LearnNewsExamplesServices
Frontmatter
id14687
titleServing-cost measurement program: inference duty cycle + hardware-option economics (no numbers until measured)
stateOpen
labels
enhancementaiarchitecture
assigneesneo-opus-grace
createdAtJul 4, 2026, 6:16 AM
updatedAt1:54 AM
githubUrlhttps://github.com/neomjs/neo/issues/14687
authorneo-fable-clio
commentsCount0
parentIssue14442
subIssues
15014 Serving-cost meter: deterministic endpoint-owner sampling and provenance report
subIssuesCompleted1
subIssuesTotal1
contentTrust
projected
quarantined0
signals[]
blockedBy[]
blocking[]

Serving-cost measurement program: inference duty cycle + hardware-option economics (no numbers until measured)

Open Backlog/active-chunk-3 enhancementaiarchitecture
neo-fable-clio
neo-fable-clio commented on Jul 4, 2026, 6:16 AM

Context

Operator-mandated research V-B-A (2026-07-04): every cost conversation tonight ran on gut-feel figures (his flagged first; the author's struck second — D#14683 correction on record). The rule going forward: cost claims are [UNMEASURED] until a named measurement exists. A chat model + an embedding model resident around the clock is a real unquantified load; whether an own server beats cloud beats customer metal is unknown.

The Problem

Three economic unknowns block honest pricing AND honest architecture choices (they are research items, not modeling exercises): (1) the always-on inference duty cycle — what one institution-day consumes (chat + embedding, idle vs active phases) on reference hardware; (2) hardware-option economics — own server vs cloud instance vs customer-provided metal, measured under the same workload; (3) the actual current serving bill for the deployed middleware/front-door surface (read the console, not the vibes).

The Architectural Reality

The measurement substrate half-exists: the #14672 fresh-install harness's phase-log pattern extends naturally to steady-state metering (resource sampling around the orchestrator + provider endpoints — the localModels/openAiCompatible seams in the config template are the instrumentation points); METRIC-node emission (#14430 discipline) gives every figure its falsifying query. Boundary: measurement METHOD + reference-hardware results are public-safe; pricing derivations from them live in the private substrate only (standing rule).

The Fix

One PR: a steady-state metering harness (samples compute/memory/token-throughput around the resident models over an N-hour institution-day window on named reference hardware) + the three measurement runs recorded as METRIC nodes with falsifying queries + a one-page results doc (numbers with provenance, zero extrapolation).

Acceptance Criteria

  • Duty-cycle measurement: one institution-day on reference hardware, idle/active phases separated, provenance-stamped.
  • Same workload measured on ≥2 hardware options (whichever are actually available — availability documented, no modeled third options).
  • The live middleware serving bill read and recorded (console evidence).
  • Every published figure carries its falsifyingQuery; NO derived pricing in public substrate.
  • Cross-family review.

Out of Scope

Pricing decisions (private substrate, L6 orbit) · optimizing the loads (measure first) · cloud-tenant product architecture (#14456's world).

Related

#14672/#14676 (the instrumentation pattern this extends) · #14442 (the private-side consumer) · D#14683 (the correction that mandated this) · #14430 (the discipline). Sweeps: no equivalent ticket (searches 04:02Z + this hour's filings). Origin Session ID: fa2a6fd5-7488-4af6-a0d2-3855c86003e4 Retrieval Hint: "serving cost measurement duty cycle hardware economics unmeasured"

tobiu referenced in commit ba6645a - "feat(ai): serving-cost meter — the measurement instrument for the always-on inference load (#14687) (#15013) on 3:00 AM