What is the adopted GB300 IT load per hall?

Adopted GB300 basis: 880 NVL72 racks per hall x 142 kW per rack = 124,960 kW rack IT per hall (499.84 MW facility rack IT). One rack IS one NVLink domain at GB300 — there is no split-domain footprint. Total IT including fabric switches, OOB and storage/mgmt electrical load is 134,763.2 kW per hall (539.05 MW facility) — this total-IT figure, not the rack-only figure, is the PUE denominator.

Why is the PUE 1.165 and not lower?

The PUE is derived bottom-up: chiller electrical input comes from a Carnot-fraction COP over the actual evaporator/condenser lift (never a nameplate figure), plus measured pump, fan, UPS-loss, distribution-loss and auxiliary terms. At the design-day ambient the largest non-IT term is the chillers (~25,811 kWe facility, ~29% of non-IT). Facility load / Total IT = 628,087 / 539,053 = PUE 1.165 design-day (1.158 annual bin-weighted, 1.250 at the 36 C worst bin). The 1.12 target is a design aspiration; the engine reports the +0.045 gap rather than closing it.

How is CDU count sized at GB300?

CDU_running = ceil(liquid_heat_kW_per_hall / CDU_unit_kWth). Liquid heat per hall = 124,960 kW rack IT x 0.85 = 106,216 kW. CoolIT CHx1000 units are rated 1,000 kWth each, so 107 run per hall + 1 standby = 108 installed per hall, x 4 halls = 432 installed facility-wide.

AI/HPC Data Hall — NVL72 Cockpit — Technical Manual & Methodology

01 Purpose & engineering basis

The AI Data Hall Cockpit is a deterministic digital-twin dashboard for a GB300 NVL72 HPC campus operating direct liquid cooling. It answers five engineering questions: what is the IT load per hall and per facility, how is heat split between liquid and air paths, what hydronic flow and how many CDUs are required, what is the honest bottom-up PUE across the year, and what does the electrical distribution look like under the adopted load.

Basis of design is adopted, not locked-and-forgotten: the DCAI_MODEL object holds authored leaves only — every number carries a // source: line and an evidence class — and is deep-frozen at page load; no tab can mutate it. Every KPI on every tab reads from the same DCAI_CALC.snapshot — Rule 1 (one source of truth) of standarization/ACCURACY_VALIDATION.md. The governing hardware reference is the NVIDIA GB300 NVL72 specification (one rack = one 72-GPU NVLink domain, 142 kW/rack); thermal and facility methodology follows ASHRAE A2 and ISO/IEC 30134-2 (PUE) and ISO/IEC 30134-8 (CUE_IT).

02 Inputs — ADOPTED basis + tab selection

The primary input block is the deep-frozen DCAI_MODEL — authored leaves only, spec gb300-500mw-2026-09-06, adopted 2026-09-06. Tab-selection controls (view mode, hall selection) change which derived quantities are displayed but do not alter the underlying constants. All values below are published in data/dcai-parameters.json and verified by tools/test-dcai-engine.mjs.

Parameter	Symbol	Value (adopted GB300)	Source
Facility data halls	`facility.halls`	4	js/dcai-model.js (ADOPTED)
NVL72 racks per hall	`compute.racks_per_hall`	880	data/dcai-parameters.json
Racks per NVL72 domain	`compute.racks_per_nvl72_domain`	1	NVIDIA GB300 NVL72 spec — one rack IS one NVLink domain, no split-domain footprint any more
Racks per facility	`compute.racks_facility`	3,520	data/dcai-parameters.json
GPU per rack / CPU (Grace) per rack	`compute.gpu_per_rack / compute.cpu_per_rack`	72 / 36	NVIDIA GB300 NVL72 spec
GPU per facility	`compute.gpu_facility`	253,440	data/dcai-parameters.json
NVSwitch per rack	`compute.nvswitch_per_rack`	18	NVIDIA GB300 NVL72 spec
IT load per rack (ADOPTED)	`power.rack_it_kw`	142 kW	NVIDIA GB300 NVL72 spec (PUBLISHED)
Rack IT per hall	`power.rack_it_hall_kwe`	124,960 kW	880 × 142
Rack IT per facility	`power.rack_it_facility_mw`	499.84 MW	124,960 × 4 / 1000
Nameplate IT label	`power.nameplate_it_mw_label`	500 (LABEL only — never a denominator)	Owner decision, plan cheerful-cuddling-mitten.md
Total IT per hall (rack + fabric + OOB + storage/mgmt)	`power.total_it_hall_kwe`	134,763.2 kW	data/dcai-parameters.json
Total IT per facility (the PUE denominator)	`power.total_it_mw`	539.05 MW	data/dcai-parameters.json
Liquid capture ratio	`heat.liquid_capture_ratio`	0.85	ADOPTED, direct liquid cooling
TCS supply / return	`design.planes.p07_tcs_supply_c / p08_tcs_return_c`	40 / 50 °C	ADOPTED — warm, dry-only heat rejection
Water density	`rhoKgPerL`	1.0 kg/L	STANDARD physical constant
Water specific heat	`cpKjPerKgK`	4.186 kJ/kg·K	STANDARD physical constant
LV voltage (line-line)	`distribution.voltage_ll_v`	400 V	data/dcai-parameters.json
Power factor	`distribution.power_factor`	0.96	data/dcai-parameters.json
CDU model / unit rating	`equipment.cdu_model / equipment.cdu_unit_kwth`	CoolIT CHx1000, 1,000 kWth	PUBLISHED vendor rating
Chiller unit / COP basis	`equipment.chiller_unit_kwth / design.planes.p18_cop_air_path`	4,000 kWth; COP derived from a Carnot fraction over the actual lift, air-path COP ≈ 5.50 at design (NEVER a nameplate figure)	data/dcai-parameters.json, js/dcai-engine.js copFromLift()
UPS frame rating	`equipment.ups_frame_kw`	1,250 kW (2N)	data/dcai-parameters.json
Transformer unit rating	`equipment.transformer_unit_mva`	2.5 MVA	data/dcai-parameters.json
CRAH unit rating	`equipment.crah_unit_kwth`	200 kWth	data/dcai-parameters.json
PUE design band (target)	`pue.band_min / pue.band_max`	1.12 – 1.25 (target 1.12)	data/dcai-parameters.json
Data hall geometry (L × W × H)	`geometry.hall_length_m / width_m / height_m`	62 × 31 × 5.5 m = 1,922 m²	data/dcai-parameters.json

Retired basis (GB200): this page was originally built around a GB200 NVL72 split-domain basis-of-design (BASELINE-DECISION.md, locked 2026-05-17): 27 NVL72 domains per hall, 2 rack-positions per domain at 66 kW each = 132 kW per domain, 54 rack-positions per hall, 14.256 MW IT per facility (Scenario A), CDU sized at 350 kW end-of-row units, and a nameplate chiller COP of 6.8. That basis was retired 2026-09-06 by owner decision (plan cheerful-cuddling-mitten.md) in favour of the GB300 NVL72 basis documented in this section — one rack IS one NVLink domain at 142 kW/rack, with no split-domain construct at all. The retired files, js/datahall-model.js and js/datahall-calculations.js, remain on disk byte-frozen and still pass their own 57 worked-example tests (tools/test-datahall-calc.mjs) as the retirement record — they are historical, not the current basis, and are the only place on this page where GB200 numbers describe anything other than a retired reference.

03 Calculation methodology

All formulas below are implemented verbatim in js/dcai-engine.js. Function names are in parentheses. Source references cite the module's own header comments and tools/test-dcai-engine.mjs, which asserts the balance identities below on every run.

Rack and IT load — `compute()`

Rack_IT_kWe = racks_per_hall × rack_it_kw Total_IT_kWe = Rack_IT_kWe + fabric_kWe + OOB_kWe + storage/mgmt_kWe Adopted basis: 880 × 142 = 124,960 kW rack IT per hall (499.84 MW facility). Total IT per hall adds fabric-switch, out-of-band and storage/management electrical load: 124,960 + 7,040 + 264 + 2,499.2 = 134,763.2 kW/hall (539.05 MW facility) — this total, not the rack-only figure, is the PUE denominator.js/dcai-engine.js compute() · tools/test-dcai-engine.mjs

Rack density — `racks_per_nvl72_domain`

1 NVL72 rack = 1 NVLink domain = 72 GPU + 36 Grace CPU + 18 NVSwitch The adopted basis retires the split-domain footprint entirely (see §02 "Retired basis" note): racks_per_nvl72_domain = 1. There is no "2 rack-positions per domain" arithmetic any more.NVIDIA GB300 NVL72 spec

Facility load — `compute()`

Facility_kWe = Total_IT_kWe + cooling_kWe + upsLoss_kWe + distLoss_kWe + aux_kWe In compute(m), published as design.electrical.facility_kwe. The denominator for PUE and CUE_IT.js/dcai-engine.js compute()

PUE — `op.pue`

PUE = Facility_kWe / Total_IT_kWe Computed once per weather bin inside operatingPoint(). ISO/IEC 30134-2 definition. Design-day 1.165, annual bin-weighted 1.158, worst-bin (36 °C) 1.250 — all DERIVED (cyan chip), never the TARGET (1.12–1.25 band). Rule 4: marketing target ≠ derived value.js/dcai-engine.js operatingPoint() · ISO/IEC 30134-2

Liquid heat / air heat — `compute()`

Liquid_heat_kWth = Rack_IT_kWe × liquid_capture_ratio (0.85 direct liquid) Air_heat_kWth = Rack_IT_kWe × (1 − liquid_capture_ratio) + fabric + OOB + storage/mgmt + UPS-loss + dist-loss + aux Per hall: 124,960 × 0.85 = 106,216 kW liquid; the residual air path is NOT a simple 15% split — it also carries the electrical load that never enters the liquid loop (fabric switches, OOB, storage/mgmt, UPS loss, distribution loss, auxiliary), which is why it publishes as 35,509 kW air per hall rather than the naive 18,744 kW.js/dcai-engine.js compute() line ~311

Hydronic TCS/CDU flow — `flowM3h()`

Flow_m³/h = Q_kWth × 3600 / (ρ_kg/m³ × Cp_kJ/kg·K × ΔT_K) In flowM3h(qKwTh, rhoKgPerM3, cpKjPerKgK, deltaTK). Q is the LIQUID CAPTURED heat (85% of rack IT), not full IT load — the UI states this basis explicitly. TCS runs warm (40/50 °C supply/return) and dry-only, which is what keeps WUE truly 0.00 rather than merely unmeasured.js/dcai-engine.js flowM3h() · design.flows.tcs_m3h

CDU count — `ceilCount()`

CDU_duty = ceil( Liquid_heat_kWth_per_hall / CDU_unit_kWth ) CDU_installed = CDU_duty + standby In ceilCount(duty, unit). CoolIT CHx1000 units are rated 1,000 kWth. ceil(106,216 / 1,000) = 107 CDUs running per hall + 1 standby = 108 installed per hall, × 4 halls = 432 installed facility-wide — N+3 style single-unit standby, not an N+1 block.js/dcai-engine.js ceilCount() · equipment.cdu_installed_facility

CRAH residual cooling — `airM3s()`

perCRAH_kWth = Air_heat_kWth / CRAH_duty_units Flow_m³/s = Q_kWth / (ρ_air × Cp_air × ΔT_K) In airM3s(qKwTh, rhoAir, cpAir, deltaTK). 178 CRAH duty + 1 standby = 179 installed per hall (200 kWth/unit), 716 facility.js/dcai-engine.js airM3s() · equipment.crah_installed_facility

Chiller COP — `copFromLift()`

Carnot = (evapC + 273.15) / (condC − evapC) COP = min(copMax, carnotFraction × Carnot) In copFromLift(evapC, condC, carnotFraction, copMax). There is no nameplate COP anywhere in this engine — the ~5.50 air-path COP at the design ambient FALLS OUT of the actual evaporator/condenser temperature lift and a Carnot fraction (~0.30–0.32), and moves with ambient across the year (worse at the 36 °C worst bin). This COP is NEVER calibrated or back-solved to hit a target PUE.js/dcai-engine.js copFromLift() · design.planes.p18_cop_air_path

Bottom-up PUE across the weather year — `compute()`

for each weather bin: op = operatingPoint(model, bin.ambientDbC, fixed) annual_bin_weighted = Σ(facility_kWh) / Σ(IT_kWh) across 8,760 h free_cooling_cliff_c = tcsSupplyC − cduApproachK − dryCoolerApproachK In compute(m). The liquid path free-cools whenever ambient + dry-cooler approach ≤ TCS supply − CDU approach; below the 34 °C cliff the chiller is bypassed on the liquid side. The adopted design-day ambient sits ON that cliff (0.0 K margin). Design-day PUE 1.165, annual bin-weighted 1.158 (98.2% of hours free-cool), worst-bin (36 °C) 1.250, target 1.12, gap +0.045 — reported, never closed by tuning an input.js/dcai-engine.js compute() · pue.free_cooling_cliff_ambient_c · pue.gap_to_target

Electrical — current, kVA, busway

I_A = kW × 1000 / (√3 × V_LL × PF) [required current] kVA = kW / PF [apparent power] group_kW = racks_per_group × rack_it_kw [RPP group load] V = 400 V, PF = 0.96. One RPP group = 22 racks × 142 kW = 3,124 kW ≈ 4,697 A, carried on a 5,000 A busway trunk (93.9% loaded) — 40 groups per hall (4 per row × 10 rows), 80 RPP per hall.js/dcai-engine.js compute() · distribution.group_kw · distribution.busway_trunk_a

Room volume & density — `hall_volume_m3`, `it_density_kw_per_m2`

Volume_m³ = length_m × width_m × height_m Density_kW/m² = Rack_IT_kWe_per_hall / hall_area_m2 Geometry: 62 × 31 × 5.5 = 10,571 m³ per hall (1,922 m² floor area). Gross IT density = 124,960 / 1,922 = 65.0 kW/m². Fire-agent sizing (if implemented) requires further inputs per NFPA 2001 — room volume alone is insufficient.data/dcai-parameters.json geometry.*

04 Constants & data sources

Every constant lives in the deep-frozen DCAI_MODEL with a // source: comment and an evidence class (PUBLISHED / ADOPTED / ASSUMED / STANDARD / LABEL). No economically-material literal lives inside a function body. Key hardware constants:

Equipment	Model / Reference	Rating / Value	Source
UPS	1.25 MW modular frame, 2N topology	1,250 kW per frame; 135 frames/hall/feed, 1,080 total	data/dcai-parameters.json equipment.ups_frame_kw
Transformer	Cast-resin distribution transformer	2.5 MVA; 33 per hall per feed, 262 facility total	data/dcai-parameters.json equipment.transformer_unit_mva
Busway	RPP-group trunk, 5,000 A rated	~4,697 A required per 22-rack group (93.9% loaded)	data/dcai-parameters.json distribution.busway_trunk_a
Generator	4 MW class MV diesel generator (generic — no specific model is named in the current basis)	4,000 kW/unit; 169 running (N+2), 171 installed	data/dcai-parameters.json equipment.generator_model
CDU	CoolIT CHx1000 (row CDU)	1,000 kWth; 107 running + 1 standby = 108/hall, 432 facility	data/dcai-parameters.json equipment.cdu_model
Chiller	Water-cooled centrifugal, COP derived (not nameplate)	4,000 kWth/unit; 36 running design day, 142 worst bin, 143 installed	data/dcai-parameters.json equipment.chiller_unit_kwth
CRAH	Air-side residual-heat unit	200 kWth/unit; 178 running + 1 standby = 179/hall, 716 facility	data/dcai-parameters.json equipment.crah_unit_kwth
Dry coolers	Facility dry-cooling loop (no evaporative term — WUE stays 0.00)	1,000 kW/unit; 593 running design, 631 installed	data/dcai-parameters.json equipment.dry_cooler_unit_kwth

Facility electrical (design-day operating point)	kWe (facility, 4 halls)	Share of non-IT
Chillers (largest non-IT term)	25,811	29.0%
Dry-cooler / cooling-tower fans	15,925	17.9%
UPS loss	19,551	22.0%
Pumps (TCS + HTW + CHW + CDW)	14,824	16.6%
CRAH fans	4,626	5.2%
Distribution loss	6,547	7.4%
Auxiliary (lighting, controls, security)	1,749	2.0%
Total non-IT (design-day)	89,034

05 Outputs

Output	Engine function	Unit	Basis chip	Interpretation
Rack IT load per hall / facility	`snapshot.power.rack_it_hall_kwe`	kW	ADOPTED	880 × 142 kW = 124,960 kW/hall (499.84 MW facility). Never randomised.
Total IT per hall / facility (PUE denominator)	`snapshot.power.total_it_hall_kwe`	kW	DERIVED	134,763.2 kW/hall (539.05 MW facility) — rack IT + fabric + OOB + storage/mgmt.
Liquid heat	`snapshot.heat.liquid_hall_kwth`	kW	DERIVED	Rack IT × 0.85; drives CDU count and TCS flow.
Residual air heat	`snapshot.heat.air_hall_kwth`	kW	DERIVED	Rack IT × 0.15 plus fabric/OOB/storage/UPS-loss/dist-loss/aux; drives CRAH sizing.
TCS total flow	`snapshot.design.flows.tcs_m3h`	m³/h	DERIVED	Based on liquid captured heat (85%), not full rack heat — label required.
CDU running / installed count	`snapshot.equipment.cdu_duty_per_hall / cdu_installed_per_hall`	count	DERIVED	ceil(liquid_kWth / 1,000). 107 running + 1 standby = 108 installed/hall, 432 facility.
CRAH flow per unit	`snapshot.design.flows.crah_air_m3s`	m³/s	DERIVED	Per active CRAH; 178 running + 1 standby = 179 installed/hall.
Busway trunk loading	`snapshot.distribution.busway_loading_pct`	%	DERIVED	~4,697 A required per 22-rack RPP group on a 5,000 A trunk = 93.9% loaded.
Transformer apparent power	`snapshot.equipment.facility_kva`	kVA	DERIVED	Total IT_kW / PF; 262 × 2.5 MVA transformers facility-wide.
UPS loading %	`snapshot.equipment.ups_loading_pct`	%	DERIVED	~79.9% of the 2N 1,250 kW frame fleet — plausible with margin.
PUE (bottom-up, design/annual/worst)	`snapshot.pue.design_day / annual_bin_weighted / worst_bin`	—	DERIVED	1.165 / 1.158 / 1.250. Shows the honest derived value across the weather year, never the target.
Free-cooling cliff	`snapshot.pue.free_cooling_cliff_ambient_c`	°C	DERIVED	34 °C — the adopted design day sits ON the cliff (0.0 K margin).
Data hall volume	`snapshot.geometry.hall_volume_m3`	m³	ADOPTED	62 × 31 × 5.5 = 10,571 m³ per hall.
Gross IT floor density	`snapshot.geometry.it_density_kw_per_m2`	kW/m²	DERIVED	124,960 kW / 1,922 m² = 65.0 kW/m²; 880 NVL72 racks at 142 kW each.
Retired platform reference	`RZDataHallRackDensity.studyReference('gb200-legacy')`	rack / kW / MW / kW/m²	RETIRED REFERENCE	The pre-2026-09-06 split-domain footprint on its own smaller hall geometry — see §02 "Retired basis" note for the figures. Historical only, cannot alter the adopted baseline.

Release acceptance	Executable evidence	Required invariant
Alarm query and first-out	`tools/test-datahall-ai-alarm-query.mjs`	All filter families compose; first-out is earliest per incident; export follows the filtered set.
Rack-density profiles	`tools/test-datahall-ai-rack-density.mjs`	Adopted basis is derived from the model leaves; the retired basis (see §02) remains a named reference on its own geometry and cannot change the baseline.
Electrical semantic truth	`tools/test-datahall-ai-electrical-topology.mjs`	Normal/fault scenarios reconcile all 880 racks per hall and animation consumes semantic state.
Fire cause/effect	`tools/test-datahall-ai-fire-cause-effect.mjs`	Zoned commands, feedback, inhibition and reset authority validate; BMS/DCIM is monitor-only.
Shared Design Studio	`tools/test-rz-design-studio.mjs`	Current/current-plus-study registration, validated issue metadata and accessible dialog lifecycle remain intact.
Operator UI integration	`tools/test-datahall-ai-operator-ui.mjs` plus browser journey	Controls bind to the pure modules, dialogs remain accessible, and no document overflow or console error is introduced.

06 Worked example — adopted GB300 basis

Numbers reproduce the engine output exactly for the adopted GB300 basis (142 kW/rack, one rack = one NVLink domain). Examples 1–10 are illustrative walk-throughs consistent with the balance identities asserted by tools/test-dcai-engine.mjs. Below is a compressed trace of ten examples.

Ex1 — Rack IT load & count: 880 NVL72 racks/hall (one rack = one NVLink domain, no split footprint). 880 × 142 kW = Rack IT = 124,960 kW/hall = 499.84 MW facility.
Ex2 — Total IT (PUE denominator): 124,960 + 7,040 fabric + 264 OOB + 2,499.2 storage/mgmt = 134,763.2 kW/hall = 539.05 MW facility. The 500 MW figure is a LABEL on the rack-only envelope, never this denominator.
Ex3 — Cooling split: Liquid = 124,960 × 0.85 = 106,216 kW/hall. Air = 124,960 × 0.15 + fabric + OOB + storage/mgmt + UPS-loss + dist-loss + aux = 35,509 kW/hall — not a naive 15% split, because everything outside the rack that never enters the liquid loop rides the air path.
Ex4 — TCS flow: Facility flow 38,449.2 m³/h ÷ 4 halls = 9,612.3 m³/h/hall. Flow basis = liquid captured heat (85% of rack IT), not full IT load; TCS runs warm and dry-only (40/50 °C), consistent with WUE 0.00.
Ex5 — CDU count: ceil(106,216 / 1,000) = 107 CoolIT CHx1000 running/hall + 1 standby = 108 installed/hall; × 4 halls = 432 installed facility-wide.
Ex6 — CRAH count: ceil(35,509 / 200) = 178 CRAH running/hall + 1 standby = 179 installed/hall; × 4 halls = 716 installed facility-wide.
Ex7 — RPP group & busway: 22 racks/group × 142 kW = 3,124 kW/group; I = 3,124 × 1000 / (1.732 × 400 × 0.96) = 4,697 A on a 5,000 A busway trunk = 93.9% loaded. 40 groups/hall (4 per row × 10 rows), 80 RPP/hall.
Ex8 — UPS & generators: 540 UPS frames/feed × 1,250 kW = 675,000 kW installed per-feed capacity; total IT 539,052.8 kW ÷ 675,000 = 79.9% UPS loading (2N, plausible). Facility electrical load ≈ 628,087 kW: 169 running (duty) + 2 standby = 171 installed 4 MW-class generators (N+2).
Ex9 — PUE across the weather year: Facility 628,087 kW / Total IT 539,053 kW = design-day PUE 1.165; annual bin-weighted 1.158 (98.2% of hours free-cool); worst-bin at 36 °C ambient 1.250. Free-cooling cliff = 34 °C ambient, and the adopted design day sits ON that cliff (0.0 K margin). Target 1.12 → gap +0.045, reported, never closed by tuning the COP.
Ex10 — Geometry & density: 62 × 31 × 5.5 = 10,571 m³/hall (1,922 m² floor). Gross IT density = 124,960 / 1,922 = 65.0 kW/m². Fire-agent design requires additional NFPA 2001 inputs; room volume alone is insufficient.

PUE reading: The derived PUE (1.165 design-day, 1.158 annual, 1.250 at the 36 °C worst bin) reflects a Carnot-fraction chiller COP under a free-cooling regime that covers 98.2% of the year, not a nameplate figure applied to 100% of the heat. The 1.12 target is a design aspiration; the engine reports the +0.045 gap rather than closing it. Rule 4 (ACCURACY_VALIDATION.md): target ≠ derived value — the cockpit shows them in different colours and chips.

07 Accuracy validation — 6 rules

Codified in standarization/ACCURACY_VALIDATION.md and enforced by tools/probe-accuracy-validation.mjs. Gate any change that touches the cockpit pages.

Rule	Requirement	Failure example
`Rule 1` — One source of truth	Every KPI on every tab reads from the same `DCAI_CALC.snapshot` — never from a separate variable or inline literal.	Dashboard showing the retired per-domain figure (see §02) while the electrical tab shows 142 kW/rack simultaneously.
`Rule 2` — No Math.random on basis KPIs	IT load, PUE, Facility load, GPU count, CDU count, UPS/transformer/generator ratings: deterministic. Reload 20× — must be identical every time.	Random jitter on TCS flow or dashboard rack-load causing different values each load.
`Rule 3` — Every metric carries a denominator	PUE = Facility ÷ Total IT; UPS loading = load ÷ rated; transformer = kVA_load ÷ kVA_rating. Denominator shown in tooltip or inline.	PUE shown without stating whether the denominator is rack IT (499.84 MW) or total IT (539.05 MW).
`Rule 4` — Target ≠ derived value	Target PUE ≤ 1.12 (design band, blue chip) and derived PUE 1.165 design-day / 1.158 annual / 1.250 worst-bin (engine output, cyan chip) are different numbers with different colour meanings.	Showing "PUE 1.12" (target) in green as if it were the operating result.
`Rule 5` — Terminology matches the adopted engineering basis	Use "NVL72 rack" — one rack IS one 72-GPU NVLink domain at 142 kW, so no disambiguating "rack position" language is needed or correct any more. Never use "rack-pos", "physical rack position", or the retired per-domain/per-position kW figures (see §02) as if current.	Labelling a rack as a "physical rack position" or citing the retired per-domain kW figure as the current one.
`Rule 6` — Basis chip on every critical KPI	Each KPI carries: MEASURED / DERIVED / TARGET / SIMULATED / ADOPTED / MANUAL. Click opens the shared basis drawer with formula, inputs, output, scope, denominator, source.	A PUE tile with no chip — operator cannot determine if the value is a target or a derived result.

08 References & standards

NVIDIA Enterprise Reference Architecture — components — governing hardware reference for the adopted 72-GPU / 36-Grace-CPU / 18-NVSwitch GB300 NVL72 rack at 142 kW.
data/dcai-parameters.json (spec gb300-500mw-2026-09-06, adopted 2026-09-06) — Single immutable parameter registry for the whole cockpit: 4 halls, 880 NVL72 racks/hall, 142 kW/rack, 85% liquid capture, TCS 40/50 °C dry-only, PUE design band 1.12–1.25.
js/dcai-model.js — authored leaves, every value typed with a // source: line and an evidence class (PUBLISHED / ADOPTED / ASSUMED / STANDARD / LABEL).
tools/test-dcai-engine.mjs — balance-identity acceptance tests (IT = rack + fabric + OOB + storage/mgmt, facility = IT + non-IT, PUE = facility / IT, weather bins sum to 8,760 h, and more) run on every change to the engine.
ASHRAE A2 / ASHRAE 90.4 — Inlet temperature envelope for HPC cooling; energy efficiency for data centres.
ISO/IEC 30134-2 — Data centre key performance indicators: Power Usage Effectiveness (PUE) — definition, measurement scope, and boundary.
ISO/IEC 30134-8 — Carbon Usage Effectiveness relative to IT equipment energy (CUE_IT).
Chiller COP methodology — derived from a Carnot-fraction over the actual evaporator/condenser lift (copFromLift()), never a vendor nameplate figure, and never calibrated to hit a target PUE.
Engine sources — js/dcai-engine.js (DCAI_CALC, pure derivation engine) + js/dcai-model.js (DCAI_MODEL, deep-frozen authored basis) + js/dcai-parameters.js (shared registry for the basis drawer).
standarization/ACCURACY_VALIDATION.md — 6 accuracy rules enforced by tools/probe-accuracy-validation.mjs.
ISA-18 alarm-management series — lifecycle, operator response, rationalization and first-out context for the simulated alarm workspace.
IEC 62682:2022 — management of alarms for process industries; terminology and acceptance reference, not a conformance claim.
js/datahall-model.js + js/datahall-calculations.js (RETIRED 2026-09-06) — the byte-frozen retired-era pair, kept only as the retirement record; see §02 "Retired basis" note.

09 Cockpit controls & view map

This operating map is additive to the locked methodology above. It identifies every visible control family in datahallAI.html and the state it is allowed to change. None of these controls writes to field equipment.

Control / view	Hook	Expected behavior	Engineering boundary
Portfolio / DC Solutions / PRD / Manual	`.bbtn`	Navigate without changing the locked scenario.	Documentation is public; the cockpit keeps its existing tier gate.
Basis of Design	`bodTrig`, `bodDrawer`	Open the calculation-audit drawer populated from `DCAI_CALC`; export its printable PDF.	It is an auditable design basis, not a field test certificate.
Generate Design / Design Studio	`genDesignTrig`, `RZDesignStudio`	Open the shared dialog, display one current engine snapshot and provenance, then allow only the registered document types and current/current-plus-study scopes.	The study appendix is labeled non-adopted; it never mutates the snapshot. Focus is trapped and returned; failure is visible.
FAQ / Theme	`faqTrig`, `themeToggle`	Open the FAQ dialog; switch/persist the light/dark palette.	Neither action changes model values.
Primary tabs	`button[data-t]`	Select Dashboard, Data Hall, Room Layout, Rack, Cooling, Electrical, Network, Fire, BMS, or Alarms & Events; one `.pn` remains active.	Tab changes presentation only.
Rack-density platform study	`platformProfile`, `platformComparisonBody`	Keep the adopted values as the baseline and optionally show the retired reference (see §02) on its own geometry.	Selection is read-only presentation; the retired reference cannot alter any operating KPI.
Sidebar collapse / reopen	`sideTog`, `sideReopen`	Widen the active diagram on desktop and remember the preference.	Does not hide alarm meaning or alter data.
Floor selection / Overview	generated floor targets, `floorBack`	Open a tagged floor plan and return to the building view.	Geometry is a design visualization, not a surveyed as-built.
Hall selection	`button[data-dh="1…4"]`	Select DH-1 through DH-4 and update active hall context.	All halls share the same adopted basis unless a future model declares a variance.
Electrical sub-tabs	`button[data-ep]`	Select facility Overview or DH-01…DH-04 SLD.	Only one `.ep` panel is active.
Electrical operating scenario	`electricalScenario`, `electricalPathSummary`, `electricalTimeline`	Evaluate normal, utility/source, busway and rack-feed events; render served/2N/degraded/lost totals, path state and ordered event timeline.	Semantic topology determines animation. De-energized paths never remain visually active; all meaning survives reduced motion.
SLD equipment	`data-sld-click`, `data-rz-line`, `data-rz-breaker`	Open the appropriate SLD mimic or right inspector for MV, RMU, transformer, MSB, UPS, busway, rack PSU or cooling feed.	Selection is read-only; no breaker command leaves the browser.
SLD presentation	`sldZoomIn`, `sldZoomOut`, `sldZoomFit`, `sldPrint`, `sldResultsToggle`, `sldMimicClose`	Zoom, fit, print, show/hide results or close the selected mimic.	Results are modeled annotations, not IEC 60909/IEEE 1584 studies unless explicitly sourced.
Rack / component detail	generated rack/tray/component targets, `rackModal`	Open the selected tag, type and component diagram.	Component inventory follows the locked reference architecture.
CDU / chiller / cooling tower	`data-cdu`, `data-ch`, cooling-tower targets	Open a tag-specific process HMI with status and values.	Local pressure/flow/temperature movement is simulated sensor-class data.
Pump / TCS / chemistry equipment	generated equipment targets, `eqHmi`	Open the type-specific equipment HMI.	Displayed setpoints/nameplates require tag-level source records before field use.
In-rack CDU / CRAH	`data-ircdu`, generated CRAH targets	Open rack hydronic or air-side detail.	Heat/flow basis remains the adopted basis; jitter is not a sizing input.
Corridor services	FWS, leak, cable-tray targets	Open corridor-service HMI for the selected family.	No service-isolation or safety command is implemented.
Battery ESS	`data-bat`, `batHmi`	Open the 480 VDC system/unit view for the selected hall and unit.	Runtime/SOC/thermal values are simulation unless identified as engine basis.
Fire cause-and-effect matrix	`fireScenario`, `fireZone`, `fireCauseEffectBody`	Select an initiating event/zone and inspect affected notification, lift, access, AHU/CRAH, smoke-control, suppression, EPO, generator, CCTV and BMS/DCIM outputs with required feedback.	FACP is command authority. BMS/DCIM only publishes/monitors the event; the browser sends no command.
Alarm query workspace	`alarmFilterForm`, `alarmResultsBody`, `alarmDetail`	Filter by date/time, point/tag, system, severity, lifecycle, quality, value comparator, state transition, event/action/text or saved view; select a row for evidence.	Records are immutable simulated fixtures, not a field historian.
First-out / CSV	`alarmFirstOut`, `alarmExportCsv`	Show the earliest event per incident and export the current filtered record set with provenance fields.	Export is evidence from the simulation only; it is not an acknowledged site log.
KPI basis cards	`.k[data-basis]`	Click, Enter or Space opens formula, inputs, output, scope, denominator, source, mode and timestamp.	Uptime and alarm drawers explicitly identify design placeholders.
Close controls / Escape / scrim	modal-specific close IDs + shared modal controller	Dismiss the active surface and return focus to its trigger.	Focus must never fall silently to `body`.
Scroll to top	`scrollTopBtn`	Return to the header without altering current scenario.	Reduced-motion preference removes nonessential smooth animation.

10 Telemetry, alarm & modal operating map

The cockpit deliberately contains three point classes. Basis/derived values are repeatable engine outputs, sensor-class values may jitter within declared ranges, and placeholder values require a future event or field source. Read the data-mode and quality label before acting on any number.

Point family	Representative hook	Class / cadence	Alarm or quality reading
Facility inventory and IT	`DHE.halls`, `dkIt`, `sPit`	ADOPTED; immutable.	4 halls, 880 NVL72 racks/hall, 124,960 kW rack IT/hall, 539.05 MW total IT facility (PUE denominator).
Heat split / flow / CDU / CRAH	`sDLC`, `sAir`, `sCDU`, `sCRAH`	Derived; immutable.	106,216/35,509 kW liquid/air per hall, 9,612.3 m³/h TCS flow, 107/108 CDU and 178/179 CRAH duty/installed per hall.
Electrical capacity	`reqCurrentA`, transformer/UPS/busway fields	Derived; immutable.	4,697 A per RPP group, 93.9% busway loading, 79.9% UPS loading, 5,000 A busway trunk.
PUE / WUE / CUE / IT / GPU / racks	`dkPue`, `dkWue`, `dkCue`, `dkIt`, `dkGpu`, `dkDom`	Derived or ADOPTED; 4 s reassertion for dashboard fields.	Final text is deterministic; source/denominator lives in the basis drawer and tooltip.
Uptime / active alarms KPI	Uptime card, `dkAlm`	Design placeholder.	99.99% and zero active are not a persisted outage/event history.
TCS and room sensors	`sTS`, `sTR`, hot/cold temperature/RH IDs	Sensor-class jitter; 4 s cadence.	Jitter is allowed only around 35/45°C or declared room bands; it must not alter locked ΔT sizing.
Occupancy/access	`pplCount`, `lastEntry`, `dkPpl`, `dkEntry`	Sensor/browser-time simulation.	Not a physical access-control record.
Fire/leak/maintenance counts	`sbSmoke`, `sbHeat`, `sbVesda`, `sbRope`, `sbDisabled`, `sbMaint`	Mixed static and sensor-class simulation.	Do not infer code compliance or maintenance work orders.
FWS instruments	`fwsPI`, `fwsTI`, `fwsFM`	Sensor-class jitter.	Ranges are illustrative; every HMI reading needs a simulated/quality chip.
Alarm strips	`.dh-alarmbar`	Rule evaluation every 4 s.	Critical/warning checks cover inlet temperature, CDU margin, UPS loading, TCS ΔT and stale percentage; maintenance is two scheduled items.
Equipment modal summary	`.dh-modal-alarmline`	Same alarm rule result at open.	Summary-first state must match the active panel strip.
BMS service health	`bmsGwOnline`, `bmsAlmActive`, `bmsAlmAck`, `bmsAlmCleared`	Static service baseline plus shared rules.	16/16 gateways is modeled; “cleared” currently reuses the maintenance count and is not a true event lifecycle.
Line / breaker quality	`data-rz-line`, `data-rz-breaker`, inspector Live tab	Local rendered state.	Seven breaker states and data-quality chip must remain distinguishable without color alone.
Global data mode	`body[data-rz-data-mode="simulated"]`	Persistent provenance.	The word Simulated must remain visible even when values animate.
Alarm-query records	`RZDataHallAlarmQuery`	Immutable simulated event fixtures.	Date/point/system/severity/lifecycle/quality/value/state/event/action/text filters compose; first-out selects earliest per incident and CSV exports the filtered set.
Electrical scenario truth	`RZDatahallAIElectrical.evaluateScenario()`	Pure semantic topology evaluation.	Served + lost = 54; 2N/degraded are mutually interpretable availability classes; CSS animation is a consumer, never the source of truth.
Fire cause/effect	`RZDatahallAIFireCauseEffect.evaluateEvent()`	Immutable zoned training matrix.	Rows expose command authority, expected feedback, inhibition and reset authority. FACP controls; BMS/DCIM monitor-only.
Rack-density comparison	`RZDataHallRackDensity`	Adopted baseline plus a retired-only reference.	Adopted = 880 racks × 142 kW/rack; the retired reference sits on its own smaller hall geometry (see §02). Neither value is field telemetry.

Threshold and lifecycle source: the panel alarm rules use rack inlet >27°C warning / >30°C critical, CDU margin <15% / <5%, UPS load >80% / >95%, TCS ΔT >13 K / >15 K, and stale points >1% / >5%. The query workspace adopts ISA-18/IEC 62682 lifecycle and first-out terminology, but its records and thresholds remain deterministic review fixtures—not approved field setpoints, historian evidence or alarm-system conformance.

11 Glossary & RZExplain

The shared RZExplain scanner attaches contextual definitions to marked engineering terms. The table also records the exact meaning used by this cockpit.

Term	Cockpit meaning	Do not confuse with
NVL72 rack	One physical rack IS one 72-GPU / 36-Grace-CPU / 18-NVSwitch NVLink domain at 142 kW — no split-domain footprint in the adopted basis.	The retired "domain spans two rack-positions" construct (see §02 retired-basis note).
Direct liquid cooling	Cold-plate path capturing 85% of rack IT heat.	Total facility cooling electrical power.
TCS	Technology Cooling System at 40/50°C, warm and dry-only.	Facility-water/chilled-water loop.
FWS	Facility Water System coupled through CDU heat exchangers.	Rack-side coolant circuit.
CDU	Coolant Distribution Unit — CoolIT CHx1000, 1,000 kWth design rating, 107 running per hall.	Chiller or CRAH.
CRAH	Air-side unit serving the residual heat outside the liquid loop.	Liquid-cooling primary path.
PUE	Facility power divided by total IT power; 1.165 design-day is derived, 1.12–1.25 is a target band.	A marketing target presented as measured output.
WUE	Water per IT energy; 0.00 only for the declared dry-only baseline.	Domestic/process water or an adiabatic configuration.
CUE	Carbon per IT energy; grid factor × PUE in the current calculation.	Raw facility-kWh grid emission factor.
2N	Each UPS feed can carry the protected load; normal display assumes equal sharing.	N+1 parallel capacity.
Basis / sensor / placeholder	Immutable engine fact / bounded simulated reading / value awaiting authoritative source.	Three interchangeable meanings of “live”.

01 Purpose & engineering basis

02 Inputs — ADOPTED basis + tab selection

03 Calculation methodology

Rack and IT load — compute()

Rack density — racks_per_nvl72_domain

Facility load — compute()

PUE — op.pue

Liquid heat / air heat — compute()

Hydronic TCS/CDU flow — flowM3h()

CDU count — ceilCount()

CRAH residual cooling — airM3s()

Chiller COP — copFromLift()

Bottom-up PUE across the weather year — compute()