Which U.S. state has the cheapest electricity for data centers?

Among the markets in this index, Texas has the lowest industrial electricity price at 6.33 cents per kWh as of May 2026, followed by Tennessee at 6.45 and Iowa at 6.62. California is the most expensive at 20.20 cents per kWh. These are state average industrial retail rates from EIA; no large data center pays the state average, and real contracts are negotiated.

Are data centers raising electricity prices?

The evidence is mixed and depends on market design. Industrial electricity prices rose 29.7 percent in Ohio, 16.0 percent in New Jersey and 14.0 percent in Virginia between May 2025 and May 2026, all PJM states, against a U.S. industrial average increase of 5.1 percent. Over the same period Texas industrial prices fell 3.9 percent while absorbing the fastest load growth of any major U.S. grid. PJM's independent market monitor attributes 38 percent of the latest capacity auction cost to data centers, which also means 62 percent is not attributed to them. A separate causal study by EPRI-affiliated researchers found data center growth associated with slightly lower residential prices over 2015 to 2024.

How is grid stress measured in this index?

Grid stress is a percentile rank combining seasonal peak headroom at 50 percent weight, inverse hour-to-hour ramp volatility at 30 percent, and inverse day-ahead demand forecast error at 20 percent, all computed from EIA-930 hourly balancing-authority data. Headroom compares the trailing 30-day mean of daily peaks against the 99.9th percentile hour in the same calendar window in prior years, so a winter-peaking grid is not credited with headroom for being measured in August.

Standing index of U.S. data center power economics

AI Data Center Power Cost & Grid Stress Index

Q: What does 1 MW of AI compute cost to power?

At PUE 1.3 and full utilisation, 1 MW of IT load draws about 949,000 kWh a month. At the U.S. average industrial rate of 8.71 cents per kWh that is roughly $82,657 a month, or about $991,888 a year, energy only. The same megawatt costs about $720,860 a year in Texas and about $2,300,376 in Santa Clara, California.

Where in the United States is it cheapest, cleanest and least risky to add the next megawatt of AI load, right now? Thirty-one metros across twenty-four distinct state-and-grid units, scored on federal hourly grid data, published industrial power prices and EPA emissions rates. Every input is public. Every limitation is listed.

PCI-1 · v1.0revision —grid observed through —

Jump to the index Claim ledger Inference ledger

Finding · The divergence

PJM states got double-digit price increases. Texas got a price cut.

Change in the average retail price of electricity to industrial customers, May 2025 to May 2026. The four steepest increases are all PJM states. The two steepest decreases, Texas and Arizona, are the grids that added new generation fastest. We have not seen this number reported anywhere.

Industrial electricity price change, May 2025 → May 2026

Percent change. Amber is an increase, green a decrease. The U.S. industrial average rose 5.1%.

Source: EIA Electric Power Monthly Table 5.6.A, May 2026 and May 2025 columns, industrial sector. Preliminary, Form EIA-861M. One state per row; markets sharing a state share a value.

The calculator

What does 1 MW of AI compute cost to power in your market?

Energy only. Pick a market, set your IT load, utilisation and PUE, and the model runs the state industrial retail rate against your facility power. These are scenario estimates, not quoted prices. No hyperscaler pays the state average. The gap between this number and a negotiated contract is the thing worth arguing about.

Market

IT load MW

Utilisation %

PUE

Electricity price ¢/kWh

Grid carbon intensity lb CO2e/MWh

Modelled annual electricity cost

—

Facility power

—

Energy / month

—

Cost / month

—

Cost / day

—

CO2e / year

—

Cost per MW-year

—

Finding · The spread

The same 10 MW costs three times as much to power in Santa Clara as in Dallas

Modelled annual energy-only electricity cost for a 10 MW IT load at PUE 1.3 and 100% utilisation, using each state's industrial retail rate. Darker bars cost more.

Annual electricity cost, 10 MW IT load

U.S. dollars per year, energy only. Sorted cheapest first.

Source: EIA Electric Power Monthly Table 5.6.A, industrial sector, May 2026 (released 2026-07-23). Model: MW x 1,000 x PUE 1.3 x 730 h x 12 months x state rate.

Finding · The trade-off

Cheap power and a relaxed grid rarely come together

Every metro plotted by cost score against grid stress score. The top-right corner is where you would want to be. Almost nobody is there.

Cost score vs. grid stress score

Percentile rank across the 24 distinct state-and-grid units, 0-100. Higher is more favourable. Coloured by grid family.

Grid stress is computed at the balancing-authority level from EIA-930 hourly data: seasonal peak headroom, ramp volatility and day-ahead forecast error. It cannot see a local substation constraint. The Local Interconnection Reality column is for that. Overlapping points are nudged 9px right (about 1.1 score points) so stacked metros stay visible; hover shows how many metros share a position. Metros sharing a state and a grid share a score by construction.

The index

All 31 metros, 24 distinct grid-and-price units

Click any column to sort. Scores are computed over unique state-and-grid units, then joined onto every metro sharing that unit, so the ranking cannot be shifted by listing more metros on one grid. The Metros / unit column shows where that is happening. The AI Power Readiness score is 0.45 × cost + 0.35 × grid stress + 0.20 × carbon. Those weights are our judgement, and the tool lets you change them.

Filter

#	Market	State	Grid (BA)	¢/kWh	Price YoY	10 MW / yr	eGRID lb/MWh	Observed lb/MWh	Peak headroom	Grid load 2yr	Cost	Stress	Carbon	AI Power Readiness	Metros / unit	Local interconnection reality

Cite & press

Quote it, check it, call us.

Cite the story

Cite the dataset

Data CC BY 4.0. Charts may be reproduced with the source line. Full guidance on the cite & press page.

Check it. Every number here is either computed by our published code from the federal data named above, or traced to the exact sentence of its source in the claim ledger. Every opinion is in the inference ledger with what would prove it wrong.

Corrections. Every change to this story is logged on the corrections page; nothing is edited silently.

On deadline. Joe Robison, publisher: joe@greenflagdigital.com. Same-day answer on any figure. Embargoed early access to major releases for beat reporters.

Disclosures. Run by AI desks under a human publisher; funded by an affiliate hosting site that does not advertise here; the first public deployment of a research harness its operator runs for organizations. All of it on the about page.