AI data-center demand could outpace power-plant investment: uncertainty and 'cannibalization' of revenues mean private developers may underbuild generation, leaving wholesale electricity prices higher and consumers bearing the cost.

Main Finding

Rapid, lumpy growth in AI data-center electricity demand can raise wholesale electricity prices substantially, but the same price signals that justify new generation also reduce the future revenues of entrants (a “revenue cannibalization” effect). Under realistic ERCOT/Texas-calibrated parameters, stochastic uncertainty in data-center arrivals and long, lumpy build times for generation weaken private incentives to build capacity at the pace needed to stabilize prices — so prices may rise and remain elevated unless policy or contracting corrects coordination and timing failures.

Key Points

• Conceptual framing: the paper links reference demand from two groups (traditional users and hyperscaler data centers), available accredited generation capacity, and a market-clearing price determined by D(I,X,P)=S.
• Demand heterogeneity: hyperscalers are modeled as much less price-sensitive than traditional users (choke prices A1=$70/MWh, A2=$150/MWh; normalized at reference price P0=$30/MWh).
• Deterministic benchmark:
  • If supply is fixed, data-center growth can substantially raise prices (example: price could double from $30 to ~$60/MWh over 10 years under calibrated assumptions).
  • Linear supply growth at different rates can either offset price pressure or, if fast, cause prices to fall (illustrating cannibalization).
• Stochastic model:
  • Data-center build-outs modeled as a compound Poisson process: arrivals N^µ_t with lognormal jump sizes Yk (calibrated to Texas campus capacities, mean 0.225 GW, sd 1.54 GW) to capture rare very large facilities.
  • Supply additions modeled by independent Poisson processes by technology j (natural gas, coal, solar, wind, large-scale nuclear, SMRs) with technology-specific intensities λ_j and per-arrival capacity δ_j (Table 1).
  • Calibration: reference parameters chosen to match ERCOT/Texas (P0=$30/MWh, I0=55 GW, X0=8 GW; targeted data-center growth c_X=6 GW/yr → µ≈26/yr).
• Monte Carlo results (N=1000, horizon 2025–2031):
  • Terminal price mean ≈ $33.43/MWh, SD ≈ $5.87/MWh; distributions show non-negligible probability of higher prices (and also of prices falling below baseline if supply additions cluster).
  • Representative sample paths illustrate wide dispersion in supply, demand, and prices due to lumpy arrivals.
• Investment incentives and cannibalization:
  • The authors formulate generation expansion as a stochastic control problem: a revenue-maximizing investor chooses investment intensity under uncertainty in demand growth and build completions.
  • Successful investment reduces future market prices and therefore erodes its own (and incumbents’) expected revenues. When this cannibalization and uncertainty dominate, optimal private investment can be insufficient or delayed even when aggregate demand growth argues for faster build-out.
• Policy-relevant observation: even large price spikes do not guarantee timely, sufficient entry because of long lead times, permitting and interconnection frictions, skewed project sizes, and value cannibalization.

Data & Methods

• Analytical setup:
  • Market clearing: total accredited supply S_t equals price-responsive demand D(I_t,X_t,P_t), with D split into D1=I_t F1(P) and D2=X_t F2(P). F1, F2 chosen decreasing, normalized to 1 at P0.
  • Deterministic scenarios: closed-form/ numerical solutions for price paths under fixed or linearly growing supply and demand.
• Stochastic modeling:
  • Data centers: compound Poisson process X_t = X0 + Σ_{k=1}^{N^µ_t} Y_k; Y_k ~ lognormal (MLE fit to Texas capacities excluding outlier Project Matador), µ calibrated so E[annual growth]=6 GW/yr.
  • Supply: S_t = S0 + Σ_j δ_j N^{λ_j}_t with independent Poisson processes per technology j; technology parameters λ_j, δ_j calibrated from 2015–2025 capacity-addition data (SMR parameters assumed due to lack of history). Example baseline table: natural gas (λ=10/yr, δ=250 MW), solar (λ=40/yr, δ=50 MW), etc.
  • Market-clearing price solved implicitly from D(I_t,X_t,P_t)=S_t at each simulated date (long-run monthly average price interpretation).
• Simulations:
  • Monte Carlo (1000 runs) over T=6 years (2025–31), producing distributions of S_t, X_t, and P_t. Summary statistics reported (mean, SD, distribution plots).
• Stochastic control:
  • Formulate generation expansion as a controlled-intensity jump process where an investor selects investment intensity to maximize discounted expected revenue accounting for stochastic demand arrivals and price impact from capacity additions (i.e., cannibalization). The paper analyzes qualitative trade-offs and comparative statics across technologies differing in cost, size, and build time.

Implications for AI Economics

• Market signals alone may not deliver socially optimal capacity expansion:
  • Price spikes can be necessary but not sufficient to spur timely entry because project lumpy sizes, long lead times, and cannibalization reduce expected returns to investors.
• Need for coordination mechanisms:
  • Policy interventions (streamlined permitting/interconnection, transmission build-out, targeted subsidies, capacity markets, long-term contracts, offtake agreements) can reduce uncertainty, internalize externalities, and align private incentives with social needs.
• Technology-choice and timing matter:
  • Faster-to-deploy resources (solar, gas) can relieve near-term scarcity but may be politically constrained and intermittent, whereas firm technologies (gas, nuclear/SMR) have long lead times and higher up-front costs; optimal mixes depend on expected demand growth and risk tolerance.
• Data-center siting and contract design:
  • Hyperscalers’ lower price responsiveness implies they impose high reliability/value requirements; regulators and grid planners may want pricing or locational charges that reflect true network costs, or encourage data centers to sign firm long-term capacity contracts to signal demand and finance supply.
• Implications for AI deployment costs:
  • If electricity prices rise materially and persist, the marginal cost of operating large AI models increases, potentially affecting deployment decisions, geographic location of compute, and the social cost of AI scaling.
• Research and policy gaps:
  • The model abstracts from transmission constraints, retirements, short-run intermittency, distributional effects, and interactions between real-time and long-run prices — all important for designing practical interventions.
• Practical takeaway: managing AI-driven electricity demand growth requires active coordination (planning, contracting, regulatory reform) rather than reliance on decentralised price signals alone; otherwise private underinvestment and lumpy supply arrivals can leave grids strained and consumers exposed to higher bills.