# U.S. AI-only electricity: 2025 and 2026 household equivalents

Updated 22 September 2026. This replaces the previous “AI-only values not verified” gap with a specific published model projection, using the supplied Figure 1 source data. It does not turn those projections into measured consumption or update the original model's assumptions to 2026.

## Results

| Scenario | 2025 TWh/year | 2025 million customer-years | 2026 TWh/year | 2026 million customer-years |
|---|---:|---:|---:|---:|
| Low Demand | 64.7734 | 6.181258 | 117.8969 | 11.242195 |
| Low Power | 64.7734 | 6.181258 | 117.8969 | 11.242195 |
| Mid-case | 64.7734 | 6.181258 | 117.8969 | 11.242195 |
| High Application | 64.7734 | 6.181258 | 117.8969 | 11.242195 |
| High Demand | 62.8569 | 5.998368 | 115.9539 | 11.056918 |

Digits reproduce spreadsheet precision, not empirical accuracy. The narrow early-year scenario spread is not a confidence interval or a complete uncertainty assessment. Scenario names are preserved even when their early-year numerical ordering differs.

## Cooling and scope

The paper's Methods, “Assessment of the environmental footprints of AI servers,” calculates electricity from server usage and regional PUE. Figure 1c separates server and infrastructure energy; the supplied sheet gives shares 0.89045411 and 0.10954589. Figure 1e includes this overhead: no extra PUE multiplier is applied. The aggregate Figure 1c share is not treated as a verified year-specific split.

This model concerns top-tier AI servers, using mainly Nvidia DGX specifications and supply-chain assumptions. It is not a census of every AI workload, general-purpose server, or external network/storage system. Its operational boundary excludes embodied manufacturing impacts. It should not be spliced into Berkeley Lab's series as if scope and methods were identical.

## Household denominator and formula

Latest available EIA release checked: September 2026 Short-Term Energy Outlook, Table 7a, modeling completed September 3.

- 2025: 10,479 kWh per residential customer per year; historical estimate subject to revision.
- 2026: 10,487 kWh per residential customer per year; full-year forecast.

Million residential customer-years = TWh × 1,000 / annual kWh per customer.

Thus 64.7734 × 1,000 / 10,479 = 6.181258 million (2025), and 117.8969 × 1,000 / 10,487 = 11.242195 million (2026).

Both numerator and denominator describe the United States and the same calendar year. “Homes” is shorthand for utility-customer equivalents using average purchased electricity. It is not a median household, an occupied-household count, gross consumption including self-consumed rooftop generation, or total household energy including gas. No household denominator from another year was substituted.

## Provenance

- Uploaded file: xiao-figure1-source.xlsx
- Notebook input: inputs/xiao-figure1-source-9271ff9e9a3d.xlsx
- Input upload version: 95f1cb44-b70c-41ae-86ef-0833ca4803d8
- SHA256: 9271ff9e9a3d48016c12e383c7fbe81c760957c62184cd67ee002ee08d4fae48
- Extraction: Figure1(e)!C3:G4; scenario headers C1:G1; years A3:A4.
- Column B, “Reference Value,” is excluded: the figure caption identifies the reference as an older all-data-center comparison, not an AI scenario.
- No values were digitized or visually estimated from a chart.
- Supplied workbook was analyzed as the user-provided source data; no byte-for-byte comparison with a publisher download was performed.

Paper: Xiao et al., published 10 November 2025, DOI 10.1038/s41893-025-01681-y.
https://www.nature.com/articles/s41893-025-01681-y

EIA Table 7a (rolling URL; September 2026 values frozen above):
https://www.eia.gov/outlooks/steo/tables/pdf/7atab.pdf

## Reproduction

```python
import pandas as pd
data = pd.read_excel(
    'inputs/xiao-figure1-source-9271ff9e9a3d.xlsx',
    sheet_name='Figure1(e)'
).rename(columns={'Unnamed: 0': 'year'})
scenarios = ['Low Demand', 'Low Power', 'Mid-case', 'High Application', 'High Demand']
result = data.loc[data.year.isin([2025, 2026]), ['year'] + scenarios].melt(
    id_vars='year', var_name='scenario', value_name='electricity_twh'
)
result['eia_kwh_per_customer'] = result.year.map({2025: 10479, 2026: 10487})
result['million_customer_years'] = result.electricity_twh * 1000 / result.eia_kwh_per_customer
```

Chart QA: verified source rows and excluded reference series; computed all five scenarios; checked text bounds and visually inspected the rendered chart. The chart groups the four identical scenarios while retaining the fifth separately.
