Some sources failed to update; figures from the previous fetch are shown: Ember (world generation)
FORECAST TRACK RECORD
How official power forecasts have missed
Key points
- Official forecasts miss in a consistent direction. Every EIA edition from 2005 to 2015 (11 of 11) expected US power use to grow faster than it did: a median 1.0% a year against 0.3% in reality. Solar and wind were set too low in almost every edition, in the US and worldwide.
- But demand forecasts are not always too high. From the 2016 edition all 6 judged editions were too low. The real pattern is lateness: forecasts move a little each year and trail reality at turning points, first when growth stopped around 2007, then when it came back.
- Today's “surge” forecast, PJM's 3.6% a year, still means about 1.8% a year even after subtracting the worst over-forecast on record (1.8 points a year). That is far from the flat 2010s. Meanwhile the latest EIA outlook (1.3% a year) is already being outrun: it expected +1.1% for 2025, and use grew +2.1%.
The scorecard
Each past edition’s forecast for 10 (or 5) years after it was published, compared with what actually happened. “Too high” means at least 8 in 10 editions overshot; “Too low” means at least 8 in 10 undershot.
| US (EIA), 10 years on | US (EIA), 5 years on | World (IEA), 10 years on | World (IEA), 5 years on | |
|---|---|---|---|---|
| Power useworld: total generation | Too high11 of 11 editions; median +8% | Too high13 of 16 editions; median +3% | No clear lean8 of 17 too high; median -1% | No clear lean9 of 20 too high; median 0% |
| Solar output | Too low11 of 11 editions; median -83% | No clear lean5 of 16 too high; median -38% | Too low6 of 6 editions; median -73% | Too low11 of 11 editions; median -42% |
| Wind output | Too low11 of 11 editions; median -64% | No clear lean4 of 16 too high; median -34% | Too low12 of 12 editions; median -33% | Too low17 of 17 editions; median -18% |
| Gas-fired output | Too low11 of 11 editions; median -33% | Too low16 of 16 editions; median -22% | Too low14 of 17 editions; median -6% | Too low18 of 20 editions; median -6% |
| Coal-fired output | Too high11 of 11 editions; median +117% | Too high15 of 16 editions; median +46% | No clear lean10 of 16 too high; median +4% | No clear lean12 of 20 too high; median +2% |
| Nuclear output | Too high10 of 11 editions; median +5% | No clear lean12 of 16 too high; median +2% | No clear lean12 of 17 too high; median +20% | Too high16 of 20 editions; median +14% |
Two things stand out. Supply-side forecasts lean hard and the same way every time: solar and wind too low in the US and worldwide, and in the US coal too high and gas too low. Demand forecasts lean in the US (11 of 11 too high 10 years on) but not for the world as a whole (8 of 17 too high; the median miss is -1%). The world never had a US-style flat decade, because growth in China and other emerging economies carried on.
US solar forecasts include small rooftop panels, while the actual figures (EIA Monthly Energy Review) count only power plants. That difference makes forecasts look higher, so the “too low” result is on the cautious side. US editions 2005–2025 (reference case) against actuals up to 2025; IEA WEO editions 1993–2022 (central scenario) against Ember up to 2025. IEA projections are given every five years, so the years between are joined on a log scale.
Source: EIA AEO Retrospective / EIA / Lopez et al. 2025 / Ember
Power use: too high, then too low
Show the numbers
| Edition | Years compared | Expected | Actual | Actual − expected |
|---|---|---|---|---|
| 2005 | 2004–2014 | 2.09% | 0.60% | -1.49% |
| 2006 | 2005–2015 | 1.58% | 0.26% | -1.32% |
| 2007 | 2006–2016 | 1.56% | 0.25% | -1.31% |
| 2008 | 2007–2017 | 1.03% | -0.11% | -1.14% |
| 2009 | 2008–2018 | 0.84% | 0.33% | -0.51% |
| 2010 | 2009–2019 | 1.19% | 0.58% | -0.61% |
| 2011 | 2010–2020 | 0.59% | -0.10% | -0.69% |
| 2012 | 2011–2021 | 0.47% | 0.15% | -0.32% |
| 2013 | 2012–2022 | 0.86% | 0.61% | -0.24% |
| 2014 | 2013–2023 | 1.08% | 0.39% | -0.69% |
| 2015 | 2014–2024 | 0.80% | 0.55% | -0.25% |
| 2016 | 2015–2025 | 0.67% | 0.77% | +0.10% |
| 2017 | 2016–2025 | 0.48% | 0.84% | +0.36% |
| 2018 | 2017–2025 | 0.80% | 1.08% | +0.28% |
| 2019 | 2018–2025 | 0.61% | 0.72% | +0.11% |
| 2020 | 2019–2025 | 0.60% | 1.05% | +0.45% |
| 2021 | 2020–2025 | 1.20% | 1.77% | +0.57% |
| 2022 | 2021–2025 | 0.87% | 1.62% | +0.74% (early) |
| 2023 | 2022–2025 | 0.02% | 1.10% | +1.08% (early) |
| 2025 | 2024–2025 | 1.08% | 2.08% | +1.00% (early) |
The editions up to 2015 were written as if the old growth of the 1990s would come back. It did not: LED lighting, efficient appliances and heavy industry moving abroad kept US use flat from 2007 to 2021. Forecasts came down only a little each year, and by the time they reached the flat line, demand had started to grow again with data centres and EVs. From the 2016 edition every forecast has been too low.
So the error is not “always too optimistic”. It is “slow to turn”. Across the 17 editions with at least five years of results, the actual growth rate differed from the forecast by between -1.5 and +0.6 points a year.
Source: EIA AEO Retrospective / EIA
Solar and wind: too low, every time
Ten years after publication, the US forecasts had solar too low in 11 of 11 editions and wind in 11 of 11. The IEA had world solar too low five years on in 11 of 11 editions, by a median 42%. Each new edition raised the line, and reality ran past it again.
Solar and wind: the fall in costs was underestimated
A study of past IEA outlooks (Lopez et al., 2025) concluded that the WEO badly underestimated the growth of solar and wind and overestimated fossil and nuclear generation. The IEA's own convention (assuming only today's policies) combined with solar costs falling faster than expected.
Source: Lopez et al. 2025
Gas and coal: the shale boom was not foreseen
US forecasts put gas-fired generation too low and coal too high every time. Shale gas made gas cheap from the late 2000s, and forecasts never kept up with how fast gas replaced old coal plants.
Source: Lopez et al. 2025 / Ember / EIA AEO Retrospective
The grid operator: PJM
PJM runs the grid for 13 eastern states and Washington DC, the region with the most data centres. Each January it publishes a 10-year forecast of the summer peak.
Show the numbers
| Report | Expected growth (% a year) |
|---|---|
| 2013 | 1.3% |
| 2017 | 0.2% |
| 2018 | 0.4% |
| 2019 | 0.3% |
| 2020 | 0.6% |
| 2021 | 0.3% |
| 2022 | 0.4% |
| 2023 | 0.8% |
| 2024 | 1.6% |
| 2025 | 3.1% |
| 2026 | 3.6% |
The 2013 report expected the summer peak to grow 1.3% a year, reaching 175 GW in 2023. The actual 2023 peak was 147 GW, lower than the 154 GW of 2012 (2023 was a mild summer, but even the June 2025 heatwave peaked at 162 GW). Then the forecasts stayed at 0.2–0.6% a year for six reports, and jumped to 3.6% within three years as data-centre requests poured in.
The first check of the new forecasts is in. At the July 2025 peak, data centres in PJM drew about 7.7 GW against the 8.2 GW expected a year earlier (95%); in Dominion’s area in Virginia, the largest, 4,970 MW against 5,563 MW. In its 2026 report PJM tightened its vetting of data-centre requests and cut the 2027 peak by 3,735 MW (2.3%), while keeping the long-term growth.
Source: PJM, Jan 2013 / PJM, Jan 2024 / PJM Inside Lines, Jul 2025 / PJM, May 2026 / PJM Inside Lines, Jan 2026 / PJM, Jan 2026
Discounting today's forecasts
These are forecasts of the future, so we do not call them wrong because today’s numbers are lower. Instead we ask two things: if they miss the way past forecasts did, what is left? And is reality so far moving at the forecast pace?
Even if PJM’s forecast misses by as much as the worst over-forecast on record (EIA’s 2005 edition, 1.5 points a year, or PJM’s own 2013 report, 1.8 points), peak demand in the eastern US would still grow about 1.8% a year, some 20% over ten years. That is not the flat decade the grid was built for.
So far: is reality keeping pace with the latest EIA outlook?
| 2025 edition expected, 2024→2025 | Actual | Versus the forecast | |
|---|---|---|---|
| Power use | +1.1% | +2.1% | above the forecast |
| Solar output | +27.9% | +34.5% | above the forecast |
| Wind output | +4.8% | +2.8% | below the forecast |
| Gas-fired output | -4.9% | -3.3% | above the forecast |
Latest: US power sales over the 12 months to Jul 2026 were +1.2% on a year earlier; the 2025 edition expects +0.6% for 2026.
What has to hold for the high forecasts
- Data-centre requests must turn into real load. PJM has started to vet them harder, and the first check (July 2025) came in at about 95% of the forecast.
- Power plants and grid connections must arrive in time. Demand cannot outgrow supply for long: if turbines, transformers and connections are short, demand is delayed rather than met. The grid →
- AI investment must continue. If it falters, data-centre growth is the part that stops first.
AI Cycle Watch: four ways the AI boom could end → Power demand: what is driving it →
Source: EIA AEO Retrospective / EIA / PJM, Jan 2026 / PJM, May 2026 / PJM Inside Lines, Jan 2026
This page organises information; it does not recommend buying or selling anything. Data belongs to each publisher and may contain errors or delays. Make your own investment decisions.
Updated 10 Nov 2026 06:30 JST · US power data to Jul 2026 · metal prices to Oct 2026