← NM AI Research · research portfolio

Announced vs. Deliverable AI Power Demand

How much announced capacity is actually deliverable: on both sides of the meter, the generation that gets built (the PJM completion anchor) and the data-centre load that survives grid operators' screens (PJM + ERCOT). Announced is a scenario, not a baseline.

Interactive front-end over published aggregates · v1 (combines supply and load sides) · canonical concept: DOI 10.5281/zenodo.20559430 · v1 paper: DOI 10.5281/zenodo.21971645 · NM AI Research, ORCID 0009-0003-4213-7769 · CC BY 4.0

GW
De-duplication (r1)removes literal duplicate filings
Viability survival (r2)entered → signed interconnection agreement
Build rate (r3)signed agreement → in service

The completion funnel (2010–2017 cohort, 204 GW entered)

Each milestone is conditional on surviving the previous one, so the rates multiply. The steepest drop is at the System Impact Study, where a project learns its share of network-upgrade costs.

Stage% of enteredwhat it is

Where the MW die: 61% of announced MW exit before a signed agreement; another 13% exit after one (a signed IA converted to service only ~58% of the time). The gap is viability and build, not paperwork.

Realized completion by submitted cohort

Recent cohorts read low only because they have not had time to resolve, so the headline uses the mature, near-fully-resolved windows (bold). Those converge on ~16–25% built, about one announced MW in five.

Cohort% built% withdrawnmaturity

On the load side the operators deflate the headline themselves and publish the method. This view does not deflate the load: it places PJM's and ERCOT's own published vetting side by side, and brackets the result. The contribution is the cross-region, reproducible packaging, not the deflation.

PJM: the published large-load funnel (2030 horizon)

PJM accepts a request into its forecast only after a documented haircut (50% default non-firm probability, 70% utilization, 70% capacity-to-demand, minimum 36-month ramp); only the Firm subset (backed by an Electric Service Obligation or Construction Commitment) counts toward capacity. Its worked example survives at ≈48% of requested capacity.

Stage2030 GW% of submittedwhat it is

ERCOT: the cross-region check

ERCOT had almost no barrier to entry before its PGRR115 rules, so its queue was heavily padded. It publishes a realized de-rate: actual per-site peak consumption is 49.8% of requested MW. Three quantities must be kept apart: request-survival (≈50%), realized utilization (49.8%), and energized-to-date (≈2%, an age figure: the surge is too young to have built out, not a failure rate).

StageGW% of requestedwhat it is

The two-rail energisation bracket

No matured data-centre load cohort has resolved yet, so the deliverable rate is bracketed, not measured. The band spans the regime change: the lower rail is the lightly-gated queue's historical completion; the upper rail is the operators' newly-vetted acceptance.

RailSurvivesbasis

Cross-region finding: two independent operators, PJM applying a forward de-rate (≈48%) and ERCOT measuring a realized one (49.8%), land near half of requested data-centre demand surviving the first credible screen, and both now gate the forecast on an executed agreement. The two ≈50% figures are constructed differently (PJM's bundles a completion-probability factor; ERCOT's is conditional on the project being built), so the resulting claim is structural, not a numeric identity.

What this is. A realized announced-to-in-service completion rate from PJM's own resolved serial queue (8,253 generation projects, ~837 GW announced). The calculator lets you deflate any announced headline through the three measured attrition stages and watch the deliverable move. Defaults are the PJM mature-cohort base rate; the sliders are for sensitivity, not invention.

The load-bearing assumption. The mature 2010–2017 cohort is generic historical generation (wind/solar/gas) used as a completion base rate, not capacity announced to serve AI. Transferring that rate to the AI-era cohort is the named assumption (cohort comparability); FERC Order 2023 could raise completion, speculation could lower it.

Both sides of the meter (v1). The calculator and PJM evidence deflate the supply side (generation that gets built); de-duplication barely moves it (r1 ≈ 0.81–0.92): the gap is overwhelmingly viability and build, and "Built" counts only In Service (conservative). The Data-centre load view adds the demand side, where the operators do the deflation themselves: this tool packages PJM's and ERCOT's own published vetting and brackets it, rather than deflating the ~950 TWh headline itself. The two ≈50% first-screen figures (PJM applied, ERCOT realized) are constructed differently, so the cross-region claim is structural. Every output is a band with its softest input named: treat any headline as a scenario, not a baseline.

Guardrails (read before citing)

No raw PJM data is bundled here (it carries a redistribution restriction); only the published derived aggregates. The full method reads a locally-obtained PJM export; see the Zenodo record. r2 × r3 = built/entered is an algebraic identity, not an independent check. Anchors (LBNL ~14–20% national; Exelon ~22% of company pipeline) are loosely comparable, not like-for-like.

Disclosure. No third party reviewed, funded, or directed this work. Independent analysis, not investment advice.

Reproducibility. This page embeds the published aggregates (funnel.csv, cohorts.csv, and load-side dc_funnel.csv, ercot_funnel.csv) verbatim, regenerated by build.py; full reproduce.py and write-up live in the Zenodo record.