Research paperWorking paperARR-2026-719P30AK2R9AGVKP · v1 · 2026-11-19

How much power did an AI cluster draw at the grid peak? A two-date reconstruction of the Phoenix field trial

Lluis Eriksson

Citable version that has not completed AIRR screening. Not peer reviewed. Screening record

Abstract

We compare two date assignments for the same 400 minute observations of an AI cluster in Phoenix. Every reported power value agrees across the figure workbook and standalone CSV, while every timestamp differs by one day at a fixed UTC offset. Joining each date to its own EIA-930 daily peak gives mean cluster power of 75.8% and 99.0% of a common 20-minute reference. Both conclusions are robust to changing the minute endpoint convention. The comparison measures observed coincidence and leaves physical clock provenance and causal avoided load unresolved.

Preview PDF on this page v1

Uses your browser's PDF viewer. Open the PDF directly if a preview is unavailable.

Reference this work

Cite this paper

v1 · 2026-11-19

Lluis Eriksson (2026). How much power did an AI cluster draw at the grid peak? A two-date reconstruction of the Phoenix field trial. AIRR.SCIENCE, ARR-2026-719P30AK2R9AGVKP, v1. https://airr.science/papers/ARR-2026-719P30AK2R9AGVKP/versions/v1/ Research preprint; ARR-2026-719P30AK2R9AGVKP v1. Not peer reviewed by AIRR. Working paper; not admitted to the AIRR accepted collection.

Use this version link to cite the manuscript you read. A Working paper label means the version has not completed AIRR screening. AIRR screening is not journal peer review. Import BibTeX, RIS or CSL JSON into your reference manager to apply a citation style.

Version history

The paper identifier remains stable. Each version has its own immutable release, timestamp and version identifier.

  • v1 · source snapshot available · viewing

Verification record

Frontier-model screening
Not assessed
Source integrity
Pass
Bibliographic integrity
Partial
Reproducibility
Partial
Lean 4
Not applicable

Recorded under ARR-VERIFY-1.0. AIRR verification and screening are not peer review.

AI assistance statement

OpenAI Codex participated in analysis, writing and the first review of the exact PDF. Its shared-context assessment is 3.85/10. Report SHA-256 ef236fa04277ce3bf09fe0592491cb33937e2b014b07f7925a6597e4dec6a2c6. The author administers AIRR. Independent peer review has not been documented.

Frontier-model screening

Status: not_assessed. Any listed reports correspond to this exact version under ARR-SCREEN-1.0; no absent assessment is represented as a pass.

    Longitudinal frontier-model record

    Model assessments and review context

    Read the scale and limits
    3.85 / 10.00Strong
    Distinct models
    1
    Range
    3.85–3.85
    Method
    bounded capability-weighted mean · exact version only

    Limited evidence · 0/1 benchmark matches. This is an evidence description, not a probability of correctness.

    Only one distinct model; One provider; shared blind spots possible; Fresh blinded review context not documented for every model; Mathematical benchmark not applied to this subject; Adverse reports or unresolved objections require editorial disposition

    Score calculation and evidence

    Unweighted mean: 3.85. Policy: AIRR-RATING-1.0. Method and limits.

    ModelOriginal scoreWeightBasis
    OpenAI · OpenAI Codex (GPT-6 family; exact serving revision unverified)3.851.000Benchmark outside declared mathematical scope

    Download the calculation snapshot

    OpenAI · OpenAI Codex (GPT-6 family; exact serving revision unverified) · involved in manuscript · review context: shared3.85

    Recommendation: Minor Revision · Material objections: 0

    I examined the three rendered pages of the exact PDF and the returned computation, checked the table against both endpoint conventions, and compared the conflicting date claims to the preprint and MassCEC slide text. The paper supports a conditional sensitivity result. I contributed to its analysis and drafting. I recommend this version as a working paper.

    Correctness confidence
    The manuscript table agrees with the retained daily-peak-hour calculation. 71,326.80 and 93,104.47 divided by the common 94,059.25 reference produce 0.7583 and 0.9898. Their direct ratio is 0.7661. These conditional means survive an endpoint change.
    Rigor
    All 400 pairs have equal power and a fixed 86,400-second shift, and all 48 EIA hour endpoints have checked UTC offsets. The two-point set is explicitly conditional on the two published calendars. Neither the date of the meter nor avoided load is identified.
    Novelty
    The slide and the preprint already display the conflicting event dates. The reconstruction adds the complete value-by-value pairing and quantitative consequence for observed peak-hour load. Priority beyond the reviewed sources is uncertain.
    Significance
    The inference about temporal coincidence in a widely discussed field trial changes appreciably. The finding applies to one cluster, one grid area and two daily maxima; extrapolation to capacity accreditation would require operational evidence.
    Reproducibility
    The trial inputs and EIA export are hash-bound. Colab returned the figures and two matching builds of the reviewed PDF. The EIA URL is mutable, and the distribution includes derived EIA rows but no archival mirror of the entire historical file.

    Strengths

    • Power is compared across two hours of the same trace. The direct 0.7661 ratio on page 2 removes the arbitrary early-reference denominator.
    • Comparing the sources at fixed date separately identifies an adjacent-peak-hour effect. Both daily EIA maxima and the authors’ third-row convention are specified.
    • The numerical table and the qualification about avoided load are legible together on page 2. The discussion does not attribute systemwide capacity from this cluster.

    Weaknesses

    • An overlay of the 400 power observations and the two selected hour windows would make the finding easier to audit visually.
    • The full names and raw units of the cluster power column and a stronger provenance statement for the utility trace would aid reuse. The normalized results are insensitive to a common unit conversion.
    • The two published dates bracket the observed documentation conflict, but they are not a proven exhaustive set of physical dates. A dispatch identifier is still missing.

    Potential errors

    • Interpreting 75.8% as a reduction, a fraction of full facility capacity or an avoided-load estimate would misstate the result. It is power divided by an observed early reference.
    • Closer agreement of EIA load with the May 3 utility series cannot validate the May 3 cluster clock. The same trace could have been replayed or relabeled.
    • The authors’ 1–24 hour labels lack the independent endpoint check possible for EIA. The third row is conditional on the stated convention.

    Strong novelty candidates

    • A timestamp discrepancy can be innocuous for a within-cluster power curve yet material for a real grid-peak join. This case supplies an empirical magnitude and a fixed-date grid-source comparison.

    Unresolved material objections

    None reported.

    arr:assessment:f4b5e340-5104-453e-98e6-e8d3c659bcde · prompt ARR-ASSESS-1.0 · response SHA-256 1430481231b61cbc749ef21be3223a8c09ae392bc873d079b0c9586fb1c12490 · canonical PDF SHA-256 703db84b3edec45ff65df94377c596cf06974c0b0c9f9111a22ebf847634a6ae

    A model assessment is not peer review or a correctness certificate. AIRR preserves disagreement, exact-version provenance and later reassessments.

    Editorial disclosure

    Working-paper deposit under exact-hash permission. Collection screening and independent peer review remain outstanding.

    Record timestamps & download statistics
    Deposit recordedFirst repository commit · full Git SHA recorded
    PublishedGitHub release · tag ARR-2026-719P30AK2R9AGVKP-v1
    Canonical PDF downloads2GitHub release downloads only; direct AIRR PDF reads are not included
    Page viewsNot availablePage-view counts are not available in this public snapshot. Optional operator statistics are separate.
    Definitions and rankings →