Not Yet, Broadly — But the Mechanism Is Now Live: AI Data Centers and Household Electric Bills

Summary

Mostly not yet, for most households through 2024 — but the causal mechanism is now demonstrably operating in at least one major grid region, and the dollar effects are starting to show up in bills. The most rigorous study to date, published by Rutgers researcher Kiran Garimella in 2026 and covering 22,834 ZIP codes across 24 states from 2014-2024, reports no statistically significant, broad-based increase in nearby households' electric bills tied to data center openings during that period — the paper's own headline numbers range from a borderline +$2.09/month down to a non-significant +$0.90-1.84/month, and its utility-territory-level figure (the level at which rates are actually set) was indistinguishable from zero. Separately, we independently pulled EIA retail price data through 2024 and found PJM-region states (the 13-state grid hosting the largest AI data center concentration in the world) did NOT rise faster than the rest of the country over 2019-2024. But that picture is now changing: PJM's own Independent Market Monitor attributes a majority of a 9.3x spike in its capacity-market auction price (from $28.92 to $269.92, later capped at $329/MW-day) to AI data center load growth, and SemiAnalysis calculates this alone will add roughly $25-30/month — about 15% — to a typical PJM household's bill by 2026/27. A concrete case already visible: New Jersey households saw an overnight ~20% residential rate jump in June 2025 tied to this auction outcome. By contrast, Texas (ERCOT), which has comparable AI data center growth but no capacity-auction market design, has seen wholesale prices move only a few percent — evidence the effect is as much about market design as about data centers' raw power draw.

What the best available research found (Rutgers/NJ State Policy Lab, 2026)

Kiran Garimella (Rutgers) and collaborators built a panel of 22,834 ZIP-level units across 24 states, 2014-2024, linked to 2,277 operational data centers with known ZIP codes and opening dates, using ACS self-reported household electricity bill data as the outcome. These are the published paper's own numbers — we did not re-run their underlying model in this session, only checked their headline conclusion against an independent government price series (see below). As reported by the authors:

• Households in ZIP codes with a data center reported bills about $2.09/month higher than households in ZIP codes without one — small next to the ~$164 average monthly bill in the sample, and only borderline significant by the paper's own reporting.
• Restricting the comparison to ZIP codes that had similar bills before any data center opened brought that figure down to +$0.90/month, described by the authors as not statistically significant.
• After accounting for local income changes, the paper reports +$1.84/month, again not statistically significant.
• At the utility-service-territory level (the level at which residential rates are actually set by state regulators) — the cut the authors call most decision-relevant — the paper reports an unweighted figure of +$0.35/month and a household-weighted figure of -$2.00/month. The authors describe neither as distinguishable from zero.
• A check for whether OTHER ZIP codes served by the same utility as a data center also see higher bills (which would be consistent with costs being spread across the whole customer base) came back negative in the paper, offering no support for a simple cost-spreading story in this period.
• Some higher figures appeared 4-5 years after a data center opened, which the authors say is consistent with the multi-year lag of utility rate cases, but this pattern did not hold at the utility-territory level and rests on fewer long-follow-up observations.

The authors' own caveat is important: their sample is dominated by data centers built before the current AI wave, which are far smaller than the hundred-megawatt-to-gigawatt hyperscale AI campuses now being built. Their conclusion is explicitly "not yet," not "never."

The mechanism that IS now measurably operating: PJM's capacity market

PJM Interconnection — the wholesale grid operator covering 13 eastern states plus DC, including the world's largest data center cluster in Northern Virginia and major AI buildouts in Ohio and Indiana (Google, Amazon/Anthropic, Meta) — runs an annual forward capacity auction (the Base Residual Auction) that pays power plant owners to stay on standby for peak demand. This price is spread across all PJM customers' bills as a distinct "capacity charge."

That auction price rose from $28.92/MW-day (2024/25 delivery) to $269.92/MW-day (2025/26) — a 9.3x jump — and the subsequent two auctions (2026/27, 2027/28) both cleared at a federally imposed price cap of $329/MW-day, the maximum allowed. PJM's own Independent Market Monitor (IMM), an entity required by federal regulators to audit the auction, ran counterfactual simulations and found:

• Removing ALL data center load from the demand forecast would have reduced total capacity payments by 64% (a $9.33 billion reduction).
• Removing only data centers not yet energized (keeping already-operating ones) still would have cut costs by 53% ($7.74 billion).
• The IMM attributes roughly 7.9 GW of incremental 2025/26 datacenter demand and ~12 GW for 2026/27 — no other factor came close in size.

SemiAnalysis calculates that the capacity-price increase alone will add roughly $25-30 per month to an average PJM household's bill by 2026/27 (using ~880 kWh/month average usage and a 0.4 load factor) — on the order of a ~15% increase relative to the pre-AI-datacenter baseline for PJM's 67 million residents. A concrete, already-realized instance: New Jersey households saw an overnight roughly 20% jump in residential electricity rates in June 2025, which triggered state legal action and was a major issue in that year's state elections.

Important nuance: PJM's forecasting methodology (the Variable Resource Requirement curve) is itself a simulation built on PJM's own demand projections, and SemiAnalysis's independent tracking of actual data center construction timelines suggests PJM's forecasts have been running too high — the datacenter load forecast was cut by 800 MW in 2024 and another 1.1 GW in 2025 relative to the prior year's projection. So while data centers are the dominant driver PJM's own monitor identifies, the SIZE of the resulting price spike is also inflated by forecast uncertainty and PJM's specific market design, not physics alone.

The counterfactual that isolates market design from data center growth: ERCOT/Texas

Texas (ERCOT) is undergoing an AI data center buildout of comparable scale — OpenAI, Google DeepMind, and Anthropic are all building large Texas facilities, and ERCOT's own 2025 long-term forecast identified data centers as the single largest driver of incremental peak demand growth (projecting up to 77.9 GW of potential data center load by 2030). Yet ERCOT has no capacity auction — it is an "energy-only" market using real-time scarcity pricing — and Texas wholesale/futures electricity prices have moved only a few percent over the same period, nothing resembling PJM's 9.3x capacity-price spike. This is strong evidence that the PJM effect on bills is substantially a market-design story (how a specific state-supervised auction mechanism translates forecasted load growth into a cleared price) layered on top of the underlying real demand growth, not a direct, physics-driven pass-through of data center electricity consumption into rates everywhere data centers exist.

Our own check against EIA retail price data

To sanity-check the Rutgers finding independently, we pulled EIA residential retail electricity prices by state (askamerica energy.eia_electricity_prices, sourced from EIA's Electric Power Monthly, observed coverage 2006-2024) and compared the 2019-2024 percentage change in residential rates for the 13 PJM states plus DC against the rest of the country. This was a simple descriptive comparison of published average prices, not a regression.

Result: PJM states rose 25.4% on average (unweighted, 2019 to 2024), essentially in line with the 26.7% national average and even slightly below Texas's own 27.0% — and only modestly above the 21.7% average for non-PJM states. This lines up with the Rutgers finding: through 2024, there is no clear PJM-specific price gap visible in the retail data. The capacity-auction effect described above only began clearing in mid-2024 for the 2025/26 delivery year and is still working its way into rate cases and bills — EIA's most recent loaded year (2024) sits right at the front edge of that transition, which is why the gap is not yet visible in this series but is squarely implied by the auction data and the June 2025 New Jersey rate event that followed it.

Bottom line

For the average American household as of the most recent data with broad coverage (through 2024), there is no robust evidence that data centers have already driven a detectable, widespread increase in electric bills — several ways of slicing the numbers in the best available academic study all come back statistically insignificant, and our own independent state-level price check agrees. But "not yet" is doing real work in that sentence: in PJM specifically — the region hosting the largest concentration of AI data centers in the country — a well-documented, regulator-monitored mechanism (the capacity auction) has already cleared prices that are expected to add on the order of $25-30/month (~15%) to typical bills by 2026/27, with PJM's own market monitor attributing roughly half to two-thirds of that increase to data center load growth specifically, and New Jersey has already experienced a real ~20% rate jump tied to this mechanism. The comparison with Texas — comparable data center growth, minimal price impact — shows that market design, not data center electricity use per se, determines how much (if any) of this cost reaches ordinary households, and where capacity markets and slow-moving rate cases exist, the bill impact is now becoming real rather than hypothetical.

Every query behind this report

1 warehouse call ran in this session, in order. Each is reproducible against the same snapshot.

SELECT state_abbr,
  MAX(CASE WHEN price_year=2019 THEN avg_price_cents_kwh END) AS p2019,
  MAX(CASE WHEN price_year=2024 THEN avg_price_cents_kwh END) AS p2024
FROM energy.eia_electricity_prices
WHERE sector_code='RES' AND price_year IN (2019,2024)
GROUP BY state_abbr
ORDER BY state_abbr