AI and Energy - the Power and the Glory

Looking for clarity on AI, data centres and the future of energy? This week's episode of Cleaning Up is for you. Your guide will be Mike Headroom, an AI avatar based on my voice. See what you think!

This week’s episode of Cleaning Up is all about AI, data centres and the future of energy. It’s my voice, but that’s not me speaking, it’s Mike Headroom, an AI cloned version of me. Image: Liebreich Associates / All3Media

Remember Max Headroom? Well, there's a new star in town: Mike Headroom. And he has a lot to say about AI, data centres and the future of energy.

This week’s episode of Cleaning Up is an audio adaptation of my recent piece for BloombergNEF about AI and energy, titled The Power and the Glory. It was a deep dive, nearly 7,000 words, starting with the history of AI and the many false alarms about how data centres would eat all the electricity in the world. I followed this with a structured exploration of what AI uptake might look like, what power demand it might drive, and how that demand might be met. And, as my long-time readers / listeners might expect, I finished with a few cautionary words.

I don’t want to spoil the fun, but here are a few of the main conclusions:

• AI is real and transformative, though I don’t expect humans to be replaced any time soon.

• AI data centres - particularly the ones needed to train frontier models - are unthinkably big, complex and costly. Think $25 billion per GW.

• The speed with which AI is rolled out will be limited by the rate of uptake of paid services, the data centre supply chain and skills, the availability of power, and the willingness of the US to export GPUs.

• Demand for power for AI will also be limited by improvements in energy efficiency - in chips, GPUs, algorithms and applications. Already since publication, Chinese AI DeepSeek appears to have outperformed major models by orders of magnitude on cost and energy efficiency.

• Estimates for the increase in data centre power demand range from a paltry 35% global increase in between now and 2030, to a storming 250% growth in the US alone by 2028.

• On a global basis, my own estimates of power demand are at the modest end (45GW of additional dispatchable power by 2030). However, for the U.S. they are more aggressive: 30 GW of additional dispatchable power by 2030, accounting for around 9% of total power use.

• In their rush to power their AI ambitions, the tech titans have latched on to nuclear, particularly Small Modular Reactors. They will find building any new reactor design frustratingly slow and stubbornly expensive.

• The hyperscalers will build a lot of gas capacity because it’s fast and simple, but will eventually discover what all the utilities have discovered - the cheapest way to meet demand, in a way that is acceptable to local communities and regulators, will be via a mix of renewables, batteries and gas backup, plus a lot of investment in the grid.

• There will almost certainly be a market correction along the way - as has happened with every new platform technology: electricity, telephony, aviation, railways, internet, quantum computing, driverless cars, electric planes, hydrogen (repeatedly), etc, etc.

The creation of Mike Headroom

Writing the piece for BloombergNEF and recording the episode has been a huge learning experience for me.

I didn’t want to be one of those commentators who bloviates about AI without using the stuff, other than when it is invisibly embedded into other services. So, over the past year I started to use Perplexity (instead of a lot of Google searches), Elicit (to find and quickly assess academic papers), Replit (to make databases instead of excel), and Elevenlabs (to create Mike Headroom and record the piece).

In each case, there was an initial learning curve as I got to know the tool; a wow moment as I realised its capability; disillusionment, as I became aware of its limitations and tendency to make stuff up; and finally accommodation, as I learned how to use the tool and on what, and stopped wasting my time on stuff it could not yet deliver.

Recording the piece was fascinating, but also frustrating and time-consuming. To clone my voice, I uploaded around three hours of previous Cleaning Up audio-blog episodes. In order to stop others from doing the same, once it has created the voice (a few hours of processing followed by an email), Elevenlabs makes you read a random sentence to prove you are the owner of the voice. It felt like a graduation viva.

Once that was done, I created a “project” and uploaded my text. There are two problems: the AI has no real understanding of what it is reading, so its intonation and emphasis is all over the place; and it is variable, so each time you tweak a paragraph and regenerate it, it comes out differently - often with new problems. So there is a lengthy process of inserting punctuation, capitalising words, manually adding pauses, even rewriting phrases, and then regenerating multiple times, until you are happy with the results.

Even then, however, you are not done. Each paragraph could be quite different in terms of timbre, loudness, pace, tone, and so on. Oscar (our brilliant producer) went through the entire episode and wrought his magic, to make it bearable to listen to. If I were to do this again, I would probably start again with a new set of training audio, making sure it was much less variable. Or I could just wait another couple of years for the AI to improve!

Listen now to “Generative AI, the Power and Glory” on Cleaning Up.

And remember, it may be my voice, but it’s not me reading it!

So, there you have it…

I hope you enjoy the episode. Please do share it with friends, family and colleagues - AI is, quite simply, one of the most significant developments of our time, and it is vital that we get a clean signal on where it is headed.

If you are enjoying listening to/watching Cleaning Up, please recommend it to your friends, family and colleagues. Our audience is growing, and you can help! And don’t forget to subscribe to Cleaning Up’s own newsletter: CleaningUpPod.Substack.com.