Grid Lock-in and the Geography of AI

Key Takeaways

• Research Question: When grid capacity is slow to expand, can an early electricity-intensive entrant lock up scarce grid capacity and displace later, more mobile investment?
• Data: We link 5,140 U.S. data center facilities with electricity-market data, recreational cannabis ballots, labor-market records, and a new satellite measure of indoor cannabis cultivation over 2000–2024.
• Method: We exploit close recreational cannabis ballots in a panel difference-in-discontinuities design, comparing changes in counties of narrowly passing and narrowly failing states.
• Findings: Legalization increases detected indoor cultivation by 3.7 times the sample mean within five years. Active data center stock falls by 0.70 facilities per county and operating capacity by 16.2 MW. The decline begins before cultivation fully scales and reaches counties across state lines that share the same utility grid.
• Implications: Slow-adjusting grids create an opportunity cost across industries: capacity committed to one large electricity user can redirect later, more mobile investment elsewhere.

Background

Electricity-intensive industries can compete for more than power prices. They also compete for access to a grid that takes years to expand. Once local generation, transmission, or substation capacity is committed, later users may face higher costs or longer waits for connection.

We study this problem through recreational cannabis legalization. Indoor cultivation uses large amounts of electricity and expands quickly after legalization, creating a concentrated new load. Data centers also require substantial power, but their location is relatively flexible. At the same time, the two industries share little labor, land, or technology, which helps isolate competition through the grid.

We ask whether this early cannabis load changes where data centers are built. If developers expect higher future electricity costs or tighter interconnection capacity, they may avoid the affected grid before those constraints are fully reflected in current retail prices. We call this mechanism grid lock-in.

Data & Methodology

We build a U.S. county-year panel from 2000 to 2024. Data center activity comes from Aterio, which tracks 5,140 facilities across 461 counties, including their operating status and power capacity. Electricity prices and consumption come from the U.S. Energy Information Administration.

Cannabis cultivation is harder to measure consistently across states, especially where production is unlicensed or illegal. We therefore construct a satellite-based measure that combines nightlights, thermal anomalies, and rooftop characteristics, and validate it against licensed cultivation sites.

For identification, we exploit recreational cannabis ballots decided close to the 50% passage threshold. We use a panel difference-in-discontinuities design, which compares how data center activity changes after the vote in counties of narrowly passing and narrowly failing states. The main sample includes 26 ballots within ten percentage points of the cutoff.

Findings

Legalization expands cultivation and reduces data center capacity

Recreational cannabis legalization produces a significant increase in indoor cultivation. By the fifth post-ballot year, detected cultivation is 0.176 square kilometers higher per county, equivalent to 3.7 times the sample mean.

Electricity prices rise where cultivation expands. An additional square kilometer of detected cultivation is associated with a 2.2% increase in average retail electricity prices, with similar responses for commercial and industrial users.

Data center investment moves in the opposite direction. In counties of narrowly passing states, the active stock falls by 0.70 facilities per county, operating capacity by 16.2 MW, and entrant capacity by 4.4 MW.

Data centers respond before cultivation fully scales

The data center gap opens in the ballot year, while cannabis cultivation continues to expand over the following years.

The same pattern appears across counties. Electricity prices rise mainly where cannabis supply was already established before legalization. Data center activity also falls in counties with little prior cannabis activity, where prices show little immediate response.

For data center developers, future grid conditions matter at the siting stage. Expected power costs, available substation capacity, and interconnection delays can affect location decisions well before those constraints appear fully in retail electricity prices.

The effect extends across state borders through the shared grid

The decline does not stop at the state line. Counties in neighboring states served by the same utility grid lose data centers after a ballot passes next door, even though cultivation never arrives there and their own retail prices do not move.

These counties speak to the mechanism. With no local price change to respond to, the pattern points to constraints that work through the shared grid and not through retail rates. Capacity committed on one side of the line can tighten the headroom available to users that connect later on the same network, while each state still sets its own retail rates.

Higher electricity costs reshape data center siting

We estimate a discrete-choice model using 2,121 data center siting decisions. The model instruments a county’s electricity price with legalization, and that first stage puts the legalization-induced price increase at 12.1%. This captures the overall price change associated with legalization, rather than the 2.2% increase per square kilometer of cultivation reported above. A price increase of that size lowers a county’s relative probability of attracting a data center by 44%.

Without the legalization-induced price increase, more data center capacity would be located in the narrowly passing states and less in states such as Texas, Pennsylvania, and Illinois.

The model also implies an annual $15.6 billion consumer-surplus loss across 648 counties in near-threshold ballot-passing states, reflecting higher electricity costs borne by consumers that cannot relocate.

Implications

Policies that attract electricity-intensive industries can have consequences beyond the sector they target. Where grid capacity is slow to expand, the gains from one new entrant may come with two costs: later mobile investment can be pushed elsewhere, and households and other captive users can face higher electricity prices.

This makes grid capacity part of the policy trade-off. Expanding transmission can reduce the risk that one large load crowds out the next, while retail-rate protection can limit the burden on consumers who cannot relocate when the grid tightens.