What AI Is Made Of

Run 03 — Material Forensics

The fixed plant's copper, aluminum and steel balances remain OPEN. Material forensics establishes battery chemistry minima, physical construction scenarios and tested accounting bounds, but does not establish qualified plant metal base cases. No complete plant mass, finite whole-boundary metal upper bound or material completeness percentage is asserted.

PLANT-GB300-1GW-2026Q4-A still implements CFG-GB300-2026Q4. SNAP-RUN03 adds research and material accounting without changing the architecture, canonical equipment counts, operating assumptions or prior snapshots. Run 03 ends at embodied equipment stock. Civil quantities, semiconductor manufacturing inputs and extraction remain outside scope.

Before the new records

The latest Run 02 branch was pulled at 5ef3841c1b6d01aba47c7177f28d0c4c9b3b9aa2. The governing standard, prior findings, BOM, reconciliations, uncertainty method, handoffs, conflicts and registers were inspected. All 193 inherited tests and all four existing reproducers passed before any Run 03 records were created. All 680 inherited tracked files reproduced unchanged. The complete SNAP-RUN02 material-balance SHA-256 remained 8d8e7aef912ba16f0f3be3a23464024a22f5cc20e22d1e1e66cbd69b38942729.

Preflight fingerprints (repository-only reference) protect all 1,717 inherited canonical entities. The new reproducer independently recalculates that historical report and checks the fixed equipment BOM and retained evidence. The populated plant has 107 modules. Its 6,785.714285714286 rack equivalents are distinct from the 6,848 provisioned rack positions. Research projections use the former where the canonical quantity specifies rack equivalents.

What can now be attributed

Evidence / physical scopeDefensible resultAdmission and limit
Selected Li5 cabinets, both B2 populations4,924.140 t declared gross; 8,346 cabinets and 4,056,156 physical cellsInherited exact selected assembly stock; includes modules and cells
LFP active-material equivalentAt least 477.494450 tCapacity/Faraday bound; fully lithiated equivalent, not a phase assay at deployment state of charge
Reversible lithium inventoryAt least 21.006095 tSame one-electron inventory; additional electrolyte and inactive lithium remain unknown
Total elemental lithium in battery cabinets21.006095–4,467.651645 t, base unknownIntentionally loose ceiling: gross less mandatory non-lithium FePO4 framework; accommodates electrolyte/excess Li
Iron / phosphorus / oxygen within that equivalentAt least 169.032473 / 93.751841 / 193.704040 tNested elemental views; iron in the cathode is not steel; do not add to compound mass
Whole battery copper, aluminum or steelSeparately 0–4,446.645550 t, base unknownGross less minimum non-metal active-equivalent stock; maxima share a single budget
Selected chiller R513A180.830 tExisting design-qualified declaration retained exactly; not added again as a child inventory

The battery cabinet specification is 60 Ah. Eighteen series modules each contain 9S3P cells, implying 20 Ah/cell. The lower-bound equation is n_min = N_cells × 20 Ah × 3600 C/Ah / F; conventional atomic masses convert moles to the equivalent and elemental inventories. This is conditional on the selected nominal LFP capacity specification and one-electron chemistry. The rounded 31.1 kWh figure is checked rather than silently used to alter the canonical energy. Lithium moved into the anode at charge remains inside the inventory; the minimum does not require a fully lithiated cathode at deployment. See material bounds (repository-only reference) and the battery handoff.

Conditional constructions, kept separate

These are reproducible physical cases for matched reference objects and unselected construction choices. Their low/base/high values describe stated inputs or recipe assumptions. They are not confidence intervals for the selected plant and do not close its material balance.

Construction caseFixed-count projectionPrincipal unresolved transfer
Four-full-conductor 800 A copper busway589.148 t copper conductor stockSelected vendor, conductor material and section unspecified
Four-full-conductor 800 A aluminum busway214.718 t aluminum conductor stockAlternative to the copper case, using its own catalog section
Thin / intermediate / thick straight pipe shells3,553.159 / 5,463.170 / 8,091.984 t ferrous stockFinal wall, bore and metallurgy unselected; carbon steel and stainless steel retained separately
Two MGX reference busbar assemblies per rack equivalent303.281 / 304.226 / 305.171 t copper-grade stockSelected Supermicro part/count unqualified; stock designation is not an elemental copper assay
Two MGX reference manifold main bodies per rack equivalent107.646 t stainless steelMain-body-only geometry and selected part/count unqualified
Cylindrical IFR32650 cell recipe, three utilization cases162.814 / 172.882 / 186.476 t copper; 113.723 / 120.755 / 130.250 t aluminumActual high-power cell identity unknown; module and cabinet residuals retained
Same cell recipe, active-equivalent inventory477.494 / 507.019 / 546.888 tTheoretical utilization / literature utilization / formation-reserve case; not a guaranteed target interval
Comparable copper-wound 1 MVA Galaxy PDU7,789.600 t grossProduct-specific copper construction and net mass do not establish the selected generic PDU composition or ceiling

Busway conductor mass uses rho × area × length × conductor count, with 3,424 rails and 12 m selected allowance per rail. Copper and aluminum are mutually exclusive electrical material alternatives within the unchanged topology. The conductor stock labels preserve alloy/purity scope. The battery recipes are three distinct documented commercial families, each with three explicit utilization inputs; no recipe mean is promoted to an observed cell. The scenario composition (repository-only reference) retains every component, range, assumption, replacement relationship and provenance.

Results by research class

Transformers. The 24 campus 84 MVA transformers, 428 IT 5 MVA and 214 facility 3.5 MVA stepdowns were investigated separately. The 1,712 downstream 1 MVA transformer PDUs form a further distinct class. Selected conductor, winding geometry, cooling, impedance and compatible dry composition remain unresolved. Bottom-up winding and core geometry reproduced three independent published transformer designs with maximum errors of 0.209% and 1.222%, respectively. These are engineering design comparisons, not measured hardware. A loss/current-density winding shortcut failed against 18 published large transformer designs: mean absolute error 21.348%, maximum 70.746%. It is rejected, and no shortcut kg/MVA coefficient enters the base case. The transformer handoff retains parameters, source identity, validation errors and missing target inputs.

UPS and generators. The exact selected 2 MW UL UPS gross mass remains 4,710 kg, with batteries separate. A new service manual confirms internal copper busbars but provides insufficient geometry for their mass. Its option-dependent maximum shipping weight is not an uncertainty interval around that selected net mass. A CE 400 V PEP retains unresolved denominator, packaging and product mapping. Major feeder masses remain per-route-metre calculations because route lengths are unknown. C175 engine and service-part evidence identifies nested parts but does not prove a material composition for the selected 480 V alternator or genset. Gross generator mass is not steel; cast iron, alternator copper, radiator, skid, enclosure and fluids remain separate questions. See the electrical handoff.

Batteries. The exact cabinet topology and current capacity specification support the minima above. Published OEM cabinet populations validate a simple assembly mass relationship to within 0.112% on held-out populations, but its slope includes module-associated hardware and its intercept is not steel. A BYD cell teardown's area/areal-capacity relation predicts measured capacity within 0.507%; collector masses lack an independent weighing check and its technology does not establish the selected high-power cell. Literal collector thickness reconstruction from another published teardown violates measured gross conservation in three of five cells. That interpretation is rejected, with the conflict retained. LFP active material, graphite, Cu/Al collectors, ferrous casing, electrolyte, polymers and elemental views are separately tracked.

Cooling. Exact family specifications identify likely frame, compressor, evaporator, coil and fan materials but not their masses. F-series direct free-cooling fin/tube evidence is kept distinct from N-series indirect microchannel evidence; a later generic family diagram cannot establish the selected FH coil option. Pipe annular-section models validate against six published physical catalog weights with maximum error 1.141%. CDU preliminary wet and chiller legacy estimated-dry masses remain conditional, not target ceilings. Generic warm-water CRAHs and pumps lack compatible gross bounds; there is insufficient evidence to declare their metal contribution immaterial. See the cooling handoff.

Compute racks. Public MGX CAD provides disjoint copper-grade power/return busbar solids and a stainless manifold main body. CAD metadata and BRep volume checks agree; commercial copper-size mass checks have maximum error 0.303%. The 1,400 A busbar rating is lower than a 140 kW rack load at its stated DC voltage, so two assemblies are an explicit conditional reference case rather than an observed selected rack count. Rack frames, trays, PSUs, cold plates, PCBs, packages, HBM and networking are not multiplied into an invented semiconductor composition. Cold plates remain nested within B1. See the compute handoff.

Copper, aluminum and steel reporting

Material and incremental boundaryDeclared/design-qualified subtotalExpanded qualified forensic subtotalWhole-boundary outer intervalStatus
Copper, each of B1 / B2 / modeled B3UnknownUnknownLower 0; base and upper unknownOPEN
Aluminum, each of B1 / B2 / modeled B3UnknownUnknownLower 0; base and upper unknownOPEN
Steel, each of B1 / B2 / modeled B3UnknownUnknownLower 0; base and upper unknownOPEN

Unknown subtotals do not mean zero metal. The selected exact-gross UPS/STS/battery subset provides finite class ceilings; other high-leverage equipment and unrouted conductors remain unbounded. Therefore a finite whole-boundary remaining copper exposure cannot be reported. Construction scenarios retain useful quantities but do not become expanded qualified subtotals solely because their arithmetic or source-object validation is precise. Each machine-readable result reports its denominator: 1 GW normalized commissioned IT and the separately reproduced 1.009178947368421 GW provisioned IT capacity.

Accounting and next evidence

The hierarchy and uncertainty method (repository-only reference) makes knowledge level separate from source grade, evidence-use label and admission. Closure is material × boundary, with explicit capable-class coverage, high-contributor qualification, finite exposure, overlap, uncertainty and provenance criteria. The 2% screening rule cannot pass with a null base or unbounded class.

Each equipment record retains gross mass, disjoint physical components, nested elemental views and unresolved residual. Components or guaranteed physical minima exceeding compatible gross fail validation. Alternative rows replace their COMMON class record; they never add a second parent inventory. An unresolved residual is not assigned to steel. Priority ranking (repository-only reference) uses material leverage as an inference, not as a coefficient.

The highest-value next evidence is selected transformer winding/core data and conductor choices; exact PDU construction; cell SKU and high-power composition; selected FH coil and heat-exchanger geometry; selected rack mechanical BOM; 480 V alternator parts and material masses; and busway/feeder sections and routes. The new bounds can support future closure only after dominant unbounded classes are resolved. No design change is required to ask those questions.

Reproduce the final research and preservation audit with python scripts/reproduce_run03.py; run the full suite with python -m unittest discover -s tests -v. The generated manifest (repository-only reference), conservation audit (repository-only reference), closure criteria (repository-only reference) and preservation audit (repository-only reference) make these conclusions reviewable.

Final verification: 256 tests pass, with six additional transformer-model tests. All four inherited reproducers and the new Run 03 reproducer pass. Seven offline research builders repeat without changing their artifacts; canonical integration is idempotent. The final audit verifies 155 source hashes, all 217 indexed Run 03 research artifacts and all 1,717 inherited canonical entities. Earlier generated audit files are retained at their historical versions; fresh current validation and source checks are in build/run03/. Every inherited physical output remains byte-identical.

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_03_FINDINGS.md
SHA-256: 6917e7a59960df9c5859b9a1265401972ffdff2264c50814691fa07b12f410d5
Download research file

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