Who publishes this, and under what rules.
GII borrows the discipline of benchmark governance without borrowing anyone's independence claims. The Grid publishes these series, says so plainly, and makes every input public so the numbers can be checked.
The object, the use, the publisher
- Economic object. Public list-price notional daily spend on routed LLM inference, currently scoped to OpenRouter public top-50 traffic.
- Intended use. Research, market monitoring, editorial citation, and OTC framing. E0/E1 series are not suitable for contract settlement.
- Publisher. The Grid (thegrid.ai), which operates a real-time spot market for AI inference. This is a first-party publication and is labeled as such everywhere.
- Contact. Complaints, correction reports, and citation questions: [email protected].
A market operator, measuring an adjacent market.
The Grid runs a spot market for standardized AI inference. GII currently measures a different venue (OpenRouter public traffic), but the conflict is structural and permanent: a market operator publishing reference numbers for its own industry. GII's answer is not an independence claim; it is reproducibility. Public inputs, open methodology, versioned snapshots, content-hashed prints, and a recomputation package (shared on request) that runs from raw data to headline number. If The Grid's own venue data ever enters a published composite, that inclusion follows the universe-expansion protocol below and is flagged as self-referential with independent review.
The evidence ladder.
Every number on this site carries a tier badge. The tier states what the underlying spend measure can credibly prove, and therefore what contract role it can support. The ladder runs from public notional estimates (easy to publish, weakest claim) up to finance-reconciled realized spend (hardest to produce, the only candidate for settlement). GII publishes E0/E1 today and treats the upper rungs as its governed roadmap. Every published series carries its scope label and tier.
| Tier | Label | Input basis | Suitable for | Status |
|---|---|---|---|---|
| E0 | Public notional | Public volumes x public list tariffs | Research, monitoring, event-style questions, OTC framing | LIVE |
| E1 | Historical public notional | Public volumes x dated historical tariffs | Research backcasts and sensitivity analysis | LIVE |
| E2 | Transaction-priced venue | Consumed or purchased venue units at executed prices | Venue-scoped diligence and basis studies | ROADMAP |
| E3 | Attested venue aggregate | Venue-supplied aggregates with source controls | Private benchmark pilots and OTC diligence | ROADMAP |
| E4 | Finance-reconciled realized | Spend reconciled to invoices, ledgers, or settlement records | Candidate settlement input after governance review | ROADMAP |
Corrections and restatement policy
- Versioned prints. A print is never edited in place. A corrected print is published as a new version with a new content hash; prior versions remain archived.
- Error classes. Input errors (source restated upstream), mapping errors (slug/tariff joins), and calculation errors (code defects) are labeled distinctly in correction notices.
- Materiality. Corrections that move the headline print by more than 1% are announced on this page with the affected dates and both hashes. Smaller corrections appear in the data manifest.
- Restatement record to date: zero restatements across the archived snapshot history. The pipeline records a new version only when re-pulled content differs.
Methodology change policy
- Announced effective dates. Methodology changes take effect on a stated future date. No retroactive headline changes, ever.
- Pro forma backcasts under new methodology are published as labeled research series, never as silent restatements.
- Versioned specs. The methodology identifier
(
gii-v1) is bound to the versioned specification text. A change producing different numbers requires a new version identifier.
Source-inclusion protocol
The v1 series is venue-scoped by design. Before any multi-source expansion, the family pre-commits to:
- Eligibility review with an announced effective date, and a hard non-overlap rule: a source qualifies only if its traffic provably does not double-count an included source, or an explicit netting adjustment is published.
- Raw spend jumps by design; the chain-linked index does not. The chain-linked index uses the same universe in numerator and denominator across each link date.
- Transition dual-reporting: official composite, composite ex-new-source, and the new source's contribution.
- Self-referential inclusion (the publisher adding a venue it operates) is a manipulation channel and requires independent review under the settlement-grade governance roadmap before it can occur.
Cessation and renaming
A benchmark name must survive for the life of contracts that reference it. GII series identifiers are versioned in the data manifest; if a series is ever renamed, the manifest maps old identifiers to new ones with no numeric break. If a series is discontinued, GII commits to a minimum 90-day notice period, a stated replacement or fallback series where one exists, and permanent archival of all published prints.
The settlement-grade path
The manifesto pillar GII takes seriously: verification by independent monitors. Today's series are first-party and E0/E1. The published roadmap to settlement-grade series requires realized or attested spend (E3/E4), provenance attestation rather than plausibility screens, surveillance of settlement windows, and independent oversight of inclusion and correction decisions. The benchmark governance memo and settlement-grade specifications are described on the data page and shared on request.
Build against the print.
Download the public artifacts, request the recomputation package, or talk to The Grid about moving from public notional to verified spend.