📈 economics

Certificates without Electrons? Theory and Evidence on Impacts from AI-Driven Power Demand

This paper combines game-theoretic modeling and empirical evidence to demonstrate that AI-driven power demand degrades grid reliability, raises prices, and increases emissions due to timing mismatches between consumption and renewable generation, revealing that current renewable certificate strategies fail to mitigate these impacts while solutions like colocation with storage and spatial reallocation offer effective alternatives.

Original authors: Dana Golden, Aruna Balasubramanian, Niranjan Balasubramanian

Published 2026-06-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dana Golden, Aruna Balasubramanian, Niranjan Balasubramanian

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the electrical grid as a massive, delicate highway system that delivers power to everyone's homes and businesses. For years, traffic on this highway has been relatively steady. But suddenly, a new type of "super-truck" has hit the road: Artificial Intelligence (AI) data centers. These trucks are huge, they eat a lot of fuel, and they move in unpredictable bursts.

This paper asks a simple but critical question: What happens to the highway when these super-trucks arrive, even if the truck drivers claim they are "carbon neutral"?

Here is the breakdown of what the researchers found, using everyday analogies.

1. The "Green Certificate" Illusion

Big tech companies often say, "Don't worry, our AI is green!" They do this by buying Renewable Energy Certificates (RECs). Think of these certificates like "green coupons." A company buys a coupon that says, "We paid for a wind farm to generate electricity somewhere else."

The paper argues that this is like buying a coupon for a clean meal at a restaurant across town, while you are currently eating a greasy burger at home. The coupon doesn't stop the grease from dripping on your floor right now.

The researchers found a "timing wedge." AI uses electricity in sudden, massive spikes (like a sprint), but the wind and solar farms they buy certificates for generate power slowly and steadily (like a marathon). Because the AI's hunger doesn't match the green energy's delivery schedule, the grid has to rely on dirty, fossil-fuel power plants to fill the gap in the moment.

2. The Impact on Your Lights and Outlets

When these AI trucks hit the grid, they don't just use more power; they make the power "messy."

The "Bumpy Ride" (Power Quality): Imagine your electricity as a smooth, steady stream of water. AI data centers act like a giant, erratic pump that creates surges and sags. The study found that in areas near these data centers, the power quality gets significantly worse.
- The Result: It's as if the average homeowner goes from having a power outage once a year to having 1.5 to 2 outages a year.
- The Danger: This "bumpy" electricity (called harmonic distortion) is like shaking a delicate vase. It can damage appliances like refrigerators and air conditioners over time and even increase the risk of electrical fires.

3. The Price Tag

When demand spikes suddenly, the price of electricity goes up.

The Analogy: Think of a concert where everyone rushes to buy tickets at the last minute. The price skyrockets.
The Finding: In certain areas (specifically the PJM zone, which covers parts of the East Coast), the arrival of major AI models caused wholesale electricity prices to jump by up to 25%. This is the cost the grid pays to keep the lights on when AI is "sprinting."

4. The "Fossil Fuel" Reality

Even though companies claim to be green, the study found that AI activity forces the grid to burn hundreds of gigawatt-hours more fossil fuel (like natural gas).

The Scale: This is enough extra energy to power about 45,000 American homes for a whole year.
The Cause: Because the AI spikes happen so fast, the grid can't wait for the sun to come up or the wind to blow. It has to fire up the backup gas generators immediately.

5. The Solution: "On-Site" Generators

The paper offers a surprising solution. Some data centers have their own power generators right on the property (like having a personal backup generator for your house).

The Magic: When a data center has its own power, the negative effects on the grid disappear and even flip. Instead of hurting the grid, these self-powered centers actually improve power quality for their neighbors.
Why? They absorb their own massive spikes of energy demand so the main highway doesn't feel the traffic jam.

6. What If We Change the Plan? (Counterfactuals)

The researchers ran simulations to see how different choices would change the outcome:

Edge Computing: Instead of sending all the AI work to one giant, crowded data center (a "mega-mall"), what if we spread it out to smaller centers near where people live? This would drastically reduce the strain on any single part of the grid.
Moving Locations: If we moved the heavy AI training work to areas with plenty of empty power capacity (like the middle of the country) instead of crowded cities, the grid impact would be much lower.
Bigger Models: As AI models get bigger (more "parameters"), the damage to the grid grows exponentially. A model twice as big doesn't just use twice the power; it creates disproportionately worse power quality issues.

The Bottom Line

The paper concludes that you cannot fix the grid's problems just by buying "green coupons" (certificates) if the actual electricity usage is messy and mismatched. The physical reality of how and when AI uses power is what matters.

To protect the grid, we need to either:

Stop the spikes: Use on-site power generators so the data center eats its own food.
Spread the load: Move AI work to less crowded areas or smaller local centers.
Be efficient: Make the AI models smarter so they don't need to sprint as hard.

Without these changes, the paper warns that the "green" AI revolution might actually be making our power grid less reliable, more expensive, and dirtier in the short term.

Technical Summary: Certificates without Electrons? Theory and Evidence on Impacts from AI-Driven Power Demand

1. Problem Statement

The rapid expansion of artificial intelligence (AI) has driven data center electricity consumption in the United States from 1.9% to 4.4% of total demand over six years. While hyperscalers utilize Renewable Energy Certificates (RECs) and Power Purchase Agreements (PPAs) to claim carbon neutrality, the grid-level effectiveness of these mechanisms remains unclear. Specifically, there is a lack of causal evidence regarding how AI model releases affect:

Local Power Quality: Including voltage stability, frequency regulation, and harmonic distortion.
Fossil Fuel Demand: The extent to which AI workloads displace renewable generation or necessitate fossil fuel dispatch.
Wholesale Electricity Prices: The magnitude of price increases in local markets due to AI-driven demand shocks.
Reliability: The frequency and severity of outages and equipment degradation.

Existing literature relies on techno-economic assessments or micro-level energy performance studies but fails to isolate the causal externalities AI imposes on other grid users. Furthermore, a "timing wedge" exists where consumption (AI workloads) and credited renewable generation (via RECs) are mismatched in time and space, potentially degrading reliability even when annual consumption is theoretically covered by renewables.

2. Methodology

The authors employ a rigorous econometric framework to address the scarcity of granular operational data (e.g., exact training start times) and the staggered nature of AI model releases.

2.1. Theoretical Framework

A game-theoretic model is developed where data center operators choose between RECs, PPAs, and behind-the-meter colocation, while generators make entry decisions under endogenous financing costs. The model identifies a timing wedge—the mismatch between consumption and credited generation—as the central mechanism through which AI demand degrades reliability and increases emissions, even with 100% REC coverage.

2.2. Econometric Strategy: Staggered Difference-in-Differences (DiD)

To overcome the lack of specific training windows and the staggered release of models, the authors utilize a Stacked DiD with Multiple Horizons approach.

Treatment Definition: Treatment is defined relative to the release date ( $r_m$ $r_{m}$ ) of major AI models.
- Training Window: $[r_m - \tau_{train}, r_m)$
- Inference Window: $[r_m, r_m + \tau_{infer}]$
Stacked Design: To avoid bias from treatment effect heterogeneity (where already-treated units contaminate control groups in standard Two-Way Fixed Effects models), the authors construct separate "sub-experiments" for each model release cohort. Each sub-experiment uses only units that are genuinely untreated at the time of that specific cohort's adoption.
Equation: The dynamic treatment effect is estimated as:
$Y_{i,t,m} = \sum_{k=-K}^{K} \beta_k \cdot D^k_{ut} + \delta_t + \gamma_u + X'_{ut}\theta + \varepsilon_{ut}$
Where $D^k_{ut}$ indicates the event-time horizon $k$ relative to the release.

2.3. Addressing Endogeneity and Confounds

Fossil Demand: A Two-Stage Least Squares (2SLS) approach is used to address price endogeneity.
- First Stage: Electricity prices are instrumented using supply-side cost shifters (generator heat rates and natural gas spot prices).
- Second Stage: Fossil fuel demand is regressed on predicted prices and treatment indicators.
Price Effects: Price impacts are derived by combining estimated demand shifts with supply curve elasticity ( $\varepsilon_s$ ) in the PJM market.
Robustness Checks: The study employs parallel trends tests (Wald tests on pre-treatment coefficients), placebo tests (randomized treatment dates and locations), and sensitivity analyses regarding treatment definitions (e.g., varying radius, dropping ambiguous cases).

2.4. Data

The analysis integrates a novel dataset linking AI activity to grid outcomes:

Treatment Data: Aterio database for data center locations, ownership, and capacity; linked to specific model releases (e.g., GPT-3, Llama 2, Claude 2.1) via press releases and API dates.
Outcome Data:
- Power Quality: Whisker Labs Consumer Power Quality Index (CPQI) and Total Harmonic Distortion (THD) from 72 utilities.
- Demand: EPA Clean Air Markets Program (CAMPD) hourly generator data (2021–2023).
- Prices: ISO zonal wholesale prices (focusing on PJM).
Controls: Weather data (Meteostat), natural gas prices (S&P Capital IQ), and market indices.

3. Key Results

3.1. Power Quality Degradation

AI model releases cause significant deterioration in local power quality:

Consumer Power Quality Index (CPQI): Treated utility areas experience a coefficient increase of 0.3795, corresponding to a shift from typical U.S. power quality to the bottom quartile. This equates to an increase in annual outages for the average customer from ~1 to 1.5–2 outages per year.
Harmonic Distortion (THD): There is a 0.0625 percentage point increase in readings above the 0.8 distortion threshold. While seemingly small, sustained exposure accelerates equipment degradation (e.g., motors, refrigerators).

3.2. Fossil Fuel Demand

AI activity significantly increases fossil fuel generation:

Magnitude: The training phase adds approximately 6.33 MWh per generator-hour, and the inference phase adds 5.20 MWh.
Aggregate Impact: This translates to over 500 additional GWh of fossil generation near data centers, equivalent to the annual consumption of roughly 47,000 U.S. households.
Mechanism: The effect is driven by the immediate need for baseload and peak capacity that RECs do not temporally cover.

3.3. Wholesale Price Impacts

In the PJM market, AI-driven demand spikes lead to substantial price increases:

Zone Variation: Price impacts vary by supply elasticity. In the American Electric Power (AEP) zone (lower elasticity), wholesale prices increased by up to 25% during AI activity periods. In the ComEd zone (higher elasticity), the impact was smaller but still significant.
Retail Implications: While retail pass-through is lagged, the authors estimate a potential long-run consumer price rise of up to 25% in high-impact zones, as higher-cost marginal generators are dispatched to meet AI demand.

3.4. Mitigation via On-Site Generation

A critical finding is the sign reversal in power quality effects based on infrastructure:

Without On-Site Generation: Power quality deteriorates (positive CPQI coefficient).
With On-Site Generation: Power quality improves (negative CPQI coefficient).
Interpretation: Colocated generation (often with storage) absorbs demand spikes directly, preventing them from propagating to the grid and reducing reliance on marginal fossil generators.

4. Counterfactual Analyses

The paper evaluates how different technical evolutions of AI could alter grid impacts:

Model Scaling: Scaling from 2 trillion to 4 trillion parameters would increase power quality impact by ~35%. A 1% increase in model parameters correlates with a 0.15% increase in power demand.
Geographic Redistribution:
- Edge Inference: Shifting inference from concentrated hyperscaler facilities to distributed edge nodes reduces zonal demand spikes significantly (from 162 MWh/hour to 4 MWh/hour in simulations).
- Centralization: Concentrating all training in a single large facility creates extreme local demand spikes (~100 GWh), highlighting the risks of over-concentration.
On-Site Generation: Universal adoption of on-site generation would shift the net impact of data centers from a power quality detriment to a net improvement.

5. Significance and Contributions

The paper claims three primary contributions to the literature:

Methodological Innovation: It establishes a robust methodology for assessing macro-level grid impacts of AI using publicly available data, overcoming the lack of proprietary operational details through staggered DiD and instrumental variables.
Causal Evidence: It provides the first causal estimates of AI model releases on power quality, fossil fuel demand, and wholesale prices, moving beyond correlational studies or techno-economic projections.
Policy Relevance: It demonstrates that the externalities of AI are tightly coupled to procurement design and spatial organization. The results suggest that REC-only strategies are insufficient for grid stability. Instead, on-site colocation with storage and spatial reallocation (e.g., edge computing) are identified as effective levers for mitigating grid impacts.

The authors conclude that without structural changes in how AI infrastructure is powered and located, the continued scaling of AI models will continue to degrade grid reliability and increase costs for all consumers, regardless of carbon neutrality claims.