What AI Is Made Of

Mass Balance of Intelligence: Run 01 findings

Final snapshot: SNAP-RUN01-REVIEWED · evidence cutoff/access: 5 October 2026 · fixed reference: CFG-GB300-2026Q4.

Run 01 extends the equipment and production chain, but cannot close the selected-reference physical material balance. It establishes more rack components, finished GPU assemblies and embedded logic dies; obtains product-specific material declarations for several upstream equipment classes; and identifies precisely what prevents those declarations from becoming tonnes/GW. The main blockers are a selected electrical/cooling design, qualified component quantities and material BOMs, and semiconductor process yields. A small, explicit CDU design sensitivity demonstrates a reproducible material roll-up. Its mass is confined to those CDU products.

The README (repository-only reference), supplied Accounting Standard v0.1 (repository-only reference), complete Run 00 findings (repository-only reference), schemas, source register, ledger, reference, dependency graph, calculations and tests were inspected before extension. All 51 original tests passed. Every original central result reproduced exactly; the full 80-row balance and all 289 original canonical entities still match the preflight fingerprints (repository-only reference). No reference hardware, PUE, reserve or load-factor coefficient was replaced.

The repository now contains 75 sources, 258 claims and 269 retained ledger observations; the reviewed snapshot selects 263 ledger records, including 136 unknowns. There are 11 explicit evidence conflicts, 37 material-component relations, and three current physical-path views. All 74 captured-source hashes match. Source counts include protocols and repeated-source lineage; they are not independent corroboration counts. All 118 tests pass.

1. Fixed reference and physical requirements

The reference remains the Supermicro SRS-GB300-NVL72-M1/M0, with 72 B300 accelerators, 36 Grace CPUs, 18 compute trays, nine NVLink switch trays and 72 ConnectX-8 NICs per rack. It is a 2026-Q4 frontier training/research reference, normalized to 1 GW IT electrical capacity. It does not describe a measured campus or an accepted single 1 GW training fabric.

Run 00 assumptions remain unchanged: extra-rack IT reserve 2%/5%/10%; rack input 132/140/140 kW; PUE 1.1/1.2/1.3; electrical load factor 80%/90%/100%. The OEM's 132–140 kW power interval still has an unresolved meter/workload/CDU boundary. Low/high are arithmetic envelopes, not confidence intervals or jointly optimized builds. Counts are fractional normalized equivalents, not purchase orders. The complete reproduced physical table (repository-only reference) supplies low/central/high and ledger IDs.

QuantityLowCentralHighUnit / boundaryInterpretation
Rack equivalents6,4296,7867,424racks/GW, B1Unchanged conditional normalization
B300 accelerators462,857488,571534,545units/GW, B1Unchanged
Grace CPUs231,429244,286267,273units/GW, B1Unchanged
Compute trays115,714122,143133,636units/GW, B1Unchanged
NVLink switch trays57,85761,07166,818units/GW, B1Unchanged
ConnectX-8 NICs462,857488,571534,545units/GW, B1Unchanged
Physical HBM133.3140.7153.9million manufacturer GB/GW, B1Physical label retained; usable bytes unresolved
Power shelves51,42954,28659,394units/GW, B1New: eight/rack
PSU modules308,571325,714356,364units/GW, B1New: six/shelf, 48/rack
Finished GPU assemblies462,857488,571534,545packages/GW, B4 requirementNew assembly-count view of existing GPUs
Embedded good GPU logic dies925,714977,1431,069,091dies/GW, B4 requirementNew: two logic dies/GPU
Facility electrical input1.101.201.30GW, cumulative B2Unchanged PUE sensitivity
Annual facility electricity7.719.4611.39TWh/year, B2 flowExact central 9.4608
Approximate steady IT heat1,0001,0001,000MW thermal, B1 attributionExplicit local heat-conversion approximation

The Supermicro rack datasheet and NVIDIA Blackwell Ultra technical description support the hardware coefficients (Grade B). Their per-GW descendants retain Run 00's assumption ancestry and remain sensitivities with no empirical evidence grade. Finished GPU assemblies and their embedded dies are nested inventory views, not quantities to add together.

2. Electrical architecture: what is required, what is selected

The generated electrical path (repository-only reference) connects enabling supply/HV interface, HV/MV substation, facility MV/LV distribution, protection or buffer choice, PDU/busway/cables and rack power shelves. It has separate facility and backup branches. The equipment classes are supported by inspected primary designs; voltage, ratings, ownership, redundancy and installed quantities remain unselected. Dashed upstream routes show this conditional topology. A protection/buffer class does not select UPS or battery chemistry.

Vertiv's RD030 (repository-only reference), revision August 2025, provides a real GB300 reference one-line: 64 DGX compute racks at 142 kW plus 64 support racks at 14 kW, totaling 9.984 MW IT; four-to-make-three power; eight 2 MVA IT UPS units; separate 2N facility UPS; sixteen 1 MVA PDUs. These are Grade B engineering context. The rack architecture and support allowance differ from the fixed Supermicro reference. Those equipment counts are not scaled to our GW. UPS MVA, protected IT MW, facility MW and utility connection capacity remain distinct. RD030 also contradicts itself about RPP count; the disagreement remains recorded.

A Virginia SCC Edsall order (repository-only reference), 8 August 2025, approves four 84 MVA, 230/34.5 kV transformers, a four-breaker 230 kV ring bus, and a 0.9-mile connection for a projected 176 MW campus plus local/future loads. It establishes a real regional topology and shared allocation problem, not a universal transformer/GW coefficient. The 336 MVA aggregate nameplate cannot be divided by customer MW without power factor, contingency, load allocation and expansion reconciliation. Approved future in-service dates are plans, not delivered equipment.

Selected-reference electrical requirementRun 01 resultMissing evidence
Utility connection and substation MVAUNKNOWNParcel, utility study, power factor, firm-contingency and allocation plan
HV/MV and facility MV/LV transformersUNKNOWN count/ratingAccepted one-line, voltage, duty, redundancy and losses
MV switchgear/LV switchboardsUNKNOWNFeeder count, voltage/current/fault ratings, transfer layout
IT/facility UPS, batteriesUNKNOWNProtection strategy, ride-through, redundancy, chemistry and selected BOM
PDU/busway/cable quantitiesUNKNOWNBranch topology, voltage/current, route lengths and conductor choices
Backup generation/storage/fuelUNKNOWNSelected supply/reliability route and its attributable scope
Electrical lossesUNKNOWN allocationEfficiency curves and operating profile, reconciled inside PUE

One factual interpretation was corrected explicitly: Run 00's power-shelf claim limitation treated “shared 4+4” as demonstrated whole-rack redundancy yielding 132 kW usable. The source does not establish that operating/failure interpretation. Eight 33 kW shelves have 264 kW aggregate module nameplate; the same rack sheet labels total 132 kW and operating input up to 140 kW. Run 01 preserves the old claim, records the conflict, and keeps usable rating unresolved. It leaves the normalization unchanged.

The separate Supermicro shelf sheet (repository-only reference) identifies 50 V DC output, six named PSU modules, 33 kW at 6+0 or 27.5 kW at 5+1, peak 97.5% efficiency and capacitor buffering. Module-level modes do not settle rack-level redundancy. Peak efficiency is not operating loss, and integrated capacitors do not establish a battery-material BOM. Whole-shelf material mass would include its modules; those cannot be added again without reconciliation.

3. Cooling, heat and water

The generated cooling flow (repository-only reference) preserves direct liquid cooling of high-power components and the external liquid-to-liquid CDU route. It separates residual air, the facility loop, electrical parasitics, final rejection, source withdrawal, circulating coolant, consumption and usable return. No cold-plate count follows directly from GPU/CPU counts: a plate can cover multiple devices.

Liquid heat fraction, air fraction, temperature/pressure/flow requirements, selected pumps, pipe lengths, coolant inventory and final heat-rejection plant remain unknown. Supermicro portfolio “up to 98%” heat capture and a Vertiv DGX thermal split are different architecture evidence; neither becomes the exact Supermicro fraction. The persisted thermal check therefore reports an open generation/transport/rejection balance rather than claiming closure.

Supermicro's GB300 page describes an external CDU option up to 1.8 MW and eight racks. A separate in-row SKU sheet gives up to 2.2 MW, 1,560 kg dry mass, 20 kW maximum electrical power and N+1 internal pumps. SKU/revision/rating conditions are unresolved. Run 01 does not multiply the eight-rack grouping by the other SKU's mass. An explicitly isolated grouping sensitivity gives 804/849/929 whole CDU units, with zero additional whole-unit spares only for that calculation; required redundancy remains unknown (R1-C-S-CDU8).

A Carrier verified CDU EPD (repository-only reference) gives a useful second coefficient chain: a named external liquid-to-liquid unit rated 1.3 MW thermal at 4°C approach temperature, 13.5 kW rated electrical input and 1,392 kg installed product mass. Its six categories reconcile exactly, with packaging separate. The unit's qualification for the reference remains unresolved.

Isolated Carrier N-only sensitivity: provision the full 1 GW IT heat envelope with one independent train and no whole-unit spares, at that declared rating. ceil(1000/1.3) gives 770 units. This models installed nameplate provision, not actual liquid capture, operating parasitics, heat rejection or available production. The scenario protocol (repository-only reference) and named formulas preserve each parent.

CDU-only design-option materialQuantityMeaning
Stainless steel577.215 tonnesDeclared alloy category
Steel328.513 tonnesSeparate declared category
Cast iron158.474 tonnesKept separate from steel
Other undecomposed material3.311 tonnesNo element assignment
Plastics/nonmetals2.387 tonnesDeclared category
Electronics1.940 tonnesNo copper/trace-element decomposition
Total installed CDU product mass1,071.840 tonnesExact sum of those product categories
CDU electrical nameplate10.395 MWWithin PUE overhead if selected; no annual-energy claim

These are one design-option subtotal, not the reference's incremental B2 total. Air handling, facility pumps, piping, coolant, final rejection, building and redundancy additions are excluded. Copper and aluminum are not zero just because the EPD lacks explicit categories. The option uses the existing IT architecture and a compatible functional cooling class; it does not establish a qualified installation.

Supermicro's 2023 tower sheet (repository-only reference) describes a closed process loop with a separate evaporative spray circuit. Closed circulation therefore does not establish zero consumption. Capacity depends on wet-bulb and operating temperatures. Marketing gallons/minute/MW with incomplete conditions and a newer system-wide 1–50 MW range do not supply an annual Virginia water coefficient or an individual tower count.

Annual source withdrawal, consumption, usable return and WUE remain UNKNOWN. No source/basin/parcel, seasonal profile, heat-rejection route, cycles of concentration, blowdown treatment or return conditions have been selected. No water/query coefficient is investigated. Indirect generation water and dedicated source infrastructure also remain unallocated.

4. Material BOM and reconciliation

The component reconciliation (repository-only reference) and machine-readable roll-up (repository-only reference) distinguish coefficient, installed quantity, material result, boundary, source category and design option. The named material_component calculation is component_count × kg_per_unit. Unselected counts propagate null. A quantified subtotal without a selected component is null, not zero; completeness has no invented percentage denominator.

Product-specific evidenceReported material coefficientBoundary and applicability limits
Hitachi 40 MVA, 132/15.6 kV ONANCopper 5,796 kg; steel 36,573.53 kg/productB3 candidate, Italian 50 Hz product; selected US duty/count unknown
Hitachi 1.6 MVA dry transformerAluminum 1,081 kg; copper 30.8 kg/productB2 candidate, 2022 reference product, 2025 EPD; electrical steel/cast iron separate
ABB UniGear feeder panelCopper 35.22 kg; aluminum 10.57 kg/panelB2 candidate; breaker/sensors included; packaging separate; registration placeholder retained as a verification issue
Schneider Canalis KTA 2500 A configurationCopper 43.344 kg; aluminum 1,374.624 kgB2 context for exact 64 m sections plus bends/tapoffs/feed boxes and packaging; historical PEP; not kg/metre
Schneider Galaxy VXL 1,250 kW UPSExplicit copper category 8.466 kg; electronics 317.475 kg undecomposedB2 packaged-product context; external energy storage excluded; copper category is partial elemental coverage
Carrier external CDUInstalled mass 1,392 kg with six category coefficientsB2 candidate; actual reference quantity/qualification unknown; isolated N-only option above

All are Grade B technical declarations, not measured reference-project BOMs. Schneider OEM download access returned 403; the primary-authored PEP documents were inspected in distributor reproductions. Issuer, retrieval host, date and limitations remain distinct. Canalis's 2020 declaration validity expired in 2025; it is historical product evidence and triggers vintage warnings. No proxy acquires current-product applicability from its evidence grade.

Hitachi's eight listed 40 MVA material categories total 57,341.46 kg against declared 57,380 kg: 38.54 kg remains unassigned. The residual is retained, not distributed among metals. Four same-rating transformer variants have different copper masses; these are different products, not uncertainty endpoints. Whole equipment, packaging, copper categories, alloys and undecomposed electronics cannot be treated as interchangeable elemental assays.

Reference material scopeCopperAluminumSteel
B1 computeUNKNOWNUNKNOWNUNKNOWN
Incremental B2 facilityUNKNOWNUNKNOWNUNKNOWN
Incremental B3 enabling infrastructureUNKNOWNUNKNOWNUNKNOWN
Cumulative B2/B3 totalsUNKNOWNUNKNOWNUNKNOWN
B4 primary feedstock / extractionUNKNOWNUNKNOWNUNKNOWN

No reference tonnes/GW total is defensible. Missing B1 coverage includes rack frames, compute/switch trays, PSUs, busbars, boards, cables, NICs, cold plates and extra-rack fabric/storage. Missing B2 coverage includes selected transformers, switchgear, protection/storage, routes, cooling plant and structures. Missing B3 coverage includes allocated substations, transmission, generation and water infrastructure. B4 additionally requires process yields, scrap/recovery, recycled inputs and coproduct allocation. Prices, capex and revenue supply none of these physical conversions.

5. Semiconductor and manufacturing requirements

The semiconductor path (repository-only reference) distinguishes embedded finished-good inventory from manufacturing inputs and capacity. NVIDIA directly supports two logic dies/GPU. Reticle-sized does not disclose actual die area or gross usable dies per wafer. No proprietary area, guessed yield or silicon tonne coefficient is inserted.

The reviewed chain is:

embedded good logic dies
  / effective known-good-die retention through qualified assembly
  = required good-die input
  / (gross usable dies per wafer × fabrication yield)
  = logic wafer starts

Every factor has a ledger parent. The retention, gross-die map and fabrication yield are unknown, so starts remain null even if one of those later becomes known. Package starts likewise need assembly/test/rework evidence. CPU, host-memory, NIC and switch semiconductor chains remain additional missing coverage.

The accepted HBM stack/die chain remains unresolved because NVIDIA's technical article body says eight 12-high stacks while its correction says twelve stacks. Primary Micron/SK hynix 36 GB, 12-high products support a separate Grade C capacity-equivalent proxy: 288/36 = eight stack equivalents and 96 memory-die equivalents per GPU. Central arithmetic is about 3.909 million stack equivalents and 46.903 million memory-die equivalents. The reproduced proxy calculation (repository-only reference) retains source values and functions. It does not establish the B300 supplier/part, base/interface die count, DRAM wafer starts or qualification. A package footprint is not a DRAM die area; relative thinning is not absolute silicon thickness.

TSMC's 2025 annual report (repository-only reference) distinguishes company-wide 17–18 million 12-inch-equivalent wafer capacity, 15 million shipments, and planned 2026 capacity. These span nodes and wafer sizes. They do not establish available 4NP wafer starts. B300-specific fab/packaging allocation remains unknown. Primary CoWoS evidence connects silicon interconnects, copper and substrates to the production route, but provides no transferable B300 mass coefficient.

ChainRequirementIndustry capacity evidenceAvailable buyer allocation / delivery
B300 finished packagesConditional central 488,571 net assembliesPortfolio/node/packaging context; no qualified starts conversionUNKNOWN
Logic fabricationWafer starts UNKNOWNTSMC aggregate wafer equivalents; no selected 4NP shareUNKNOWN
HBM fabrication/stackingAccepted stacks/dies/wafers UNKNOWNSupplier product/expansion evidence; Run 00 allocation disclosures retainedUNKNOWN
Advanced packaging/substratesNet assembly view known; starts/BOM UNKNOWNShared CoWoS expansion/constraint evidenceUNKNOWN
Transformers/switchgearSelected counts/rating UNKNOWNReal operational factories and expansions; no SKU throughputUNKNOWN
CDUs/heat rejectionReference SKU/count UNKNOWNProduct ratings/EPDs; no production slotsUNKNOWN

Manufacturing geography is preserved without inventing batch origin. Evidence includes TSMC's Taiwan/global footprint and Blackwell-family production in Arizona; SK hynix Cheongju DRAM/HBM expansion; Micron Singapore/Idaho plans retained from Run 00; Hitachi South Boston, Virginia manufacturing and expansion; Eaton Wisconsin operations and South Carolina planned 2027 production; Siemens Fort Worth operational LV switchboards; and Carrier studied factories in Mexico, Spain and China. None allocates this order or proves qualified delivery.

Run 01 collected clock definitions and state distinctions, but does not force a commissioning date. June 2023 transformer lead-time interviews remain historical context, not October 2026 quotations. Edsall permit-to-target dates and manufacturing construction plans are separate from order-to-delivery and accepted capacity. The 16-edge Run 00 dependency graph and unresolved durations remain unchanged for Run 02.

6. Confidence, schema failures and next evidence

PortionQualitative assessment
Exact per-rack inventories, reported two-die architectureStrong primary specifications; conditional GW normalization remains explicit
EPD material coefficients for exact productsStrong technical evidence within the declared functional unit; partial categories and historical vintages matter
Electrical/cooling functional routesUseful engineering evidence; selected voltage, redundancy, counts and duty remain unresolved
Isolated CDU material sensitivityReproducible conditional example; qualification, derating and complete facility scope unresolved
Actual reference materials, water, wafer starts and buyer allocationCurrently unquantifiable from inspected evidence
Primary ore/feedstock requirementsUnquantifiable without process/recycling/allocation reconciliation

STANDARD_CHANGELOG.md (repository-only reference) documents ten evidence-driven proposals for v0.2: material identity; product/functional-unit applicability; separate design options; category versus elemental/packaging coverage; thermal/PUE semantics; explicit water flows; production stages and multiple yields; conflict relations; incremental/cumulative roll-ups; and nameplate/redundancy/allocation clocks. The governing v0.1 document is unchanged.

The software extends Git + canonical JSON + Python + generated SQLite. New conflict, material-component and physical-path relations reference the existing common ledger. Independent review exposed material/water relabeling, packaging transfer, lead-only evidence leaking into subtotals, and missing cross-geography/thermal checks. Fixes and adversarial tests are documented in RUN01_INDEPENDENT_REVIEW.md (repository-only reference). Review also appended explicit water-return/flow interpretations and assembly-retention parents. The actual initial built SNAP-RUN01 remains historically queryable; SNAP-RUN01-REVIEWED is the final report membership.

The 47-question unknowns queue (repository-only reference) identifies why each gap matters, current/missing evidence, likely resolver, effect, priority and falsifier. The highest-value next evidence is an accepted rack input/failure test and exact B1 material BOM; an owner-approved facility one-line with redundancy and route geometry; qualified cooling/heat-rejection design and source/return conditions; and B300 fabrication/assembly/stack yields plus reserved supplier allocation. These gaps are separate: selecting a design cannot supply a missing material assay, and obtaining an EPD cannot select its installed quantity.

Reproduce both runs with python scripts/reproduce.py; test with python -m unittest discover -s tests -v. The Run 01 artifact manifest (repository-only reference) and preservation check (repository-only reference) make the resulting coverage and limits auditable.

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_01_FINDINGS.md
SHA-256: 9bb56a7afd10785081f6896f2253a36fd4e30090db938f90f3f19396bcf24066
Download research file

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