How much power do hyperscalers use, and how much would it cost to go fossil-free 24 / 7 ?

I found out today that the Carbon Disclosure Project allows you to register and download their questionnaire responses. This means for companies that conveniently leave absolute energy usage figures out of their annual reporting, it’s possible to get an idea of what they might be – something I care about in the context of figuring out the assumed energy needed for a fossil free internet.

Developing an intuition for absolute energy usage figures by the dominant hyperscaler providers

I don’t want to breach the terms of service of the CDP, so I’m not publishing these verbatim – but the figures they have collated are really helpful, as it’s now finally possible for me to get an idea of the relative amounts of electricity used by the big hyperscalers.

When I use any numbers for energy below, in each case I’m rounding to the nearest Terawatt hour (TWh). This is thousand gigawatt-hours, a million megawatt hours, or a billion kilowatt hours. This is an absolutely massive unit, and it’s typically used to describe annual energy usage of countries or cities (Ireland’s energy usage was about 34 TWh in 2022, for example).

For future reference, I’m also using the answer to 8.2 – Consumption of purchased or acquired electricity in the CDP questionnaire, as the source of numbers here.

Amazon’s absolute electricity usage figures comes to about 39 TWh, and around 90% of this is listed as coming from renewable sources. I didn’t see it listed in their most recent Sustainability report detailing figures for 2022.

Google helpfully publish their own CDP response for the same year, on their own website. The same figures are about 22 TWh, and 100% listed as coming from renewable sources for the financial year of 2022.

Microsoft also publish their CDP response. Their figures are about 18TWh and 100% from renewable sources for the same year.

These figures for renewable energy are all on annual basis, not hourly.

How do they compare, and what does this tell us?

Google and Microsoft are pretty close in usage – Microsoft is maybe 20% lower than Google. This gives us an idea of how much Microsoft has shifted from being a company that sells licenses for people to run on their own computers, to becoming a cloud company that bundles the software with the use of its own infrastructure now instead.

Interestingly, Amazon’s total electricity consumption is almost the same as the Google and Microsoft combined. However it’s worth bearing in mind that Amazon is not the same as AWS – there is a massive set of fulfillment facilities to think about too, that is used by the retail part of Amazon, so we can’t attribute all of those 39 TWh of it to AWS data center facililties.

With this in mind, it’s not at all surprising that Amazon is now the largest corporate buyer of renewable energy in the world given its size – to hit 90% renewable on an annual basis, they’ve needed to buy 35 TWh of ‘clean’ energy. For context, this is more than all of the electricity that Ireland used the same year.

So, if you wanted 24/7 fossil-free energy powering all these cloud providers, you might sum these annual figures to end up at 80TWh of energy that you would need to have matched on an hourly basis from fossil free sources.

As mentioned before, we don’t have a meaningful split between Amazon and AWS, but all these providers have grown since 2022, so 80TWh, in the absense of other data doesn’t feel too outrageous a number to shoot for.

How much would this cost?

Once again, large tech firms don’t disclose how much they pay for energy. (Somewhat educated) guesswork time again.

Let’s assume a pretty generous figure of 100 USD per megawatt hour (MWh) to procure round the clock 24 / 7 renewable energy. Prices can range from 30 USD to well over 100 depending where you are, and what time of day you’re buying, but as an average, this shouldn’t too wacky a figure.

If there are 1,000,000 megawatt hours in a terawatt hour, then the maths is fairly simple – one hundred times one million means you’re looking at a hundred million USD per terawatt hour.

If all those zeros are a pain to think about , you can also say about 0.1 billion USD per Terawatt hour, as a very general rule of thumb.

Let’s budget this per hyperscaler, using our educated guesses, and look at their revenue reported as public companies.

FirmEnergy use (TWh)Energy cost (billions, USD)Revenue (billions, USD)Percentage of revenue
Amazon393.9554 (source)0.7
Google222.2280 (source)0.8
Microsoft191.9212 (source)0.9

What does this tell us?

A key takeway for me here is that energy costs as a percentage of revenue are not large at all for hyperscalers. I could imagine them easily doubling the amount spent per megawatt hour to green their supply, matching every hour of use much more quickly, without endangering their business in the slightest.

Seeing it in a table here makes it all the more stark.

I haven’t tried comparing what these firms spent on energy versus what they spent on shareholder buybacks in the same year, but if you’re curious here’s a head-start – comparing what is spent on share buybacks vs the cost of decarbonising the grid.

Update: I forgot I already had these figures in that post for the companies in question, so I’ve added them below. Seeing what is being spent by these firms on energy, which is absolutely core to their business and the functioning of these services vs just buying their own shares to make number go up puts it into perspective for me – we could be moving SO much faster than we are now.

FirmAssumed Energy cost in 2022 (billions, USD)Money spent on buybacks in 2022 (billions, USD)Multiple spent on buybacks vs spent on energy

Wait, 90% and 100% sounds pretty good – why care about 24 / 7?

There isn’t really the space in this post to go into the details, but the short version is that saying you run on green energy on an annual basis, while better than nothing, doesn’t really represent how energy systems really work – we often assume it to mean all the power is coming from green sources all the time, and that’s not the reality at all. For us to meaningfully shift, the clean energy we use needs to be timely, deliverable and additional.

I’ve now carried out literally hours of podcast interviews with experts in this fields to help folks understand some of the finer points of using green energy in tech. Here’s one with Savannah Goodman at Google, here’s another one with Dr Iegor Reipin at TU Berlin diving into the modelling of it using open source energy modelling software, and here’s one with Nina Jablonska of Energy Tag, where we unpack timely, deliverable and additional as desirable attributes of clean enegry in more detail.

It’s also acheivable on a faster timescale than the speed these firms are moving at.

Here’s an podcast interview in Volts with the CEO of Peninsula Energy who are on track to achieve 100% 24/7 by 2025, and here’s a post about Good Energy, a UK firm offering 90% hourly matching already. If we’re not moving faster, the figures above are pretty stark – it’s because these firms are prioritising other things like share buybacks over energy transition.

Interested in more?

I run a non-profit dedicated to reaching a fossil free internet by 2030, and I hope this post helps explain that we can afford to move much faster than we are right now when it comes to greening digital services.

I also did a talk along these lines as FOSDEM last year in the energy dev room talking about a fossil free internet, and some of the most interesting projects I know. If you enjoyed this post, you might enjoy that talk.

If you’d like to talk about this or work together on anything related to this, do please get in touch.