Is power load growth stalling?

What weather-normalized EIA 930 data suggests about recent power trends

Power load growth has been the biggest driver of gas demand growth in recent years, well before coal retirements began to slow. Data centers and electrification have supported both accelerating growth and an even more bullish narrative: that power demand is on a structurally stronger trajectory, and gas demand will follow.

But that assumption is starting to look less secure. Because power generation is highly weather-dependent, interpreting short-term changes in load requires careful normalization for weather and seasonality. Using daily EIA 930 data to extend the load picture beyond where EIA 923 reporting ends, I build a daily model for a real-time read on the trend that matters most for gas demand. And it suggests that after adding ~90 TWh (~1.9 Bcfed) in the first half of 2025, power load growth decelerated, and possibly even plateaued, in the second half of last year.

Building a weather-normalized model

Because power generation is highly weather-dependent, it is essential to first build a weather-based model of power loads to estimate the extent to which day-to-day variation reflects weather trends versus sustainable structural growth. Using the EIA’s 13 regions and NOAA temperature data, I model demand as a function of HDDs, CDDs,¹ and dummy variables to account for weekend and holiday effects. I also discount HDDs and CDDs if the prior week’s average temperature has been relatively moderate. Or in real-world terms, a single 80-degree day after a week of temperatures in the mid-60s probably won’t trigger widespread air conditioning usage, whereas consistent 80-degree temperatures would. This gives us a first-pass estimate of what load would have been like as a function of weather conditions.

Then, using the coefficients from each model, along with the difference in a given day’s temperature and normal weather for that day of the year, we can estimate a weather-normalized daily power generation time series. However, weather normalizing relative to typical temperatures on that day wouldn’t tell us much, because power generation levels are so structurally different throughout the year. In other words, a 4,000 TWh run-rate² day in March would be extremely bullish, whereas a 4,000 TWh run-rate day in July would be catastrophically bearish.

Understanding the structural trend, then, means we need to normalize each day’s demand relative to an average day for the year as a whole. Normalizing to that average day — which is two-sevenths of a weekend day and a small fraction of a holiday in addition to average HDD and CDD levels — then gives us an indication of what a given day’s power generation level tells us about what we would expect power demand to be over the course of an entire year. Figure 1 shows the (annualized) raw demand in gray alongside the weather-normalized run-rate demand in pink. The 30-day moving average (in teal) is even less noisy.

Figure 1 | Annualized Lower-48 power demand

However, this run-rate power demand estimate still has some underlying seasonality, indicating that households or businesses are less likely to use electricity at different times of the year, even after accounting for differences in weather. As shown by the black line in Figure 2, power demand remains seasonally weak in spring, even after accounting for lower HDD and CDD levels. Adopting different HDD or CDD thresholds degrades the model’s overall explanatory power, so it is likely that these consistent errors reflect underlying behavioral differences from consumers.

Figure 2 | Power demand model errors

The year-to-year consistency in these errors, though, means that we can explicitly incorporate a seasonal adjustment to get a better estimate of run-rate power loads. The upward shift in errors from year to year illustrates the underlying structural strengthening in power demand levels.

But we can view this underlying demand more easily by using the seasonality in Figure 2 to adjust the run-rate power demand from Figure 1. The result is the purple line in Figure 3, which indicates where the latest power demand data suggests annual power demand levels will end up, accounting for seasonal and weather factors.

Figure 3 | Run-rate Lower-48 power demand

From this chart, we observe a few important trends:

• Between late 2024 and the second quarter of 2025, we see a consistent pattern of strong growth through the first half of last year: ~90 TWh, or ~1.9 Bcfed.

• Consistent results with the annual model: between 2023 and 2024, I estimate ~75 TWh (~1.6 Bcfed) of growth, which is consistent with the ~90 TWh from the annual model based on EIA 923 data. However, the daily model allows us to isolate the timing of this growth, which occurred most significantly in late 2023, with more modest growth before and afterward.

• However, this growth, at best, decelerated and may even have reversed slightly in the second half of 2025. This weaker recent trend may reflect supply-chain challenges for data centers in particular. But it could also reflect a weakening in overall economic activity, with leading indicators pointing to decelerating growth.

Where the growth is coming from

Regionally, ERCOT has led the way, accounting for ~60 TWh of growth since the beginning of 2023. PJM, meanwhile, accelerated sharply in the first half of 2025 but has been flat since. In the Southwest, where electric customers are backstopping a major expansion of Transwestern, power demand growth has been consistent. In percentage terms, Southwest demand growth has been in line with ERCOT’s, whereas PJM’s is less impressive, given the ISO’s large territory.

Figure 4 | Run-rate power demand growth since January 2023

Given the importance of power loads to gas demand growth, translating the EIA 930 data into a real-time estimate of power loads, net of weather and seasonal impacts, will be an important leading indicator of gas market tightness or looseness. Whether this plateau reflects data center supply-chain delays or broader economic softening, it’s a trend worth watching. It’s one I’ll track closely in the months ahead to see whether the recent plateau will reverse or presage a shift back toward flatter loads.

1. Heating-degree days and cooling-degree days, defined as the difference in a day’s average temperature and 65 degrees, but zero if the difference is negative.

2. Multiplying daily generation by 365 for an annualized run-rate figure