AI for AI: Optimizing Additional Infrastructure Build-out to Power Artificial Intelligence Data Centers
Alexander Crosier, Kyle Onghai, Ronnie Sircar
q-fin.GN · cs.AI · econ.GN
Quick verdict
Models ERCOT AI load vs grid build-out via stochastic control; cannibalization deters investment. Sound math; illustrative costs, monopoly framing limit force.
Summary
The paper builds a stylized electricity-market model calibrated to ERCOT/Texas (reference price $30/MWh, traditional demand 55 GW, data-center demand 8 GW) in which price clears supply against two demand groups with distinct price elasticities (choke prices $70 and $150/MWh). A deterministic analysis shows that with no supply growth, projected 6 GW/year data-center load doubles average wholesale prices from $30 to $60/MWh within a decade, and traditional consumers would be priced out only after ~22.3 years. Replacing smooth growth with compound-Poisson arrivals (lognormal increments, mean 0.225 GW, fitted with a goodness-of-fit test) and per-technology Poisson capacity completions (Table 1, EIA data) yields Monte Carlo distributions: mean terminal 2031 price $33.43/MWh, standard deviation $5.87 over 1000 paths. The core contribution formulates generation expansion as a stochastic control problem: a single revenue-maximizing investor chooses the intensity of capacity completions, deriving an HJB equation verified in Appendix B and solved by a semi-implicit Euler scheme (Appendix C). Numerically, and in an explicit analytic example, optimal investment intensity falls to zero as capacity accumulates even while data-center demand keeps arriving, because each completed plant depresses the price earned on the whole portfolio (cannibalization). Under their calibration, prices still rise to roughly $46-49/MWh over six years. Main limitations: cost parameters are self-described as illustrative, the single-investor framing maximizes cannibalization, and no code is released.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
4 strengths · 4 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- With no supply response, ERCOT-scale data-center growth (6 GW/yr reference load) roughly doubles average wholesale prices from $30/MWh to $60/MWh within ten years.
- Modeling load and capacity arrivals as compound-Poisson processes turns point forecasts into wide terminal-price distributions (mean $33.43/MWh, sd $5.87, N=1000, 2025-31).
- A revenue-maximizing investor's optimal completion intensity falls to zero as capacity accumulates even while data-center demand keeps arriving, because each new plant cannibalizes the price earned on the whole portfolio.
- Under the paper's calibration, prices still rise to roughly $46-49/MWh over six years despite optimal investment.
- In an additive-demand analytic example, complete cannibalization makes never investing optimal, illustrating the mechanism in its extreme form.
Strengths
Rigorous mathematical treatment unusual for a policy-adjacent paper
Remark 2.1 proves existence/uniqueness of the clearing price and comparative statics via the implicit function theorem; Appendix B supplies a thinning construction, admissible-control class, DPP, supermartingale principle, and a full verification theorem (B.4) with multi-technology extension (B.6).
Calibration grounded in public data with explicit parameter tables
Tables 1-3 give per-technology Poisson intensities and increments calibrated to EIA 2015-2025 data; the data-center jump law is MLE-fitted to Texas campus capacities (mean 0.225 GW, sd 1.54 GW) with a Monte Carlo goodness-of-fit test, and the arrival intensity mu = 6/0.225 is tied to the 6 GW/yr ERCOT trend.
Elegant analytic extreme case that isolates the mechanism
Honest engagement with contradicting empirical literature
Weaknesses
Headline quantitative results rest on admittedly illustrative cost parameters
Single representative investor maximizes cannibalization by construction
Internal inconsistency between Sections 2.4 and 3
Thin statistical reporting and full-dispatch revenue assumption
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv 2609.08166, q-fin.GN) builds a three-layer framework connecting AI data-center load growth, generation capacity, and market-clearing electricity prices, calibrated to ERCOT/Texas: reference price P0 = $30/MWh, traditional reference demand I0 = 55 GW, data-center reference demand X0 = 8 GW, with demand-response functions featuring choke prices A1 = $70 and A2 = $150/MWh (Section 2.2). Layer 1 is deterministic growth (Section 2.3): with no supply response, 6 GW/year of data-center load doubles average wholesale prices from $30 to $60/MWh within ten years, and traditional consumers would be fully priced out only after about 22.3 years (dropout time tau in Appendix A). Layer 2 makes arrivals stochastic: compound-Poisson data-center demand (lognormal increments, mean 0.225 GW, sd 1.54 GW, MLE-fitted to Texas campus capacities with a goodness-of-fit test) and per-technology Poisson capacity completions calibrated to EIA 2015-2025 data (Table 1). Layer 3 (Section 3) endogenizes investment: a single revenue-maximizing investor controls the completion intensity of new capacity, yielding an HJB equation verified in Appendix B and solved by a semi-implicit Euler scheme (Appendix C). Headline claims: (i) prices double within a decade absent supply response; (ii) stochastic arrivals turn point forecasts into wide distributions (mean terminal 2031 price $33.43/MWh, sd $5.87, N = 1000); (iii) revenue cannibalization drives optimal investment intensity to zero even as demand grows; (iv) under their calibration prices still rise to roughly $46-49/MWh over 2025-31.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the paper's main finding about AI data centers and electricity prices?
Using a Texas/ERCOT-calibrated model, it finds that without new generation, projected data-center load growth (about 6 GW/year) roughly doubles average wholesale prices from $30 to $60/MWh within a decade; and when a profit-maximizing investor does build, each new plant lowers the price earned by its entire existing portfolio, so optimal investment slows or stops even as demand grows. Under the paper's calibration, prices still rise to about $46-49/MWh over six years.
What does 'revenue cannibalization' mean in this paper?
Electricity prices are set by supply-demand clearing, so every completed generator depresses the market price that all generators receive, including the investor's own existing plants. When that price decline outweighs the extra sales from the new unit, building more capacity destroys the investor's own revenue, and optimal investment ceases even while demand keeps growing.
What data does the paper use?
Public sources only: ERCOT load forecasts, EIA capacity-addition records (2015-2025), Texas Tribune data-center campus listings (lognormal fit: mean 0.225 GW, standard deviation 1.54 GW, excluding the Project Matador outlier), PJM market-monitor reports, and press reporting. No code or data archive is stated to be released.
Is the model calibrated or just illustrative?
Mixed. Demand levels, growth rates, arrival intensities, and capacity increments are calibrated to Texas/ERCOT public figures, but the investment-cost parameters (beta and per-technology rho in Tables 2-3) are explicitly described by the authors as reduced-form and illustrative, so the investment-path and price forecasts under optimal control should be read qualitatively.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure images are not extractable from the PDF text; all readings attributed to Figures 1-10 rely on their captions and in-text numerical descriptions only.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.