Hub Dollars — V3 coverage review, 18 October 2026

The selected-family total now includes two independently checked additions:
Mistral Large 3 base: 41B active × assumed 10/18/30T tokens -> 1.968 / 7.9704 / 29.52 MUSD.
Nemotron Nano 2 base: 12.31B × reported 20T tokens -> 1.18176 / 2.65896 / 5.9088 MUSD.
Nano 2 is below the 30B priority threshold but remains in the overall scope.
These are rental-equivalent 6ND estimates, not reported invoices. Sources in reviewed-additions.json.
52 selected families: 284.61974 / 844.33319 / 3496.722795 MUSD.
Documented-input central subset: 462.082948 MUSD. Prices and FLOP proxies remain assumptions.
Organization figures remain partial; Cosmos and other omissions are NOT included in this total.

Claude's new inventory is valuable, but its claim of complete per-model estimates is not accepted:
- Actual registry has 694 unique rows, not 694 plus five; 186 have no central estimate.
- Family totals were assigned to a single arbitrarily chosen repository; these are not model budgets.
- Equal parameter counts or architecture do not establish copied weights or shared training.
- Zero for suspected copies/merges can hide distillation, routing or subsequent training.
- Declared-parent and variant filters overlap, so subtraction was not a valid coverage funnel.
- BF16/HF suffixes do not establish a redundant format and must not exclude sole canonical releases.
- Curated producer lists, missing size metadata and download eligibility prevent exhaustive claims.

The revised registry retains ALL 39,572 >=30B IDs in the snapshot, plus 81 catalog/audit candidates.
39,653 rows have explicit CONDITIONAL scenarios. 1,453 producer-list/audited rows from 124 namespaces
are searchable in the interface. These are repositories, NOT distinct models/training runs.
Base scenario: 6*N_active*D with year-based token priors and prices 0.8/1.8/4 MUSD per 1e24 FLOP.
Increment-only scenario: 20/100/500B additional tokens with the same price assumptions.
Repository creation year is only a weak proxy for a training generation.
The two scenarios are alternatives, NOT additive. Inherited parent training is excluded from the
increment scenario, not asserted to be zero. For unresolved active counts, the number is explicitly
DENSE-EQUIVALENT, not a defensible MoE estimate. Generic proxy unsuitable for some video/hybrid models.
Exact source-reviewed model components are retained separately from generic scenarios where mapped.
No registry aggregate is published until training-event identities and overlaps are reviewed.
No shape-based ancestry is accepted; metadata parent declarations remain claims to verify.
Format labels are candidates, not exclusions. Community rows are retained in full downloadable files.

Cosmos 3: primary report verifies token counts; runtime remains inferred. The reported throughput is a
joint image/video benchmark, and GPU counts conflict between prose and table caption. Image/video rates
may be concurrent; treating them as exclusive alternatives is provisional. The sensitivity file is a
research diagnostic, not an accepted dollar estimate or a confidence interval. Do not add Claude's
27M central figure or its full 465M omissions sum to the total. Cosmos 1 cluster capacity similarly is
not measured per-model training spend. Keep omitted mass separate and unresolved.

Next work: verify training lineages for large producer releases, replace generic token priors with
primary budgets, reconcile family membership with individual training events, and resolve missing sizes.