Originally posted at The Orthosphere on 12th November, 2026

Six companies dominate the construction and use of data centres: Amazon, Microsoft, Meta, Google, Oracle and Apple. Data centres have been around for a while, supporting search and cloud services, not to mention global scale surveillance. An article in Nature1 speculates that AI data centres could account for 1% of global electricity demand by 2030. The article attempts to determine the share of demand of each of those companies, and projects demand to 2030 based on three scenarios; conservative, neutral and optimistic.
Figure 3 aggregates the historical and projected values for all six companies. ChatGPT was released in November of 2022. At that time, the aggregate consumption of the six was 100 TWh. This I take to be the annual energy consumption. The figure shows projections for 2030 of between 240 and 290 TWh for these companies.
The International Energy Agency (IEA) produced an updated report2 3 earlier in 2026 about the impact of AI on energy demand. Their figures are for the the whole data centre sector, rather than being narrowly focussed on the major players.
Our updated projections see electricity consumption from data centres roughly doubling from 485 TWh in 2025 to 950 TWh in 2030, accounting for around 3% of global electricity demand by that date. Electricity consumption from AI-focused data centres grows much faster than overall data centre electricity consumption, tripling in this period. … In the Base Case, the electricity consumption of these AI-focused data centres increases by more than threefold to 2030, reaching around 465 TWh by 2030.
Grok 4 is trained at the Colossus 2 data centre. I asked Claude about the energy requirements for this one set of models (and Grok 4 is a collection of models.) It gave me references from 2025, for the training of the the initial Grok 4 model.4 This is just one of the companies and models competing for consumers.
The best available estimate is about 310 GWh (0.31 TWh) of electricity, from Epoch AI, an independent research group.
How Epoch gets there
It estimates Grok 4 used 246 million H100-hours, extrapolated from what is known about Grok 3 and assumptions about how far training was scaled up.
It assumes H100 SXM GPUs with a maximum draw of 700 W, and that average power during training is typically around three-quarters of peak.
GPUs alone would then account for about 130 GWh (246M h × 0.7 kW × 0.75). Epoch’s 310 GWh implies a multiplier of roughly 2.4 for servers, networking, cooling and other facility overhead. That multiplier is my back-calculation, not a figure Epoch states.
Related estimates from the same analysis
Cost is about $490 million, mostly hardware.
Water use is about 750 million litres. Emissions are about 150,000 tons of CO2-equivalent, because Colossus ran mainly on natural gas turbines. A figure of roughly 73,000 tons has also been cited from Stanford’s 2026 AI Index. The two use different methods, so they disagree by about a factor of two.
What about us? What are the energy requirement for our “training”?
For adult men with light to moderate activity, the average recommended energy allowance is approximately 2,900 kcal/day (at a reference weight of 79 kg), while for women it is about 2,200 kcal/day (at a reference weight of 63 kg). These estimates use an activity factor of about 1.6-1.7 times REE. After age 50, energy needs decline further due to reduced lean body mass and lower activity levels.
https://agriculture.institute/food-fundamentals-cpo/understanding-human-energy-requirements/#adults-19-50-years-and-beyond
What are the equivalent watt-hours for these energy allowances? One kilocalorie is equivalent to 1.162222 watt-hours.5 I’ll focus on the men, for reasons which will become obvious, and trim the average a bit. 2,700 kilocalories/day is a shade over 3,000 watt-hours per day. For a year, then, the figure is about 1,100,000 watt-hours, 1,100 kilowatt-hours (KWh), or 1.1 megawatt-hours (MWh)
Energy requirements in adolescence are higher, in old age lower. Let’s arbitrarily take an active learning lifetime from, for the sake of round numbers, 14 to 44, i.e. 30 years. Given the average yearly energy requirement of 1.1 MWh per year, and in spite of the fact that learning happens all of the time, the period of intensive active learning is not more than 6 hours per day, the active learning energy consumption per person is 275 KWh per year. (But see below, in the discussion of St Thomas Aquinas.)
Put 1,000 such people together for a year and they will consume 275 MWh of intensive learning energy. Over the 30-year period of active learning for these same people, they will consume 8.25 gigawatt-hours (GWh) of energy. 10 such groups of people will consume 82.5 GWh over their total active learning lifetimes.
I’m throwing some numbers at the wall here to paint a picture in spaghetti of the energy requirements of a cohort moving through the various faculties of the mediaeval universities: Bologna, Oxford, Salamanca, Sorbonne, Cambridge, Padua, Coimbra and later Prague. A new cohort of students, at least, arrives every year, so let’s look, simple-mindedly, at 30 years of such cohorts and arrive at a total value, for 30 cohorts spread over 60 years, of about 2,500 GWh or 2.5 terawatt-hours (TWh).
The great flowering of philosophy and theology occurred primarily during the High Middle Ages over a period of about 80 years. All of the figures given here (except for the basics of human energy consumption) are wildly speculative and no doubt hugely imprecise, but for the sake of argument, let’s say that such an budget is indicative of the energy levels that powered the university populations that sustained this golden age. A similar reckoning could be made, over a longer period for the Scientific Revolution.

The innovators were a tiny minority; the universities then trained agents who carried the innovations out from the universities. But one innovator in particular stands out: St Thomas Aquinas. Thought to have been born in 1225, he died in 1274. His early education began at about five ears of age, and at about fourteen years, be commenced at the studium generale in Naples. Although he, famously, ceased to write at the end of 1273 following another mystical experience, let’s say that his period of intensive learning covered a period of 35 years, and set the average caloric requirement of the dumb ox of Sicily at around 3,500 kilocalories/day. If we assume that his active learning occupied twelve hours of his day, and make an allowance for sleeping, his intense mental activity would require at least 2,000 kilocalories per day. However, human intellectual activity is not something that can be considered in isolation from the whole person, and the whole person needs sleep and worship and recreation, so I’ll keep to the first figure. (The same adjustment should then be made for the cohort numbers above; in that case multiply by 4.) Roughly speaking then, he would require 4 KWh/day, around 1.5 MWh/year and about 52 MWh over his active learning lifetime.
St Thomas Aquinas. 52 MWh.
Data centres are increasingly controversial because of their voracious appetites for electricity and water, not to mention their construction footprint. The generalised lunacy about CO₂ also demands that, in countries suffering from the sickness, vast areas be cleared for solar panels and wind farms to absolve guilty climate consciences.
I’ve been struck by the scale of the inputs required to generate imitations of human conversation, memory, reasoning, and the production of certain artefacts that flows from these. Yet all of these things, and much, much more, flow naturally from us underpowered human beings.
- https://www.nature.com/articles/s44458-026-00152-5 ↩︎
- https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary ↩︎
- https://www.iea.org/files/3179f7f8-01f6-4dd6-bffa-c9f7b73f1dc9/KeyQuestionsonEnergyandAI.pdf ↩︎
- https://epoch.ai/data-insights/grok-4-training-resources ↩︎
- https://www.xconvert.com/unit-converter/kilocalories-to-watt-hours ↩︎
