# U.S. AI electricity and household equivalents: September 2026 update

Research cutoff: 22 September 2026. Main comparison year: 2025; 2026 is a projection. All numerical household equivalents below match the electricity numerator's year and U.S. region.

## Result and boundary

A newer U.S. all-data-center comparison is supported. A comparable 2025/2026 U.S. AI-only annual total including attributable facility overhead was not verified in the public primary evidence retrieved. This is a verification limit, not proof that no model or dataset exists. Missing AI values must not be plotted as zero or replaced with all-data-center totals.

## Inputs and calculations

Formula: million residential customer-years = electricity in TWh × 1,000 / annual kWh per residential customer.

| Year | Electricity definition and source | TWh/year | Denominator kWh/customer/year | Million customer-years |
|---|---|---:|---:|---:|
| 2025 | All data-center IT equipment, IEA Table A.4 | 173 | 10,479 | 16.5092 |
| 2025 | All data centers including facilities, IEA Table A.4 | 224 | 10,479 | 21.3761 |
| 2025 | All data centers including facilities, Berkeley Lab Reference Case, derived | 234.24 | 10,479 | 22.3533 |
| 2026 | All data centers including facilities, Berkeley Lab Reference Case, derived projection | 302.1696 | 10,487 | 28.8137 |
| 2024 | AI-specialized servers, approximate visual reading of supplied Figure 9 | 50 | 10,356 | 4.8281 |
| 2024 | Those AI servers plus allocated facility overhead | 57.25 | 10,356 | 5.5282 |

The Berkeley Lab values are reconstructed from its rounded 192 TWh 2024 baseline and rounded Reference Case growth rates: 192 × 1.22 = 234.24; then × 1.29 = 302.1696. Report approximately 234 and 302 TWh, not the extra arithmetic digits as source precision. These are separate model results, not the endpoints of a statistical uncertainty interval.

The historical AI calculation is 50 × 1.145 = 57.25 TWh. The factor is Berkeley Lab's estimated 2024 PUE for facilities serving AI equipment, not the all-data-center PUE of 1.45. PUE assigns cooling, power-distribution losses and other facility infrastructure proportionally to the selected IT load. It does not add separately allocated external networking or storage. This is therefore a servers-plus-overhead estimate, not a complete measurement of all AI-related electricity. The 50 TWh is a chart-reading estimate of the AI segment thickness, not its stacked upper boundary. An illustrative 45–55 TWh reading range would yield 4.98–6.08 million customer-years including overhead; it is not a confidence interval.

## What exactly is missing for a newer AI-only result?

For each of 2025 and 2026, the missing verified inputs are:
1. U.S. annual AI-specialized-server electricity in TWh, with scope and model vintage.
2. A matching-year AI-facility PUE or directly reported attributable facility electricity.
3. For a complete AI facility footprint, attributable external storage/networking electricity and a clear accounting rule.

Berkeley Lab Figure 9 shows 2018, 2024 and 2030 equipment splits, not 2025/2026. Its 2030 AI share must not be applied to 2025. The IEA regional table supplies total and IT-only U.S. demand but not a U.S. annual AI split. Neither global AI shares nor all-center growth rates establish U.S. AI-only electricity.

Additional candidate sources checked included Gartner's June 2026 public release (global equipment breakdown), Xiao et al.'s 2025 Nature Sustainability study, and Schneider Electric's 2025 scenario report. These do not justify asserting that no forecasts exist. Their exact 2025/2026 U.S. values and compatible overhead boundaries were not verified for this calculation; Schneider's full PDF could not be retrieved through the available web access. No unverified numbers from secondary reports or snippets were plotted.

## Measurement and household definitions

- Data-center totals are modeled estimates, not a nationwide sum of AI-meter readings.
- EIA September 2026 STEO Table 7a reports mean residential electricity usage per customer of 10,479 kWh for 2025 and 10,487 kWh for 2026. Use 2025 as a historical estimate subject to revision and 2026 as a full-year forecast, incorporating available year-to-date information.
- The retained 2024 denominator is EIA final annual Table 5A's 863 kWh per month × 12 = 10,356 kWh/year.
- “Homes” here means residential utility-customer equivalents. Customer accounts are not identical to occupied households or individual people.
- This is a mean purchased-electricity comparison, not median household consumption, total household energy including gas, or a survey estimate of gross electricity use including self-consumed rooftop generation.
- A median or occupied-household alternative requires matching 2025/2026 data and its exact population/energy boundary; no unverified denominator was substituted.
- Denominator sensitivity is inverse: with a denominator 10% lower, the count rises 11.1%; with one 10% higher, it falls 9.1%. This is illustrative algebra, not a measured alternative.
- Equivalence concerns annual electricity quantities, not a claim that these households are literally powered or displaced. Upstream grid losses and embodied manufacturing/construction energy are not added.

## Primary sources

1. IEA, Key Questions on Energy and AI, published 16 April 2026, Annex A Table A.4:
https://www.iea.org/reports/key-questions-on-energy-and-ai
PDF: https://iea.blob.core.windows.net/assets/3179f7f8-01f6-4dd6-bffa-c9f7b73f1dc9/KeyQuestionsonEnergyandAI.pdf

2. Berkeley Lab, United States Data Center Energy Usage Report: 2025 Update, Figure 9, section 3.5 and conclusion:
https://www.energy.gov/documents/united-states-data-center-energy-usage-report-2025-update
User-supplied Figure 9 image: inputs/figure9-check-4d90ae646acc.png; upload version 856b52dc-d062-4787-8dce-3f86c0e61f43.

3. EIA, September 2026 Short-Term Energy Outlook, Table 7a (modeling completed 3 September):
https://www.eia.gov/outlooks/steo/tables/pdf/7atab.pdf
https://www.eia.gov/outlooks/steo/
The current table URL rolls forward; the values used are frozen above.

4. EIA, 2024 Electric Sales, Revenue, and Average Price, Table 5A:
https://www.eia.gov/electricity/sales_revenue_price/

5. Candidate newer AI forecast sources, not used as numerical inputs:
https://www.gartner.com/en/newsroom/press-releases/2026-06-10-gartner-says-data-center-electricity-demand-to-grow-26-percent-in-2026
https://www.nature.com/articles/s41893-025-01681-y
https://www.se.com/ww/en/insights/sustainability/sustainability-research-institute/powering-sustainable-ai-in-the-united-states/

## Reproduction

```python
household_kwh = {2024: 10356, 2025: 10479, 2026: 10487}
rows = [(2025,173), (2025,224), (2025,192*1.22),
        (2026,192*1.22*1.29), (2024,50), (2024,50*1.145)]
for year, twh in rows:
    print(year, twh, twh*1000/household_kwh[year])
```

The chart uses a common zero-based horizontal scale with separate year panels. The image was rendered with matplotlib, checked for text outside the canvas, and visually inspected. No AI-only 2025/2026 marker is present because the value is unverified.
