requires a transformer — and that the July 2024 event was grid-to-load coupling (ride-through), not caused by GPU load transients.
a single 230 kV fault
Corrected the July 2024 Northern Virginia incident across eight chapters and Appendix F: per NERC's incident review it was a six-fault, 82-second 230 kV disturbance sequence initiated by a failed surge arrester (~1,500 MW of data-center load lost), not a single fault.
PUE 1.1 at 40% MFU vs PUE 1.2 at 90% MFU
Replaced the implausible 90%-MFU comparison with realistic values and the governing relation: useful output per facility MWh scales with (active duty × MFU × goodput) ÷ PUE.
Delta, 1,168 GPUs
Updated the NCSA Delta resilience study citation to its current version: 1,056 GPUs, 11.7 million GPU-hours over 2.5 years (arXiv 2503.11901).
800 VDC moves 157% more power through the same copper cross-section
Scoped NVIDIA's 157% claim to its actual comparison basis (three-wire DC feed vs four-conductor 415 VAC, Oct 2025 article; the May 2025 article gave ~85% on a different basis) and rewrote the quiz to teach the basis difference instead of number recall.
Kyber NVL576 rack — 144 quad-die packages = 576 GPU compute dies
Updated the Rubin Ultra topology to NVIDIA's March 2026 architecture: NVL576 is an eight-rack MGX system (576 GPUs in one optical NVLink domain); the Kyber rack is NVL144 (144 quad-die packages = 576 dies per rack), with eight Kyber racks forming NVL1152. Appendix B's scalable-unit table was standardized on package counting.
reliability overhead as a deliberate 6–21% slice of TCO
Recast the '6–21% of TCO' figure across fifteen chapters as what SemiAnalysis actually measured: a reliability tax — the share of paid GPU-hours lost to failures and recovery, ranging ~6–21% by provider — not a deliberate TCO spending line.
fluoronitrile mix … compliant under current rules
Corrected the EU F-gas table: Regulation 2024/573 Art. 13(9) bans putting new F-gas-insulated switchgear into operation by voltage band and date (≤24 kV 2026; 24–52 kV 2030; GWP≥1 HV bands 2028/2032) — fluoronitrile blends are F-gases and are caught absent a derogation.
reporting at low load [flatters PUE]
Corrected the direction of PUE at partial load: fixed overhead makes real PUE worse as IT load falls, not better.
warm-water DLC (45–60 °C+) often does not [need a heat-pump lift]
Corrected the heat-reuse thermodynamics: a 45–60 °C return cannot supply a 60–70 °C (3G) network without a heat pump — it feeds low-temperature (4G/5G) networks directly and 3G networks with a smaller lift.
PUE category-1 (annualized, fully sub-metered)
Corrected reversed ISO/IEC 30134-2 PUE categories: Category 3 is the advanced IT-equipment-input measurement; Category 1 is the basic UPS-output figure.
NIST SP 800-88 Rev.1
Updated media-sanitization references to NIST SP 800-88 Revision 2 (final September 2025, superseding Rev. 1).
most high-risk-system obligations apply from 2 August 2026
Updated the EU AI Act timeline for the 2026 omnibus: stand-alone (Annex III) high-risk systems comply from 2 December 2027 and product-embedded systems from 2 August 2028 (Council approval 29 June 2026). Also removed an unreproducible ISO 42001 RFP-share statistic.
GB200 ~80 L/min flow, ~2.4 MW cooling capacity
Replaced the mixed GB200 cooling record with two named OEM profiles: OCP MGX Rack Rev. 1.1 (120 kW, ~85% liquid capture, up to 130 L/min) and HPE GB200 NVL72 (132 kW nominal, 115/17 kW split, 192 kW busway, 1.3 MW in-row CDU for up to 8 racks). Per-rack flow ranges are now labeled as guide derivations.
Tier III at 99.982%, Tier IV at 99.995%
Removed Tier-to-availability-percentage mappings across chapters, quiz, comparison, calculator, and decks — Uptime disavowed those figures; Tier certifies topology properties (concurrent maintainability, fault tolerance). Historical note retained in Chapters 0.4 and 12.1.
inducing a fault on the path that should still be carrying load
Removed the unsafe 2N IST instruction to improvise a second fault on the live surviving path; the test matrix now comes from the facility FMEA with defined preconditions, instrumentation, and abort criteria.
the BESS sinks the pulse peaks and sources the troughs
Reversed BESS power-flow direction corrected: to flatten grid draw the battery discharges into load peaks and recharges in troughs. The unsafe 'commissioned against a real induced grid fault' language was replaced with controlled methods (relay secondary injection, hardware-in-the-loop, planned breaker-opening tests against protected load banks).
2(N+1) — sometimes written 2N+1
Corrected redundancy notation: 2(N+1) and 2N+1 are different topologies (at N=3: eight units vs seven). Also distinguished active/active from active/standby 2N operation.
a rack busbar rated to roughly 1,400A
Corrected NVL72 busbar current: 132 kW at 48 V is ~2,750 A — ~1,400 A is the per-bar rating of the paired busbars (~2,800 A aggregate), consistent with Chapter 7.13.
~266 ML/yr water make-up
Corrected Appendix B's water arithmetic: WUE is defined per kWh of IT energy, so 50.7 MW IT × 0.5 L/kWh × 8,760 h = ~222 ML/yr (the table had applied WUE to facility energy).
~$9,000/min for cascading incidents
Fixed a unit error: $300,000/hour is $5,000/minute. The inconsistent $9,000/min figure was removed.
a hardware fault every few minutes around the clock
Corrected the 100k-GPU fault-cadence extrapolation: 7 days per 512 GPUs scales linearly to roughly one fault every ~52 minutes, not every few minutes.
together nearly 59% of all failures were GPU/HBM-related
Corrected the Llama 3 failure arithmetic: faulty GPUs (30.1%) plus HBM3 (17.2%) are 47.3% — the ~58.7% figure is Meta's full GPU-related grouping, now stated separately.
Busway / busduct lead-time floor eased: ~30–52 wk → ~27–52 wk (VAWN electrical lead-time index, June 2026: standard runs average ~27 wk; custom engineered runs at the top). Same 2026-07-12 fact-check.
H100/H200 GPU lead time re-stated: 6–12 mo → ~3–12 mo, bimodal — spot H100s improving to weeks while H200 and full enterprise systems sit at the top. Same 2026-07-12 fact-check.
800G optical transceiver lead-time floor raised: ~20–50 wk → ~36–50 wk. The 20-week floor predated NVIDIA's March 2026 $4B EML pre-buy, which pushed non-priority allocations out. Same 2026-07-12 fact-check.
HV switchgear & circuit-breaker lead time corrected: ~45–80 wk → ~100–155 wk. The old headline figure described MV-class gear; true HV breakers run ~100–155 wk (Duke Nicholas Institute / Wood Mackenzie). Prose aligned in Ch 2.1 and 2.3. Caught by the 2026-07-12 adversarial fact-check of the Lead-Time Tracker.
Project-finance calculator: the IRR solver could return a misleading result for negative-input edge cases; inputs are now clamped and the failure mode returns “—” instead of a wrong number. Dual adversarial review.
Dispatches Atom feed: entry ids were improperly escaped and the feed lacked an RFC-4287 author element; both fixed. Dual adversarial review.
Glossary: “$”-prefixed terms (e.g. $/GPU-hr) matched with wrong word boundaries, creating spurious in-text term links. Dual adversarial review.
Lead-Time Tracker: the register query missed several lead-time figures (regex under-match) and could list the same figure family twice; matching and number-family dedupe fixed. Dual adversarial review.
Rear-door heat-exchanger ceiling corrected to ~50–100 kW per rack in chapters that had drifted to lower values — Chapter 5.3 holds the canonical derivation. Caught in the dual adversarial model review (Claude Fable + GPT-5.5 high).
Every figure traced to one vendor blog was removed or re-sourced after the source failed credibility review — the register now carries the replacement citations. Editor-in-chief review pass.
Behind-the-meter gas fleet sizing corrected to the ~1 GW-class canon for the flagship deployments discussed; prior text overstated it. Editor-in-chief review pass.
NVL72-class rack weight standardized to the ~3,000 lb canon; earlier chapters carried divergent figures. Editor-in-chief review pass.
Warm-water direct-to-chip loop temperatures standardized to the 45 °C supply / 65 °C return canon; several chapters carried inconsistent pairs. Editor-in-chief review pass (446 fixes across the guide).