What AI Is Made Of

Mass Balance of Intelligence: Run 02 findings

Snapshot SNAP-RUN02 · evidence access date 7 October 2026 · fixed object PLANT-GB300-1GW-2026Q4-A · unchanged compute configuration CFG-GB300-2026Q4.

The first plant-specific material balance remains OPEN. Qualified copper, aluminum and steel subtotals are null at every boundary. Run 02 establishes 180.83 tonnes of R513A refrigerant in the selected B2 chillers, records selected equipment masses with unresolved composition, and preserves explicitly qualified gross-mass proxies and construction alternatives separately. The public evidence does not support total plant mass, total copper required, or a material completeness percentage.

This is a result about the specified reference design. It is not an observed procurement inventory or a universal material intensity for 1 GW AI facilities. The auditable calculation is installed design count × material kg per selected equipment unit. Its machine-readable balance (repository-only reference), equipment reconciliation (repository-only reference) and CSV reconciliation (repository-only reference) expose admitted coefficients, explicit unknowns, source terminology, product matching, boundary and lineage.

Preserved reference plant and preflight

The latest upstream Run 01.5 repository, commit 15155a6dbddc18afe3e4ae55f1c561d8269daa99, was inspected before material work. README, Accounting Standard v0.1, previous findings, the equipment BOM, standard changelog, unknowns, design decisions, sensitivities and canonical tables were read. All 146 inherited tests passed. The final Run 01.5 BOM reproduced byte-for-byte in JSON, CSV and Markdown. The preflight register (repository-only reference) and generated preservation checks (repository-only reference) retain the evidence.

The repository's final design governs exact counts. In particular it contains 642 module stepdown transformers (428 IT and 214 facility), 856 CDUs, 2,140 CRAHs and 1,070 chillers. These differ from the approximate counts in the Run 02 brief. The design remains 107 complete modules, six campus substations, 24 × 84 MVA campus transformers, 856 IT UPS units, 214 facility UPS units, 1,712 transformer PDUs, 1,712 IT STS units, 428 external pumps, 214 × 50 m³ tanks, 642 generators and 8,346 Li5 battery cabinets.

The denominator remains exactly 1 GW normalized commissioned IT. Installed IT capacity is 1.009178947368421 GW, and normalized compute inventory remains 6,785.714286 GB300 rack equivalents, 488,571.428571 B300 accelerators and 244,285.714286 Grace CPUs. Rack equivalents are a normalized design inventory, not purchased integer racks. No compute configuration, selected equipment, thermal architecture or plant count was replaced. Accounting Standard v0.1 and historical source, claim, ledger, design and snapshot records remain intact.

Quantified materials by boundary

All entries below are quantified portions, rather than complete physical stocks. UNKNOWN represents null. Missing scope does not contribute an assumed zero.

MaterialB1B2 incrementalB2 cumulative quantified portionB3 incremental modeledB3 cumulative quantified portion
Elemental copperUNKNOWNUNKNOWNUNKNOWNUNKNOWNUNKNOWN
Elemental aluminumUNKNOWNUNKNOWNUNKNOWNUNKNOWNUNKNOWN
SteelUNKNOWNUNKNOWNUNKNOWNUNKNOWNUNKNOWN
R513A refrigerant blendUNKNOWN180.83 t180.83 tUNKNOWN180.83 t

The R513A cumulative values carry forward the enumerated B2 stock. They do not claim that unknown B1 or B3 stock equals zero, or that the complete cumulative boundary has closed. Boundary subtotals (repository-only reference) retain unresolved scopes and is_total=false. Cumulative B2 adds B1 and incremental B2; cumulative B3 additionally adds incremental modeled B3. Cumulative values are never added to one another.

The qualified refrigerant calculation is:

1,070 selected FH3145 chillers × (89 + 80) kg R513A/chiller = 180,830 kg.

Vertiv's captured SL-70790 installer guide (repository-only reference), printed p53, Table 7.1, declares the two separate circuit charges. The guide's model decoding identifies the selected direct-free-cooling FH3 145 product and R513A blend. Although Run 01.5 selected FH3145 as a functional alternative to RD030's FF3145, this charge is exact evidence for the selected FH3145; no charge is transferred to FF3145. The subtotal is DESIGN-QUALIFIED: selected design count, exact selected product coefficient, reproducible multiplication and clear material/boundary. It does not assert measured commissioned charge or acceptance of every installation interface.

The subtotal is 180.83 tonnes / GW IT and approximately 179.185 tonnes / GW installed IT. Allocating it across normalized compute racks gives approximately 26.65 kg/rack equivalent of facility refrigerant; that stock is in chillers, not inside compute racks. All currently quantified R513A comes from this one equipment class. Its 100% contribution share describes the quantified R513A subtotal only. Copper, aluminum and steel contributor shares, top-five/top-ten concentrations and intensities remain unavailable.

Equipment mass with unresolved material composition

Gross equipment evidence is useful, but it cannot be relabeled as copper, steel, battery active material or a complete plant mass. The gross inventory (repository-only reference) separates selected equipment evidence from conditional proxies.

Incremental scope / equipmentCalculationGross massEvidence use and limitation
B2 IT UPS856 × 4,710 kg4,031.760 tSelected 2 MW UL 480 V product; batteries separate; composition unresolved
B2 IT STS1,712 × 2,177 kg3,727.024 tSelected LayerZero 1,600 A product; composition unresolved
B2 critical-cooling STS107 × 975 kg104.325 tSelected LayerZero 800 A product; composition unresolved
B2 selected electrical subtotalThree disjoint rows above7,863.109 tScoped gross equipment subtotal
B2 IT/facility battery cabinets(7,704 + 642) × 590 kg4,924.140 tSelected 18-module Li5 cabinet gross; chemistry is LFP, constituent masses unresolved
B2 selected gross subtotalElectrical plus battery cabinets12,787.249 tKnown selected gross equipment for these classes only
B1 coldplate assembliesPreserved Run 01.5 rack equivalents × 18 × 9.5 kg1,160.357143 tTier 3 PROXY; population/function match, selected installation unqualified
B3 standby generators642 × 25,000 kg16,050 tTier 3 dry gross-mass PROXY; published power-family voltage differs from selected 480 V class

The B1 count is derived from the plant-linked preserved Run 01.5 rack quantity, multiplied by 18 trays/rack. It does not substitute a differently scoped historical count for the plant lineage. LITEON's GB300 coldplate sheet (repository-only reference) reports a 9.5 kg module for four Blackwell Ultra GPUs and two Grace CPUs, matching device population and 1,400 W GPU thermal rating. It includes tubing and connectors. Its stated 2U chassis suitability differs from the selected 1U Supermicro tray; actual mounting, routing and dry/wet mass are unqualified. Low/high transfer bounds remain null. It is a conditional assembly proxy, not a copper coefficient or known exact-target mass.

The generator proxy excludes operating fluids, fuel stock and separately modeled fuel systems. The three boundary gross values above do not form a complete plant denominator. Their missing equipment and product mismatches prevent a material closure calculation. Gross equipment and nested component or material views must never be added together.

What the workstreams established

B1 compute. Exact rack/tray, frame, GPU/CPU/HBM, NVLink, NIC, shelf/PSU, busbar, cable, coldplate/manifold and PCB evidence was investigated. Public OEM carbon-footprint summaries disclose some product weights and internally collected material data, but their emissions bars are not material-mass fractions. Same-platform Lenovo weights, Supermicro MGX frame-family masses, the OCP 31 kg busbar assembly and measured A100 teardown remain context with explicit mismatch records. The GB300-labelled Supermicro 22 kg sheet describes an incompatible two-GPU/test-board configuration and is excluded. See the B1 evidence handoff (repository-only reference).

The 977,142.857143 embedded GPU logic dies remain a nested component count. Physical silicon mass needs qualified area, post-thinning thickness and disjoint logic/HBM/interposer/package scope. No semiconductor mass is inferred from memory GB, transistor count, generic wafer thickness, fabrication yield or wafer starts. The manufacturing questions remain for a later run.

Electrical infrastructure. Exact selected UPS/STS weights are retained. Transformer, alternator, busway and UPS material declarations were located, but unselected winding conductors, voltage/topology differences, incomplete manufacturing-model boundaries or inconsistent packaging denominators prevent qualified Cu/Al/steel transfers. Family checks do not establish a universal kg/MVA or kg/A law. Complete campus substations overlap their separately modeled transformers, panels and protection and receive no additive umbrella coefficient. See electrical candidates (repository-only reference) and the transformer handoff (repository-only reference).

Cooling infrastructure. Selected dry/air-cooled heat rejection remains fixed. Schneider chiller/CRAH declarations, a Carrier CDU declaration and stainless tank alternatives are retained with rating, cooling, material and configuration mismatches. LCIA extrapolation factors do not establish material extrapolation factors. Copper, aluminum, steel and stainless steel in the selected cooling equipment remain unresolved. Newer FH3145 family dry/static/operating weights have different definitions and do not resolve coolant inventory by subtraction. See the cooling handoff (repository-only reference).

Batteries. The design already selected Li5 LFP: 8,346 cabinets, 150,228 nested modules and 259.5606 MWh nominal DC energy. Chemistry does not provide element coefficients. A same-family seismic certificate's 53 lb module gives an isolated Tier 2 assembly partition scenario of 3,611.540552 t modules and 1,312.599448 t non-module residual, reconciling to 4,924.14 t cabinet gross. The certified and current cabinet versions differ, so transfer bounds remain null. Neither the module mass nor residual is a selected elemental coefficient; the residual includes conductors, controls and enclosure, not only steel. The partition remains separate from headline material subtotals. See the battery handoff (repository-only reference).

Lithium, nickel, cobalt, manganese, iron/phosphate-bearing materials, graphite, copper, aluminum, polymers and electrolyte masses remain null where exact or qualified coefficients are unavailable. LFP identity does not warrant setting Ni/Co/Mn to zero across all module and cabinet components.

Fluid capacity, inventory and annual water

The fluid output (repository-only reference) reports 18,313.806876 m³ of nominal geometric capacity: 10,700 m³ full tank volume, 6,807.045882 m³ primary straight-pipe capacity and 806.760993 m³ secondary straight-pipe capacity. DN450/DN200 are the inherited hydraulic internal-diameter approximations, not qualified manufactured bores. Fill, fittings, expansion space, equipment internals and rack loops remain unresolved. This is not an observed filled water inventory.

An isolated OEM EG30 recipe/density scenario converts this geometry into glycol and water masses, but procurement has not qualified that exact weight-basis recipe, inhibitor package, density or fill condition. Its solution parent is not added to its glycol/water children. Annual whole-site water withdrawal, consumption and makeup remain null. The selected architecture has no cooling tower or adiabatic spray circuit; this scoped absence does not establish zero whole-site water use.

Coverage, uncertainty and external comparisons

Structural coverage (repository-only reference) identifies admitted material classes and unresolved equipment, rather than inventing a material completeness denominator. The principal missing decompositions are rack internals; both transformer layers; switchboards, PDUs and facility conductors; battery cells/enclosures; cooling equipment/piping; generators and their fuel systems. Device inventories for external IT remain unqualified. Buildings, foundations and other unspecified civil/site scope are not closed by the equipment BOM. Shared grid generation/transmission, generic upstream manufacturing, mines and refineries are outside the current modeled B3 scope.

Low/base/high follow published points, genuine coefficient uncertainty and qualified transfer limits. Unavailable endpoints remain null. The refrigerant 169/169/169 kg declaration represents an OEM point, not zero physical uncertainty. No arbitrary percentage band, statistical confidence interval or missing-scope uplift is applied. Uncertainty method (repository-only reference) explains propagation, component overlap and isolated alternatives.

External copper and grid estimates are researched after bottom-up reconciliation and retained as benchmarks with their capacity denominator, boundary, vintage, geography and stock/flow semantics. They do not supply plant coefficients. With the plant copper subtotal unresolved, a numerical plant-versus-benchmark gap or error cannot be asserted. A broad benchmark cannot fill the missing denominator. See the benchmark reconciliation (repository-only reference) and evidence captures (repository-only reference).

No ore requirement or extraction allocation is calculated. Recycled-content or manufacturing-input declarations, where encountered, retain their source basis and do not silently change installed elemental stock.

Reproduce and assess the result

python -m unittest discover -s tests -v
python scripts/reproduce.py
python scripts/reproduce_run015.py
python scripts/reproduce_run02.py
python -m mbi material-balance --plant PLANT-GB300-1GW-2026Q4-A --snapshot SNAP-RUN02

The inherited suite passed 146 tests; the integrated Run 02 suite passes 193 tests. Tests exercise material identity, proxy gates, unresolved nulls, nested scope, incremental/cumulative arithmetic and historical preservation. Internal consistency cannot supply missing physical composition evidence.

The reproducer validates canonical records and local source hashes, preserves historical entities and final Run 01.5 BOM bytes, rebuilds the SQLite query layer and writes audited artifacts (repository-only reference), including the provenance register (repository-only reference). Run 02 adds first-class material-coefficient and plant-application records, source terms, proxy qualification and DESIGN-QUALIFIED evidence use without revising the governing standard. No automatic public publication or supplier messages occur.

The immediate evidence priorities are selected Supermicro rack/tray material declarations, selected transformer winding/BOM specifications, exact voltage UPS/STS and PDU declarations, current Li5 module/cell substance masses, and selected chiller/CDU/CRAH/pump/piping decompositions. These are the gates to closing the first copper, aluminum and steel balance.

Source details

Archived Run 01-03 research. These statements retain their time-of-writing context; current material estimates use Run 04. Web prose uses current site terminology; downloadable records preserve original wording. Hashes identify the downloadable records.

Source: RUN_02_FINDINGS.md
SHA-256: 87e2c7792ca1cf8f0da96467402df9d5e7a29042eb4e08581e6aafa0e85aabb1
Download research file

Optional: tell Sid you’re reading
The document is open; sharing your name/email is optional.