Artificially generated19 Terawatts
That's how much power humanity draws continuously. Every second, all year long. This report breaks down the number: where the energy comes from, how much of it is electricity, what goes into data centers, and what the sun delivers for free at the same moment.
Energy consumption is almost always reported as an annual quantity, and annual quantities mean nothing to anyone. 600 exajoulesExajoule is the unit in which the world accounts for its annual energy. An exajoule is a quintillion joules. More practical: 1 EJ = 277.78 terawatt-hours, and all of Germany consumes around 12 EJ per year. is not a number you can imagine.
Divide it by the seconds in a year, and it becomes tangible: 19 terawatts of continuous power. Humanity is a machine that permanently runs at 19 trillion watts, or 19,000 billion watts. That's the real metric, and everything else in this report is a fraction of it.
Six sources, and three of them make up 86 percent
600.45 exajoules total. The width of each section is its share.
Numbers as table
| Source | Exajoules | TWh | Share |
|---|---|---|---|
| Oil | 201.0 | 55,833 | 33.5 % |
| Coal | 166.0 | 46,111 | 27.6 % |
| Gas | 151.0 | 41,944 | 25.1 % |
| Nuclear | 31.0 | 8,611 | 5.2 % |
| Hydro | 16.0 | 4,444 | 2.7 % |
| Renewables | 35.5 | 9,847 | 5.9 % |
| Total | 600.5 | 166,792 | 100 % |
The interesting number isn't in the chart, but in the growth. Renewables increased by 3.3 exajoules in 2025, up ten percent, and solar alone provided 71 percent of that increase. For the first time outside a recession, renewables were the largest growth source in total energy supply.
And yet: the fossil block barely moves. 86.3 percent. Both are simultaneously true, and anyone who quotes only one of these two numbers is telling a story instead of a finding.
The energy transition everyone talks about is happening in 19 percent of the system
When energy is discussed, it's almost always about electricity. But electricity is only 19.0 percent of global primary energy: around 31,700 TWh out of 166,792 TWh.
The other 81 percent is combustion that never goes through a wire: transportation, industrial process heat, heating, cement and steel furnaces. There the transition is harder, more expensive and slower, and that's where the real work of the next twenty years happens.
Where electricity stands in relation to primary energy.
Four quantities from the same year, on the same axis. All values in terawatt-hoursOne terawatt-hour is one billion kilowatt-hours. A German household consumes around 3,000 kWh per year, so one TWh supplies about 330,000 households. Not to be confused with terawatt: that's power, the watt-hour is quantity..
Numbers as table
| Quantity | TWh | Share of generation |
|---|---|---|
| Total renewables | 10,730 | 33.8 % |
| Coal power | 10,476 | 33.0 % |
| Total solar | 2,778 | 8.8 % |
| Solar growth 2025 | +636 | +30 % from previous year |
| Data centers worldwide | 485 | 1.53 % |
The sun delivers nine thousand times as much
Around 173,000 terawattsTerawatt is one trillion watts, or 1,000 billion watts. A good light bulb needs 10 watts. 1 TW = 1,000,000,000,000 W of solar power constantly hit the Earth. Humanity draws 19.
The image above is rounded because the exact number doesn't help anyone: the ratio is 1 to just over 9,000. For those who need it precisely: 19.0 to 173,000 terawatts, so 1 to 9,092, or 0.011 percent.
Put another way, and this is the version to remember: what humanity converts in energy over an entire year, the sun delivers in just under an hour.
This number is not an argument for any particular technology. It's a perspective, and one that turns the usual debate on its head.
Humanity's energy problem is not a quantity problem. It's a problem of harvesting, storage, transport and timing. The raw material is there in abundance and arrives every second, whether we take it or not.
Large numbers are the point where understanding breaks down
Up to about a hundred, everyone can imagine something. After that, numbers become words, and words can be repeated without understanding them.
Particularly tricky is the jump between languages: the English billion is the German Milliarde, not Billion. Anyone who confuses this is off by a factor of a thousand. That's why in this report, every really large number has the translation next to it.
The ladder below is the entire order of magnitude in which this report operates. The highlighted step is the one in which humanity as a whole calculates.
Each step is times a thousand. That sounds harmless and is the reason why three steps further on, imagination fails: from million to trillion is the same distance as from one to million.
Data centers are 1.5 percent of global electricity. And that's the wrong question.
Actual values 2024 and 2025, baseline forecast 2030, three scenarios for 2035. The fan from 2030 is not measurement uncertainty, but an open decision.
Numbers as table
| Year and case | TWh | Share of global electricity |
|---|---|---|
| 2024 Actual | 415 | around 1.5 % |
| 2025 Actual | 485 | 1.53 % |
| 2030 Base case | 945 | nearly 3 % |
| 2035 Headwinds | 700 | under 2 % |
| 2035 High efficiency | 970 | 2.6 % |
| 2035 Lift-Off | 1,700 | 4.4 % |
636 versus 485 terawatt-hours. Solar added more generation in twelve months in 2025 than all data centers in the world consumed in the same year. That's the most honest available benchmark for the question of whether AI will break the power grid.
The answer at the global level is: no. Data centers represent less than ten percent of global electricity demand growth through 2030. Air conditioning, electric vehicles, and general electrification represent far more.
The answer at the local level is: yes, very much so. Around 80 percent of total growth through 2030 occurs in the USA and China. When a single county in Virginia or a province in Guizhou gets the load of an industrialized nation added to it, the reassuring global average doesn't help anyone there. The problem is not the quantity. The problem is the distribution and connection.
The calculation only works if the tool unlocks more energy than it consumes
The consumption side of AI is well measured. The yield side is not, and that's exactly where it's decided whether AI is a cost item or a lever in the energy system.
What is provable today differs sharply from what is announced. Both must be kept separate.
Proven
Plasma control in tokamaks. DeepMind and EPFL have shown that deep reinforcement learning can control the magnetic coils of a fusion reactor so that complex plasma shapes remain stable. Since October 2025, collaboration with Commonwealth Fusion Systems has been running to apply the same method to a reactor intended to deliver grid power.
Materials discovery. Machine learning has expanded the search space for new crystal structures by orders of magnitude. For batteries, catalysts, and solar cells, this is the bottleneck, not manufacturing.
Forecasting. Weather and load forecasting is the most unspectacular and effective application. Wind and solar power whose feed-in is known more accurately hours in advance is simply worth more in trading and requires less reserve capacity in the background.
Announced
Grid-native optimization by the second, autonomous operation of power plants, AI-designed reactor designs. Serious programs, but without reliable yield numbers. We list them here as open items, not as proof.
Steel, cement, chemicals: the physics behind the exajoules
Energy is an abstraction. Below it lies a material level, and that's the reason why the 81 percent non-electricity is so persistent.
The Potsdam Institute for Climate Impact Research models precisely this level in REMIND, its Integrated Assessment Model. Since version 3.1, four industrial sectors are individually represented: cement, chemicals, steel, and other industry. For steel, there's a process-based model inside that tracks individual production routes with their costs, energy and material inputs, instead of treating steel as a single number.
For an energy report, this is the crucial resolution: not how many exajoules industry draws, but which physical process draws them and whether it can be electrified.
Three sectors that together account for nearly half of industrial energy use.
Cement is the hard case. Around 29 percent of direct industrial emissions come from it, and a large part of that arises not from heating, but chemically when burning limestone. You can't eliminate this portion even with one hundred percent renewable electricity. It requires a different process or capture.
Quantity is not the problem
At 1 to 9,092 compared to solar irradiation, available energy is not a hard limit. Harvesting, storage, grid and timing are.
AI is energetically small and locally large
0.29 percent of primary energy globally, but 80 percent of growth in two countries. Anyone talking about AI and energy must specify at which level.
The benefit side is missing from the debate
What AI costs in energy is measured. What it unlocks in energy is claimed. Closing this gap is the most interesting open work.
Electricity is one fifth
The 81 percent combustion outside the electricity grid receives a fraction of the attention it deserves.
- Energy Institute, Statistical Review of World Energy 2026, 75th edition, June 2026. Primary energy 2025 by source. energyinst.org/statistical-review
- Ember, Global Electricity Review 2026, April 2026. Global electricity generation 2025, renewable and coal share, solar growth. ember-energy.org
- IEA, Energy and AI. Electricity consumption of data centers, base case 2030 and scenarios 2035. iea.org/reports/energy-and-ai
- Potsdam Institute for Climate Impact Research, REMIND 3.1.0, Geoscientific Model Development 17, 2024. Industrial sectors, final energy shares, electrification pathways. gmd.copernicus.org
- Google Cloud, measuring the energy demand of a Gemini query, August 2025. cloud.google.com
- Fraunhofer ISE, Energy-Charts. Reference for high-resolution grid data. energy-charts.info
McGrinsey conversions: 1 EJ = 277.78 TWh. Continuous power over 31,557,600 seconds. Per capita values based on 8.23 billion people. Percentage values rounded, totals may deviate by 0.1.
