What AI Is Made Of

Run 03 B1 compute material forensics

Fixed object: PLANT-GB300-1GW-2026Q4-A / CFG-GB300-2026Q4. The baseline gate was completed by the integrating agent before these records were created. No architecture, equipment count, historical claim, or Run 02 capture was changed.

The useful new result is a reproducible material decomposition of a narrow reference component, not a completed B1 subtotal. Public NVIDIA/OCP MGX v1.1 CAD provides separate power and return conductor solids. Combining their actual design volumes with the prescribed copper grade yields 22.347 / 22.417 / 22.486 kg of copper-grade substrate stock per reference 1400 A busbar assembly. The interval comes only from the Copper Development Association's published density interval. It does not bound selected-product transfer, manufacturing tolerances, plating, or elemental purity.

A second result is 7.932 kg of 304 stainless steel in one reference dry main manifold body, excluding caps, ports, fittings, hoses, valves, coolant and mounting hardware. This is a base-only reference-object reconstruction. Selected Supermicro manifold identity remains unresolved.

Both models remain SCENARIO ONLY when projected onto the fixed plant. No new DESIGN-QUALIFIED or expanded-forensic B1 Cu/Al/steel subtotal is admitted. The selected rack's total gross mass is still unknown, so a finite global B1 material upper bound cannot be justified.

Canonical quantities and provisioning

canonical_quantities.json is calculated directly with calculate_balance(load_data(), snapshot="SNAP-RUN02"). Base quantities are 6,785.714285714285 fractional rack equivalents, 122,142.85714285713 compute trays, 61,071.428571428565 NVLink trays, 54,285.71428571428 power shelves and 325,714.2857142857 PSU modules. These values remain conditional normalization equivalents, not purchase orders or observed installations.

The retained 5% extra-rack electrical IT reserve does not identify an external device inventory. Fully equipped but unused last-block B2 rack positions and shelf whips do not install additional B1 racks, shelves, trays or PSUs. Known GPU, CPU and embedded logic-die counts remain in the quantity capture; they do not establish semiconductor material mass.

Copper reference busbar model

The OCP server project publishes the MGX v1.1 rack design file. Its ZIP contains a 512,613,012-byte STEP file, SHA-256 ae87e29e27aded047beefc93b8dfbe5b320f0a8aa0a9525d2870ab63430484e5. The original compressed archive is preserved; the giant extracted STEP is held outside the Git tree. Compact extracted solid captures and the extraction script make review possible.

The exact disjoint leaves are:

LeafSTEP productShape representationVolume mm³
NVIDIA_PWR_BUSBAR_2_1_27024507024481,257,535.864604
NVIDIA_RTN_BUSBAR_2_1_27031077031051,257,535.864604

The OCP specification p38 prescribes C11000 H02 copper or equivalent for these power/return conductors. The Copper Development Association C11000 page supplies 0.322 lb/in³ at 68°F, with an explicit actual range of 0.321–0.323. Using exact unit conversions gives 8,885.249 / 8,912.929 / 8,940.609 kg/m³. The model is:

m_kg = 2 × 1,257,535.864604 mm³ × 10^-9 m³/mm³ × density_kg/m³.

C11000's stated minimum 99.90% composition includes silver. Therefore this is copper-grade conductor stock, not an assayed elemental-Cu mass. Any near-pure bulk-Cu normalization must remain an explicit engineering assumption. The separate silver-plating thickness is not known. Aluminium 4047 chassis grounding is identified but not quantified: mapping every support/cover CAD solid to its material is not established, and residual mass is not assigned aluminium or steel.

The reference busbar's typical 31 kg total gives an unresolved correlated residual of 8.514 / 8.583 / 8.653 kg after selecting the two copper-grade leaves. This is a source-object bookkeeping budget with an approximate gross anchor. It is not a selected-plant bound. Parent assembly volumes and repeated property declarations are not additive to the leaves.

For sensitivity, two such reference assemblies per frozen base rack count project to 303.281 / 304.226 / 305.171 tonnes of copper-grade stock. These are density-only conditional values, not the plant's material LOW/BASE/HIGH. A 1400 A assembly at 48–54 V carries roughly 67–76 kW; blindly installing one per approximately 140 kW selected rack would be unsound. Two is an explicit sensitivity mapping, not evidence that the selected Supermicro rack has two particular assemblies, the same cross-section, or the same grounding topology.

Main manifold body model

OCP specification p24 identifies the main manifold tube as stainless 304 or equivalent. The matching CAD leaf MGX_EIA_RACK_MANIF_BODY_LEFT_1_, product 4446747 / shape 4446744, has volume 1,004,023.737343 mm³. Its actual CAD bounding dimensions are 35 × 2061.05 × 60 mm. The p25 drawing's 2079.1 mm tube-assembly length has additional scope; applying it directly to the bare body would be an overlap error.

Using Outokumpu's Core datasheet nominal 304 density of 7,900 kg/m³ gives 7.931787525 kg per dry body. The conditional two-body projection is 107.646 tonnes of stainless steel at the fixed base rack count. The projection assumes supply/return reference bodies of equal volume; no exact selected Supermicro material or dimensional map was found. Stainless steel is retained as stainless steel, not elemental iron or carbon steel.

Validation and admission limits

extract_cad_solids.py extracts the exact entity graphs and independently computes BRep mass properties with Open Cascade through CadQuery. All three extracted objects are valid one-solid bodies. Copper volume errors relative to embedded CAD declarations are about 6.83×10^-9; manifold volume error is 5.32×10^-9. This verifies geometry extraction and units, not as-built material measurements.

Independent declared commercial objects provide further checks. FixMetall Cu-ETP bars, at 50×10, 60×10 and 80×10 mm, list 4.47, 5.36 and 7.14 kg/m. The same density-volume calculation differs by 0.30%, 0.23% and 0.14%. Its 304 60×40×3 mm and 60×40×2 mm tube declarations differ from the rectangular hollow-section calculation by 0.10% and 0.12%. These catalogue weights may themselves be theoretical; no weighing method is disclosed, so they are DECLARED validation objects, not MEASURED teardowns.

No validation object establishes selected Supermicro part identity. Precision does not upgrade weak product transfer. The reference density-only range is consequently kept separate from target transfer, whose LOW/BASE/HIGH remain null.

Other investigated bulk components

The new Vertiv MGX v1.1 datasheet gives a steel frame and 180 kg net rack assembly within the same 2236×600×1068 mm envelope as the frozen selected rack. It includes casters and leveling feet. This improves mechanical comparison, but different OEM/load capacity and accessories remain material gaps. The Supermicro MGX family 345 kg figure retained in Run 02 includes busbar and many reinforcing/accessory components; the figures cannot be bracketed as a target steel interval or added together.

The 188,938,199-byte MGX compute reference STEP provides a 460,939.9599409 mm³ main tray chassis volume. Many internal names contain GB2 identifiers and do not identify the selected GB300 revision. Material identity is not established by STEP defaults or a part-name suffix. Assigning that entire volume steel is therefore recorded as a rejected target material conversion. The material workflow can use this geometry when authoritative selected sheet-alloy evidence becomes available.

The GB300 coldplate supplier 9.5 kg module and matched Lenovo 29 kg compute/17 kg switch trays remain useful gross context. They overlap their nested coldplates, electronics and casings. No strong measured mass split was obtained for coldplates, PSUs, NVLink cable cartridges, internal cables or PCBs. The known false 22 kg Supermicro tray datasheet remains rejected on the Run 02 mismatch evidence; it was not revived. A100 teardown transfers remain rejected because air thermal hardware and accelerator packaging differ. Public REEL MGX listings were inspected as leads but hide quantities and include manufacturing-yield inputs; they cannot be treated as product embodied material declarations.

Highest B1 research leverage remains the 122,143-equivalent coldplate/compute-tray class and 325,714-equivalent PSU class for copper, then frame and tray sheet metal for steel. The busbar result demonstrates how CAD can create a valid reference coefficient, but its narrow stock contribution does not bound those unresolved classes. Complete B1 Cu/Al/steel closure remains OPEN.

Reproduction

python research/run03/compute/reproduce_compute.py runs the offline density, stock, residual and independent catalogue checks. reference_model_results.json is the exact captured output. For full geometry validation, unzip the preserved rack archive outside the Git tree, then run extract_cad_solids.py /tmp/<rack STEP> --output /tmp/run03-solids --report /tmp/run03-solid-validation.json. Compare volumes and solid hashes with cad_solid_validation.json; only the temporary paths differ. No macros, firmware, executable vendor payloads or vendor messages were used.

proposed_forensics.json contains one engine-ready B1 conditional scenario, keyed to the canonical R15-RACK-EQUIVALENTS plant ledger. Its two physical components use dimensional product models, separate catalogue-mass validations, and exact CAD references. REFERENCE_SCENARIO_PROTOCOL.md records the hypothetical two-assembly/two-body mapping with its own source and claims. The integrating Run 03 validator accepts the row; both models calibrate within the 1% validation threshold while transfer_admitted remains false. verification.json records the checks. This file set makes the scenario available to CLI reporting while preserving all primary target gaps.

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: research/run03/compute/HANDOFF.md
SHA-256: e71333be8a69acfcde4365f925e63b858fe4ba55fbfa27448a4e4ffc3da57c80
Download research file

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