AI is not hiking your electricity bill – yet

Consumers worry about AI is driving up their bills. But inflation and the climate transition are to blame

Recently, politicians have taken to blaming the massive AI data center buildout for the rising energy costs borne by Americans. Sens Warren, Van Hollen, and Blumenthal opened an investigation into big tech, saying:

We write in light of alarming reports that tech companies are passing on the costs of building and operating their data centers to ordinary Americans as AI data centers’ energy usage has caused residential electricity bills to skyrocket in nearby communities. Through these utility price increases, American families bankroll the electricity costs of trillion-dollar tech companies.

Bernie Sanders has called for a full nationwide moratorium on data centers. Even more moderate Dems like JB Pritzker and Josh Shapiro have begun to retreat from a previously pro-data center attitude. Maryland’s Chris Van Hollen proposed a bill to “hold data center operators accountable for their role in driving up energy prices.”

And the general public is fired up too. Seventy percent of Americans oppose data center construction in their neighborhood. Forty-three percent of Americans think data centers are a major reason for their rising energy bills.

The resource mobilization to support the AI buildout is the defining feature of our economy today. The debate over the resource costs of AI and its societal (dis)utility will be the most important political issue throughout the next decade. It is therefore of paramount importance that we level set and consider the actual facts regarding the emergence of these large loads on the grid. For instance, this long hitpiece explicitly premised on the idea that data centers drive up electricity prices provides no actual evidence that this is actually happening at scale.

In this chartbook, I’ll walk you through a few different visualizations to illuminate the issue. And you don’t have to trust me. Every single one of these charts is reproducible in Claude. All sources and assumptions are included in the Appendix and that should be sufficient for you to check my work if you like.

Sources and assumptions for all charts in Appendix

Because people forget, it’s important to remember that the US generation capacity is massive and data centers are still a relatively small part of the grid. Total generation is around 4,400 TWh/year, data centers account for four percent of that at 184 TWh/year, and AI-focused data centers at around half of that. We are talking about two percent of grid capacity. Of course, building out energy infrastructure is slow and ridden with bottlenecks, so that two percent – up from virtually nothing five years ago – can cause problems. But has it?

Fortunately, America is divided into fifty states each with slightly different approaches to energy policy, which gives us fifty natural experiments.

For this chart I divided up the fifty states into quintiles, ranked by the percentage of state electricity consumed by data centers in 2023, and took the mean residential electricity rates in those quintiles. (Note that these are real (inflation-adjusted) prices. In nominal terms, rates went up everywhere. We will cover this later). The first quartile Q1 consists of states with virtually no data centers, and Q5 includes states like Virginia and Texas that are data center powerhouses (data centers account for a full quarter of Virginia’s electricity consumption).

Sources and assumptions for all charts in Appendix

A few things are immediately obvious. The quintile featuring states with the highest data center share of grid electricity – VA, TX, NV, OR and so on – have the cheapest residential power rates, whereas states with less data center intensity are more expensive. While counterintuitive, this has an easy explanation: developers flock to states with cheap power, a culture of building infrastructure, less regulatory red tape, and more available industrial capacity.

So, if you live in a state with lots of data centers, your experience over the last decade is most likely pretty good! The data centers are there because prices are low.

But has the new buildout of AI data centers increased prices since ChatGPT was released in 2022? According to the chart, not measurably. The most data center-intense states have dealt consumers the smallest rate hikes since 2022. On average, electricity prices went up nationwide since 2022, and this isn’t unique to the states with the most data centers.

“But hang on,” you might say. “I live in one of the highest-data center-intensity quintile states, and my power bills did go up in the last decade.”

That’s true, but that’s overwhelmingly just inflation. Taking a look specifically at the fifth quintile (the 10 states with highest data center penetration):

In inflation-adjusted terms, electricity prices actually fell over the decade in the most data center dense states. So you aren’t upset about data centers. You are upset about the massive post-covid inflation. This is the sleight of hand powering policymaker attacks on data centers.

Because inflation is the cofounding factor, almost all of my charts in this piece are real, as in inflation-adjusted. This is because I am interested in measuring trends other than inflation, which was the major driver of prices over the last five years.

If you think the quintile view is too lossy, here is a per-state chart of residential power prices today versus data center share of total electricity consumption.

Sources and assumptions for all charts in Appendix

Visually, its easy to see that the big data center states – Virginia, Oregon, Nebraska, Texas, Iowa – are not among the most expensive. The most expensive states are in New England, the rust belt, and California. They have relatively limited data center activity.

But of course, this doesn’t show causality. Data centers are disproportionately placed in states with cheap power. It’s not that the presence of lots of data centers drives down power prices (more on this later). This is just a snapshot view.

But it is certainly possible to add large loads without resultant residential energy tariffs. Let’s look at the two of the most critical states for the AI/data center buildout: Texas and Virginia.

Texas is a great case study because it has undergone an absolutely gargantuan data center buildout, from traditional DCs to Bitcoin mining and now to AI. ERCOT’s large load task force which tracks every data center over 75 MW gives us some visibility into this new source of demand.

Residential electricity prices in Texas grew significantly in nominal terms, but only modestly in real terms. This can be surprising to many, but power markets are highly dynamic and multivariate.

First, the obvious factor is supply. Texas added massive quantities of generation over the period, in particular from solar, wind, and battery storage. ERCOT’s market design is also a little different from most other RTOs/ISOs. When wind and solar flood the market, wholesale prices plummet, often into the negatives; retail providers can take advantage of this and offer residential consumers cheap fixed rates. Texas also relies on abundant natural gas generation and natgas prices were generally low over the period.

And lastly, data centers aren’t necessarily competing with residential customers. Data centers often locate themselves in remote locations with an abundance of power (transmission is capacity constrained, so you get islands of cheap power in remote locations that can’t deliver power to urban centers). Hyperscalers are also increasingly going “behind the meter” and collocating generation with their data centers so they can bypass the grid entirely. For instance, OpenAI’s Stargate project involves a hybrid system in Abilene where local generation supplements the grid, and a fully behind the meter campus in Shackleford county which doesn’t rely at all on the public grid. These hyperscalers are incentivized to scale up power as fast as possible, and often this means building their own generation, rather than waiting for grid interconnect (which can take years).

Virginia similarly has added massive quantities of data center demand to its grid in the last decade. Residential power prices have been muted – though trending upwards – over the period. PJM had significant spare capacity over the last decade and this was able to accommodate new data center loads. Hyperscalers are also ideal grid customers. They draw power consistently and predictably, rather than spiking at specific moments (like households with AC). They pay transmission charges and fund interconnection upgrades. They also improve the system load factor, which can actually benefit retail customers.

How is this possible? How can the addition of a new large load lower costs for residential customers? It’s because grids have extremely high fixed costs (wires, substations, transformers, transmission) and relatively low marginal costs. Most of the expense is in building and maintaining the grid. So if you can spread more of these fixed costs across more kilowatt-hours sold, residential rates can decline. All the better if the new buyer is predictable and smooth, rather than spiky and correlated like residential power is.

This isn’t very intuitive, so think about it like the economics of airlines.

The cost inherent in flying is the fleet of planes, the staff, maintenance and the airport fees. If flights are consistently half empty, every passenger is going to have to pay a high ticket price to fund operations. If a new buyer in comes in and promises to buy 1/3 of all the seats for the next 10 years, and the planes are consistently full, the ticket prices for the remaining passengers might be flat or actually fall. Adding net new passengers is relatively cheap – soda, jet fuel, etc – compared to the overall fixed costs of flying.

Of course, this only works to a point. When the grid is completely at capacity, new large loads, regardless of how benign they may be, can strain the grid and require considerable new investment. Virginia is at that point now.

Unfortunately, there is no good source detailing the exact scale of new AI additions to the grid, but it is possible through guesswork to identify the 10 most AI-exposed states. Their real residential electricity rates are shown below.

(The EIA has not provided final 2025 data so we have to use projections.)

Real rates are creeping up in most of these states, but not to the extent you would expect given the breathless press coverage and hostile pronunciations by policymakers. And of course, rates can go up for any number of reasons. According to Brattle, the main factors influencing residential power prices over the last five years are:

• Distribution costs (poles, wires, substations) – for instance, California with very expensive wildfire hardening

• Fuel and wholesale gas prices

• Transmission expansion

• Extreme weather and resilience spending

• Load growth (both helps and hurts)

• State policy choices (decarbonization, permitting, rooftop solar subsidies, labor costs)

I want to be very clear and acknowledge that AI data center load growth has indeed begun to affect rates in some cases (as is evident by the 2025 uptick in the chart above), particularly in the mid-Atlantic. The degree to which it is affecting rates is overstated. Nevertheless, AI strains have become legible in some states:

• PJM interconnect explicitly blamed the AI data center buildout (much of it in Northern Virginia) with significantly higher capacity auction prices in 2025. However, capacity is only a small share of residential retail bills. PJM estimates a 1.5-5% price impact for residential customers. The wholesale market signal is absolutely real, but the retail impact is small. This hasn’t stopped Maryland legislators from being outraged, portraying it as Maryland ratepayers subsidizing Virginia’s data center buildout (Maryland is part of PJM).

• MISO and ERCOT also foresee data centers contributing significantly to future load, but do not characterize them as affecting residential prices today

Most of the fuss is coming from New England, which is purely political posturing from policymakers on the left (there are virtually no AI data centers in NE), or from PJM (VA/MD/PA/NJ) where large loads are the crux of a broader debate about cost-sharing for installations which primarily benefit Northern Virginia.

The fact remains, new large loads are simply not the dominant driver of residential price increases as of today. Let’s zoom out again and consider the totality of evidence over the last five years.

Chart adapted from the original published by Brattle

This is the single most important chart in this entire piece so it’s worth considering it carefully. I have organized states by data center share of electricity usage (colors) and visually represented states by the size of their grid (bubble size). The X axis shows total load growth over the last five years and Y axis shows the inflation-adjusted change in residential electricity prices over the period.

If new large loads were causing price hikes, you’d expect to see a positive correlation on the chart – lots of datapoints clustered at the top right. Instead we see the opposite.

States with lots of load growth over the last few years – Virginia, North Dakota, Nebraska, Nevada, Texas, Iowa, Arizona – all have effectively flat real prices. Some have negative real price change. All of the “high” or “very high” data center share states have muted residential energy price growth; at worst, one or two cents per KWh.

States with huge residential price increases – California, New York, Massachusetts, Maryland – have relatively little load growth and low data center grid penetration.

Selection effects are visible here too. The hyperscalers picked Texas, Nevada, Oregon, Washington, and North Dakota because these states had cheap power. But even so, these huge new loads did not meaningfully drive up power prices in those states.

Put simply, this data does not support the claim that AI data center additions are the primary driver of residential energy price increases. Modest price hikes are evident in moderately data center-exposed states in the mid-Atlantic region (PJM interconnect). Those states are paying for Virginia’s buildout, and the growth is visible there.

Will the data centers bite eventually? It’s certainly possible. Scarcity is developing in the mid-Atlantic and Virginia, as well as MISO (Iowa, Indiana, Illinois). The Carolinas and the pacific northwest are also risks. But it may not necessarily play out this way. The big hyperscalers have almost infinite cash to spend on the buildout with 2025 AI capex reaching $448b. They could offset the residential impact of price increases for 2-3% of their spending, if compelled. They are also building their installations off grid or behind the meter, insulating consumers from price impact.

The cash and willingness is there. The question is whether policymakers will work with the hyperscalers to fund the buildout of our grid infrastructure – which they can afford to do – or opt for easy populist slogans and villainize them and demand moratoriums.

Consumers are justified in being concerned about a large new presence on the grid, but there’s no real evidence this has dramatically shifted electricity prices so far. And it’s not guaranteed that these additions actually hit residential ratepayers in the future, either.

It’s also worth level-setting and reminding people that AI queries are getting vastly more efficient. Through improved hardware and better models, the amount of power required for a standard query is falling dramatically.

Holding a given unit of intelligence fixed, the actual power required to handle a query has declined by a factor of 200 since 2022.

Sources and assumptions in Appendix

Of course, AI data centers in the aggregate are drawing more power as the demand for intelligence is seemingly unbounded, but it’s worth remembering this. The power cost to produce the commodity has collapsed in the last four years. I am not claiming that AI power demand will fall. But it is surely encouraging to see that the resource we all benefit from – intelligence – is getting cheaper to produce at a rate that exceeds almost any industrial commodity in modern history.

So if not AI, what is driving the very real residential power cost increases? We’ve already established that inflation is the most obvious factor. Leaving that aside, the story is different for each grid/state, since they are managed in such disparate ways. I don’t have space to do a full analysis, so if you are interested see this presentation from the LBNL and Brattle group.

But there is one very simple variable that cleanly shows power price discrepancies: state politics, especially regarding decarbonization or the climate transition.

Before you complain, this chart takes the median of blue/red states, so it excludes crazy outliers like Hawaii that would bias the blue state data upwards. (Some people might demand I weight it by population – that’s actually much worse for the blue states.)

Blue states as a group have higher – and faster rising – residential electricity rates. This is a simple fact. The correlation of electricity price growth to state politics is shockingly high, if you look at the state by state view.

Why is this happening?

Some states, like California or Hawaii, have their own idiosyncratic issues like fire hardening or remoteness which worsen things significantly. And resource endowment (presence of local resources like hydro, gas, coal, or wind and solar) explains some of the variance. And highly-urbanized blue states often have older grids that are more expensive to maintain.

But it is nevertheless true that blue states, by and large, pursue energy policies that directly lead to higher household electricity prices.

1. Denuclearization

Six major nuclear plants were retired in the Northeast and Midwest between 2013 and 2022, notably Indian Point in NY, San Onofre in CA, and Pilgrim in MA. These losses were backstopped with more expensive natgas generation. This unforced error (nuclear, of course, is carbon neutral) directly translated into higher residential rates and higher grid carbon intensity in those states.

2. Pipeline opposition

The Northeast – the epicenter of high power prices in the US – has systematically blocked pipelines over the last fifteen years. Multiple pipelines were approved and then shot down politically. These would have imported cheap Appalachian natural gas to the most expensive energy market in the country. New England has filled the hole by… importing LNG from Trinidad and Qatar. Instead of piping gas 200 miles from the Marcellus shale gas field, the largest in North America. New England’s winter prices are catastrophic. This is the direct consequence of environmentalist lobbies in NY and MA.

3. Renewable portfolio standards

Blue states have put in place aggressive clean energy requirements, while increasing load through electrification. These states mandate that utilities reach certain renewable energy thresholds. New York for instance mandates 70% renewables by 2030 and 100% zero emission by 2040. California requires 60% by 2030 and 100% by 2045. California is betting heavily on utility-scale solar plus battery storage. New York and New England are pursuing offshore wind, with some meagre solar and hydro imports backstopping the portfolio. The trouble is that wind and solar are intermittent, require backstopping (from batteries, natgas, or demand response programs), and lots of expensive transmission. Offshore wind is fundamentally expensive and not likely to suffice on its own. Getting intermittent and unpredictable renewable generation to match demand profiles while keeping residential costs low is a significant challenge, and one that no country or polity has achieved to date. Several European countries have achieved significant non-hydro renewable penetration on their grids, but none of them have done this AND managed to keep power costs low (by American standards).

4. NIMBYism and labor costs

For grids to modernize and grow, you have to be able to build plants, transmission, substations, and so on. Blue states tend to mire energy projects in environmental review, activist lawsuits, municipal opposition, and so on. Transmission in particular is vital to renewable transitions because urban demand centers are often not co-located with sources of wind or solar. Yet in blue states, long-range transmission is politically toxic and hardly ever gets built. New York and California are great examples of places with theoretically abundant energy resources that is trapped by the political challenge of building transmission. Additionally, labor costs stymie grid buildouts. Blue states tend to require more expensive union labor and stricter rules. Put simply, blue states are trying to run a large-scale grid overhaul while making it virtually impossible to build new infrastructure at scale. Given the greater grid demands of electrification, is no surprise that electricity costs in many blue states are high and rising.

And these aren’t conspiracies. These are the stated policies of political leaders in California and the North East. This is what voters ask for and this is what they get. Blue states consistently prioritize union labor, environmental standards, clean energy, denuclearization, rigid property rights, beautification, and the veto power of activists and NGOs. This mechanically leads to slower energy infrastructure construction and as a consequence higher energy prices.

So when politicians like Elizabeth Warren complain about AI data centers driving up prices, not only are they overstating the reality, but they are shifting the blame from their own costly policies to a new unknown factor which people are nervous about. The rising energy costs borne by households decompose into inflation and decarbonization/NIMBYism. They have little – so far – to do with new AI data centers. The load demands of these data centers will bite in the coming years, but this transition should be welcomed. The hyperscalers are more than able to fund the required energy buildout themselves. If state policymakers work with them rather than demonizing them or demanding moratoriums, ratepayers can be net beneficiaries of the AI capex boom, rather than its victims.

---

Appendix: sources and assumptions

Every single chart in this piece was built with Claude Opus 4.7. All underlying data was carefully checked. To avoid accusations of bias or data manipulation, I want to give you the ability to independently re-derive every single chart in this piece. The sources and assumptions data should be sufficient to fully reconstruct each chart.

Figure 1: US electricity generation vs. data center consumption

• Total US electricity generation: EIA Monthly Energy Review (Nov 2025 release) and EIA “U.S. electricity generation in 2025 hit a record” (Mar 2026). Net generation, all sectors, including small-scale solar. 2025 figure of 4,430 TWh is preliminary.

• All US data centers, 2014–2023: LBNL 2024 U.S. Data Center Energy Usage Report (Shehabi et al., DOE-commissioned). Hard anchors: 58 TWh (2014), 76 TWh (2018), 176 TWh (2023). Intermediate years interpolated using LBNL-reported CAGRs (7% for 2014–18, 18% for 2018–23).

• All US data centers, 2024–2025: 2024 from IEA estimate (183 TWh) cross-referenced with LBNL Dec 2024 update (range 184–231 TWh); midpoint ~200 TWh used. 2025 extrapolated along LBNL’s 13–27% projected growth corridor toward their 2028 range of 325–580 TWh.

• AI-related portion, 2017 + 2023: LBNL accelerated-server breakdown — ~2 TWh (2017) and ~40 TWh (2023). 2018–2022 interpolated.

• AI-related portion, 2024–2025: 2024 from AImultiple analysis citing IEA/EPRI (53–76 TWh range, midpoint 65). 2025 extrapolated.

• Pre-2017 AI: assumed ≤2 TWh; reflects negligible AI-specific GPU footprint before transformer-era training at scale.

• “AI-related”: GPU-accelerated server consumption attributable to AI training and inference. Excludes general cloud compute, video streaming, search, crypto.

Figure 2: Real residential electricity rates by data-center-intensity quintile, 2015-2025

• Residential electricity prices, 2015–2024: EIA State Energy Data System (SEDS) Table ET3, residential electricity column (Form EIA-861). Final values, October 2025 release.

• Residential electricity prices, 2025: EIA Electric Power Monthly, monthly state retail prices (Form EIA-861M), compiled in NEADA’s November 2025 Energy Price Update, Tables 1–2 (August snapshots).

• CPI deflator: BLS CPI-U, annual averages (series CUUR0000SA0).

• Data-center intensity: EPRI, “Powering Intelligence” (May 2024), Table A1 — 2023 data-center share of state electricity, 44 reporting states.

• 2025 state values extrapolated from 2024 baseline × (Aug 2025 nominal / Aug 2024 nominal); assumes the August YoY ratio approximates the calendar-year YoY ratio.

• Quintiles built from 50 jurisdictions (49 states + DC; Hawaii excluded) on 2023 data-center share of state electricity; 10 states per bin.

• Six states excluded from EPRI Table A1 (AK, AR, DE, MS, VT, WV) plus DC imputed at 0% DC share — all land in Q1.

• Quintile assignment is fixed per state (by 2023 DC share), not re-computed each year.

• Quintile line = mean of member states’ real prices. Hawaii excluded as a structural outlier (~40¢/kWh, non-contiguous petroleum-fired grid).

Figure 3: Residential electricity prices, nominal vs real, Q5 (highest data-center intensity), 2015-2025

• Residential prices, 2015–2024: EIA State Energy Data System (SEDS) Table ET3, residential column / Form EIA-861, October 2025 final release.

• Residential prices, 2025: EIA Electric Power Monthly (Form EIA-861M), August snapshot compiled in NEADA’s November 2025 Energy Price Update.

• Inflation deflator: BLS CPI-U, annual averages (series CUUR0000SA0).

• Q5 state selection: EPRI, Powering Intelligence (May 2024), Table A1 — 2023 data-center share of state electricity consumption.

• Q5 = 10 states with the highest 2023 data-center share: VA, NV, ND, NE, IA, OR, TX, IL, AZ, WY.

• Quintile membership fixed at 2023 share; not recomputed annually.

• Series shown is the unweighted mean of the 10 member states’ residential rates (no population or sales weighting).

• Real prices = nominal × (CPI-U 2024 / CPI-U year); base year 2024.

Figure 4: State electricity prices and data center share of electricity

• Residential electricity prices (Aug 2025, ¢/kWh) — NEADA Energy Price Update (Nov 2025), citing EIA Form EIA-861M

• Data-center capacity and annual energy use by state (2024 historical) — EPRI, Powering Intelligence: Updated U.S. Data Center Scenarios (Feb 2026), public dashboard export

• State total electricity consumption baseline — EIA, Electric Power Annual 2023, scaled forward 1% to 2024

• DC share of state electricity = 2024 EPRI TWh ÷ estimated 2024 state total (2023 EIA total × 1.01)

• “Operational” data-center capacity = EPRI Nominal IT Capacity; peak load is typically ~half of nominal

• Aug 2025 residential prices are EIA preliminary figures; final values may revise by ±0.2¢

• EPRI 2026 estimates revised some 2023 state figures vs. the May 2024 EPRI paper — notably ND ↓, OH and IN ↑, PA and MI ↓ — reflecting methodology updates, not actual swings

• States with ≈0% bubbles (AR, WV, VT, MS, DE, LA) have negligible operational data-center load as of 2024; several have large announced projects not yet built

Figure 5: Texas large flexible load and residential electricity prices, 2018-2025

• Bars 2022–24: ERCOT Large Flexible Load Task Force monthly reports — approved interconnection capacity for loads ≥75 MW per site

• Bars 2025: EIA Short-Term Energy Outlook (Sept 2024) baseline of 9,500 MW; low/high scenarios are 6,500 / 14,200 MW

• Bars 2018–21: reconstructed from LBNL 2024 Data Center Energy Usage Report (Texas state estimates), public crypto miner filings (Riot, Marathon, Core Scientific, Bitdeer, Cipher), and contemporaneous press coverage

• Prices: EIA Electric Power Monthly (sales-weighted Texas residential), deflated by BLS CPI-U annual averages to 2024 dollars

• Y-axis is approved capacity — MW cleared to operate via ERCOT’s study process, not metered peak demand and not nameplate

• 75 MW per-site threshold is ERCOT’s; smaller crypto sites and sub-threshold data centers are not counted

• Pre-2022 ERCOT had no comprehensive tracking of these loads; those bars are best-effort reconstructions, not measurements

• Single load category — no crypto / traditional / AI split, because ERCOT does not categorize loads this way and any breakdown requires analyst judgment

• Behind-the-meter loads (on-site gas microgrids serving data centers, e.g., Shackelford Stargate) are excluded; actual Texas computing load is higher than the bars show, and the gap is growing

Figure 6: Virginia data center load vs. Virginia residential electricity prices, 2018-2025

• MW (left axis): Dominion Energy historical data center billing demand, 2018–2024 actual + 2025 forecast (4,149 MW), from 2025 testimony in an Indiana IURC regulatory filing. 2024 = 3,584 MW cross-checks with Feb 2025 SCC testimony.

• Prices (right axis): EIA Form 861M Historical State Data (HS861M 2010-.xlsx), sales-weighted annual residential averages for Virginia. 2018–2024 are Final; 2025 is Preliminary with all 12 months reported. 2024 value of 14.41¢ matches EIA’s annual Table 5A exactly.

• Real-dollar deflation: BLS CPI-U all-items annual averages, 2024 anchor (313.689); 2025 ≈ 320.0 per BLS December release.

• 2025 MW is Dominion’s own forecast, not measured; recent forecasts have tracked actuals within a few percent.

• 2025 price is EIA Preliminary; final value may revise by a few tenths of a cent.

• Dominion territory covers roughly two-thirds of Virginia’s data center footprint; the price series is statewide residential across all utilities. Geographic match is close but not perfect.

• AI vs. non-AI workloads are not separated in the MW bars — these are total data center billing demand.

Figure 7: Real residential electricity rates, top 10 Al data-center states, 2015-2025

• Annual values 2015–2024: EIA State Energy Data System (SEDS) and Form EIA-861, sales-weighted residential average price for each state, finalized

• 2025 estimates: EIA Electric Power Monthly state residential series, Jan–Aug 2025 (8 monthly observations averaged)

• Inflation deflator: BLS CPI-U, annual averages, base year 2024

• State selection: EPRI 2024 “Powering Intelligence” report, Table A1 (state share of US data center load as of 2023), supplemented with announced and operational AI hyperscale campuses through May 2026

• “Top 10 AI data-center states” combines an existing-load metric (EPRI 2023 share) with a forward-looking signal (announced AI campuses) — different methodology would shuffle the bottom of the list

• All prices deflated to 2024 dollars using national CPI-U, not state-level cost-of-living indices — regional inflation differences are ignored

• 2025 YTD-through-August captures 8 of 12 calendar months; final SEDS values will include Sep–Dec 2025 and are expected to land somewhat higher for PJM-exposed states (VA, OH) where capacity prices stepped up June 1, 2025

• Residential rates only — commercial and industrial trends differ

• Sales-weighted averages mask wide intra-state variation between utility service territories

• Final 2025 SEDS data publishes late 2026; until then, all 2025 figures on this chart are preliminary estimates

Figure 8: Load Growth vs. Real Residential Electricity Price Change, 2020-2025

• Prices & sales: EIA Electric Power Monthly, Tables 5.6.B (residential prices) and 5.4.B (total retail sales). 2020 final, 2025 preliminary.

• Inflation: BLS CPI-U annual average.

• Data center load: LBNL 2024 Data Center Energy Usage Report, EPRI 2024, JLARC Dec 2024 (Virginia), IEA 2024 state shares, Liu et al. (arxiv 2024).

• Prices are annual-average residential, ¢/kWh, expressed in 2025 dollars (deflator = 1.239).

• Load growth is all-sector (res + commercial + industrial), 2020 → 2025.

• 2020 baseline is mildly COVID-suppressed, inflating growth ~1–2 pp vs. a 2019 baseline.

• 2025 EIA data are preliminary (final Oct 2026).

• State data center load is estimated, not measured — EIA’s 2024 mandate to collect it was withdrawn. National ~200 TWh in 2025; VA most reliable, small-share states roughest.

• Color bins: Low <2%, Medium 2–5%, High 5–10%, Very High >10% of state electricity.

Figure 9: Energy per query for a fixed amount of intelligence

• Capability threshold: MMLU 70%, the score GPT-3.5 hit at launch. Held constant across all six points so each represents the same task difficulty.

• Methodology: at each anchor, identify the cheapest model meeting MMLU 70% on commercial serving infrastructure of that moment. Estimate per-query energy from active parameters × bytes per parameter (precision) × memory bandwidth × power, at typical batch sizes, with PUE 1.3.

• Calibration anchors: HazyResearch’s Intelligence per Watt study (Saad-Falcon et al., Stanford, 2025), which directly measured a 5.3× IPW improvement over 2023–2025; and InferenceMAX (SemiAnalysis, Dec 2025), which gives measured energy per query for currently optimized cloud serving.

• Implied rate: ~120× over 3.5 years ≈ 3.8×/year. Consistent with HazyResearch’s ~2.3×/year measured rate plus compositional effects (distillation, MoE, lower precision).

• Query length: assumes ~500 output tokens. Longer outputs (reasoning models, agentic flows) scale roughly linearly.

• Energy ≠ price: API prices fell ~10×/year over the same period, faster than energy, because margins also compressed. The energy claim is the more conservative one.

• Weakest point: gpt-3.5-turbo (Q1 23). OpenAI hasn’t disclosed its parameter count; the ~20B distilled estimate is widely cited but unverified.

• MMLU’s limits: a multiple-choice knowledge test, not a complete proxy for intelligence. A model can match MMLU 70% while losing on long-context reasoning, tool use, or anything outside the benchmark’s distribution.

• 2026 hardware assumes B200 deployment: optimized cloud serving. A real-world deployment on older hardware (H100, A100) would land closer to ~75 J instead of ~25 J.

Figure 10: Real residential electricity rates, blue vs. red states (2024 election), 2015-2025

• Annual residential rates 2015–2024: EIA State Energy Data System (SEDS) and Form EIA-861, sales-weighted state averages, finalized

• 2025 rates: EIA Electric Power Monthly, state residential series, Jan–Aug 2025 averaged (8 monthly observations)

• Inflation deflator: BLS CPI-U annual averages, base year 2024

• State election grouping: 2024 presidential statewide vote winner (AP, official state canvasses)

• Each state counts as one observation regardless of population — this is a “state policy regime” framing, not a “typical resident” framing; population-weighted alternatives shift blue median up ~4¢

• DC included in blue group (n=20); 31 states in red group

• Sales-weighted state averages mask large intra-state variation between utility territories (e.g., Dominion VA vs. NOVEC VA, PG&E CA vs. SMUD CA)

• 2025 figures are preliminary; final SEDS values publish late 2026 and will incorporate Sep–Dec 2025 data not yet available

• 2024 election grouping is applied retroactively to the full 2015–2025 series — states that flipped (e.g., MI, PA, WI, GA, NV) are classified by their 2024 outcome throughout, which is anachronistic but consistent

• Residential rates only; commercial and industrial sector dynamics differ

Figure 11: Real change in residential electricity prices, 2019-2025 (state by state)

• Residential prices, 2019: EIA Electric Power Annual 2019 (final release), state-level average residential rates.

• Residential prices, 2025: EIA Electric Power Monthly, Table 5.6.A, February 2025 (preliminary), used as a 2025 proxy.

• Inflation deflator: BLS CPI-U, 2019 annual average vs. February 2025 monthly value.

• Dot color (party): 2024 presidential election results, AP / state certifications.

• Endpoint convention: 2019 = annual average; 2025 = February monthly snapshot, treated as a full-year proxy. This mixes an annual figure with a single month and assumes February rates approximate the calendar-year average.

• Real % change = (2025 nominal / 2019 nominal) × (CPI-U 2019 / CPI-U Feb 2025) − 1.

• Preliminary February 2025 EIA values are subject to later revision.