Artificially generatedThe Flat Line
What a data center really does to the German power grid.
Energy 01 measured the world at annual resolution: 19 terawatts continuous power, data centers accounting for 1.5 percent. At annual resolution, that looks harmless. Germany provides the counter-test in fifteen-minute intervals, and there the harmlessness dissolves. Not because of the quantity, but because of the form: everything in the grid now fluctuates wildly, a data center does not. We recalculated the year 2025 ourselves in 35,040 individual measurements.
1.5 percent becomes 4.6 as soon as you look closely
In Energy 01 the global figure was: data centers consumed 485 terawatt-hours of electricity, that's 1.53 percent of global electricity. The solar increase in a single year was larger than the entire data center consumption of the world.
For Germany, the same calculation looks different. Data centers here consumed around 21.3 terawatt-hours in 2025. German electricity consumption in the same year, according to our own analysis of the fifteen-minute data, was 465.8 terawatt-hours.
That makes 4.6 percent. Three times the global share. Germany is closer to computing than the global average would suggest.
Renewables Have Made the Problem Larger Than Consumption
We pulled the complete quarter-hourly data from Fraunhofer ISE for 2025 and analyzed it ourselves. Exactly 35,040 measurement points, no gaps.
Two numbers from this, and their relationship is the core of this report.
Consumption fluctuates between 33.1 and 76.0 gigawatts. A range of 42.9 gigawatts between the quietest nighttime hour and the peak. That's a lot, but it's everyday reality, and the grid has been able to handle it for decades.
Residual load fluctuates between minus 6.7 and plus 68.5 gigawatts. A range of 75.2 gigawatts.
Residual load is what must still come from conventional power plants, storage, or abroad after subtracting wind, solar, and other renewables. It is the task that the grid actually has to solve.
And it fluctuates three-quarters more than consumption itself. Renewables have not made the task smaller, they have made it more volatile. Anyone talking about data centers in the grid is talking about this curve, not about an annual total.
All 35,040 quarter-hours, sorted by size. On the left, the hour with the highest demand for non-renewable power, on the right the lowest. This representation shows not when something happens, but how often.
| Metric 2025 | Minimum | Maximum | Range | Average |
|---|---|---|---|---|
| Grid Load | 33.1 GW | 76.0 GW | 42.9 GW | 53.2 GW |
| Residual Load | −6.7 GW | 68.5 GW | 75.2 GW | 30.2 GW |
| Solar | 0 GW | 50.4 GW | 50.4 GW | 8.0 GW |
| Wind, Onshore and Offshore | – | 51.5 GW | – | 15.0 GW |
| Data Centers, Connected Capacity | constant, no range | 2.98 GW | ||
The sunniest day is the most turbulent, not the calmest
From the 365 days, we picked out the one with the highest and the one with the lowest solar yield. The result contradicts intuition, and it explains the entire problem in one image.
Solid line shows residual load, dotted line shows total consumption. Both days on the same scale to make them comparable.
| Day | Solar yield | Consumption | Residual load | Daily range of residual load |
|---|---|---|---|---|
| June 20, 2025 | 455 GWh | 38.1 to 53.1 GW | −1.5 to 44.7 GW | 46.2 GW |
| January 5, 2025 | 9 GWh | 41.5 to 60.5 GW | 10.9 to 34.2 GW | 23.3 GW |
And in the middle, the only line that never moves
Now both together. Each vertical line below is one day of the year 2025 and shows how much the residual load fluctuated on that day: from the lowest to the highest quarter-hour.
In winter, the lines are short and positioned high: high demand, little movement. In summer, they are long and reach below the zero line: the midday sun pushes the task to zero, the evening brings it back.
The horizontal band is the German data center connection capacity, 2.98 gigawatts, on the same scale. It does not move. Not in summer, not at night, not on weekends.
This is precisely where the question of this report lies. A band that never moves is either the best or the worst thing that can happen to a grid. Both depend on a single characteristic, and that comes in section 07.
Data centers are supposed to absorb the surplus green electricity. The surplus is enough for ten days.
You hear this argument everywhere: Germany is throwing away green electricity, data centers could take it. We calculated how much is actually available.
The residual load was negative in 121 hours in 2025. The energy below this zero line, i.e., the real nationwide surplus, adds up to 254 gigawatt-hours.
That is 0.055 percent of annual consumption. A data center with one gigawatt of continuous capacity would be supplied by this for ten and a half days. After that, the entire year's supply is exhausted.
Measured against the actual demand of German data centers of 21,300 gigawatt-hours, this surplus covers 1.2 percent. As a business model, that does not work.
All in gigawatt-hours 2025, same scale. The bar length is the true value, not compressed.
| Quantity | GWh | What it means | Source |
|---|---|---|---|
| Surplus | 254 | Renewables delivered more than all of Germany consumed. Physically surplus. | Own calculation from 35,040 quarter-hours |
| PV curtailment | 2,704 | Electricity was available and needed, but could not be transported. A location problem. | Bundesnetzagentur, Redispatch 2025 |
| Data centers | 21,300 | Actual annual demand. 84 times the surplus. | Bitkom and Borderstep 2025 |
577 hours in which electricity cost money when you generated it
If a thesis about the power grid is correct, you must see it in the price. We therefore also pulled the exchange prices for 2025 and evaluated them using the same method.
Result: In 577 hours the day-ahead price was negative. Anyone who fed in during these hours had to pay for it. That's 6.6 percent of the year.
The published number is 573 hours. Our deviation of 0.7 percent comes from the fact that Germany switched from hourly to quarter-hourly contracts on October 1, 2025, and the data points therefore need to be weighted differently. We cite both numbers instead of smoothing one of them.
More interesting than the average is the range. The lowest price of the year was minus 250 euros per megawatt hour, the highest was plus 583. Between the cheapest and most expensive moment of the year there are 834 euros.
For a data center that draws the same power around the clock, this range is purely a cost question: it inevitably buys the average. For one that could shift its load, the same range is the business model.
The difference between the two is not a question of technology in the server room, but of the type of work that runs there.
Day-ahead, bidding zone Germany and Luxembourg. Daily average smoothed, the extremes of individual quarter hours are far above and below.
| Key figure 2025 | Value | Source |
|---|---|---|
| Average price | 90.47 €/MWh | own calculation |
| Lowest value | −250.32 €/MWh | own calculation |
| Highest value | 583.40 €/MWh | own calculation |
| Range over the year | 833.72 €/MWh | own calculation |
| Hours with negative price | 577 h | own calculation, published 573 |
| Year-on-year comparison | 457 h (2024) | Market reports |
Training can wait. An answer cannot.
Whether a data center helps or harms the grid depends on a single question: May the work take place later?
Training a model is batch work. It runs for weeks, nobody is waiting for an intermediate result, and it doesn't matter whether a particular hour is calculated tonight or tomorrow morning. Such load can be shifted, and shiftable load is more valuable than any power plant for a grid with fluctuating generation.
Answering a query is the opposite. A person is waiting. Delay here is not a technical detail, but the difference between a usable and an unusable product.
And it is precisely this second part that is growing. As long as AI was mainly being trained, a large part of the load was shiftable. The more AI is actually used, the smaller the shiftable portion becomes. The expansion therefore not only makes the load larger, but also more rigid.
Four sentences to get the picture right
First: 4.6 percent is not an emergency. German industry consumes many times that amount. Anyone portraying data centers as the main cause of grid problems hasn't looked at the scale.
Second: the form is still the issue. A constant load of 3 gigawatts in a system whose task varies between minus 7 and plus 68 gigawatts is something different from 3 gigawatts of industry with shift work and factory holidays.
Third: our numbers are Germany 2025. A grid with a different generation mix has different curves. France with nuclear power, Norway with hydro, or Texas with wind look completely different. The method is transferable, not the result.
Fourth: curtailment is not an argument for data centers. It is an argument for transmission lines. A data center at the right grid point can indeed absorb curtailed energy. At the wrong grid point, and 37 percent of German capacity is connected to a single one, it exacerbates precisely the bottlenecks that lead to curtailment.
If you're buying computing capacity
Don't ask the provider about the green electricity share on paper, but about the grid point and whether they can shift load. The first is accounting, the second is physics.
If you're computing yourself
Scheduling training and batch runs at night or midday costs nothing and hits exactly the hours with the lowest prices. With a spread of 834 euros over the year, this is not symbolic.
If you're talking about the numbers
Separate surplus from curtailment. The surplus was 254 gigawatt-hours, curtailment was eleven times that. Anyone lumping both together inevitably arrives at the wrong solution.
The number that counts
Not annual consumption, but connected load. 21.3 terawatt-hours sounds like quantity, 2.98 gigawatts describes what the grid must provide every second.
Everything this report is based on
Seven sources and one proprietary analysis. The raw data is public, anyone can verify the calculation. That's exactly why it states below which query delivers which time period.
- Fraunhofer ISE, Energy-Charts: public interfaceThe foundation of Fig. 01 to 03 and most numbers in this report. We pulled calendar year 2025 in four quarterly queries, a total of 35,040 quarter-hours without gaps, and summed load, residual load, solar and wind ourselves. The interface is free, without key and explicitly intended for such analyses.api.energy-charts.info/public_power?country=de&start=2025-01-01&end=2025-04-01
- Fraunhofer ISE, Energy-Charts: exchange pricesFoundation of Fig. 05 and the 577 negative hours. Bidding zone DE-LU, entire year 2025. Contains the switch from hourly to quarter-hourly contracts on October 1, 2025, which must be considered when counting.api.energy-charts.info/price?bzn=DE-LU&start=2025-01-01&end=2026-01-01
- Bitkom and Borderstep Institut: Data Centers in Germany, Update 202521.3 terawatt-hours consumption in 2025 versus 20.0 the previous year, 2,980 megawatts connected load with plus 250 versus 2024, the expectation of 5 gigawatts by 2030, and the AI share from 530 to 2,020 megawatts.bitkom.org/sites/main/files/2025-11/bitkom-studie-rechenzentren-in-deutschland-2025.pdf
- Borderstep Institut: Market Developments, November 2025The regional distribution on which section 08 is based: well over 1,100 megawatts in the greater Frankfurt area, over 37 percent of national capacity at one point, followed by Bavaria with 420 and North Rhine-Westphalia with 378.borderstep.de/2025/11/12/rechenzentren-in-deutschland-aktuelle-marktentwicklungen-update-2025/
- Bundesnetzagentur: Grid Congestion Management and Redispatch 2025Around 30 terawatt-hours total volume, of which 2,704 gigawatt-hours curtailed photovoltaics with a doubling compared to 2024. Plus the shift of bottlenecks to the distribution grid, which explains why curtailment is a location problem, not a quantity problem.bundesnetzagentur.de/DE/Fachthemen/ElektrizitaetundGas/Versorgungssicherheit/Netzengpassmanagement/Engpassmanagement/Redispatch/start.html
- Agarwal et al.: To Defer or To Shift? Flexibility of AI Data CentersThe numbers from section 07: 3 to 21 percent lower grid costs with shiftable AI load, saturation of the effect at 20 to 30 percent shiftability, and the finding that the shiftable share shrinks with growing utilization.arxiv.org/abs/2604.05376
- MCG Research: Energy 01, The Energy Balance of HumanityThe framework that this report fills. 19 terawatts continuous power of humanity, 86.3 percent of it fossil, data centers at 1.53 percent of global electricity. This also contains the global sources: Energy Institute, Ember, IEA and the Potsdam Institute.mcgrinsey.com/magazin/19-terawatt-energiebilanz-der-menschheit/
Proprietary analysis McGrinsey: All values for load, residual load, solar, wind, surplus, daily ranges and prices are calculated from the raw data of Energy-Charts ourselves, not adopted. Energy sums from power times 0.25 hours per measurement point. The surplus is the area below the zero line of residual load. The price hours are weighted by contract length, because Germany switched to quarter-hours on October 1, 2025. Where our calculation differs from a published number, both are stated in the text. Residual load is a balance sheet quantity for all of Germany and says nothing about whether the electricity actually arrives where it is needed. That is precisely the difference between section 05 and curtailment.
Energy 01 ended with the promise to make exactly this narrowing: world numbers as a framework, a grid area in fifteen-minute intervals as proof, and the question of what a data center actually does within it. The answer is: as quantity little, as form much, and as location decision very much.
Energy 03 takes on the opposite direction. If shiftable load is so valuable, why is it hardly traded? What does a grid operator actually pay today for an interruptible load, and why is that not enough?
Related: Data 02 on the question of what happens when a measuring instrument can no longer see the world it is supposed to measure.
