Part one of our Keep Calm and Count the Kilowatts series showed how AI prompts are only a small portion of a person’s daily power use. But with OpenAI alone now handling over 2.5 billion prompts a day from hundreds of millions of active users, how does the power use look on a global scale?
The big picture
Looking at the total context is where doom and gloom articles often miss explaining how AI fits into global resource use. So for an easy comparison, I’ve collected together the annual power use for a range of relevant and interesting uses and categories.
It’s pretty clear that AI is a tiny fraction of overall electricity use on a personal or global level and it’s beaten by tech as unassuming as Wi-Fi. That doesn’t mean it’s not still a huge amount of power – if Wi-Fi was a country, its electricity use would rank it in the top 50.
AI is also going through a period of very rapid growth, which is normal for transformative new technology. Take televisions for example – by the 80s, they were consuming similar amounts of electricity to AI today.
In fact, the electricity consumed by the average TV-watching family in the 80s was comparable, on a per-person basis, to the power needed for all the world’s data centers today.
AI may be a new rising star, but it’s still only 10% to 20% of the 415TWh a year of power used by data centers overall. Over 50% is for enterprise and government use (think corporate cloud storage, online banking, and digital government services), while streaming video takes about 15% and the 10 trillion+ smartphone pics in the cloud (3 trillion of those are my dog in slightly different poses) is about 0.2%.
The overall growth is not slowing down either and data center power use is expected to double by the end of the decade, with a significant chunk of demand driven by AI. It’s hard to know what the future will hold, but even if AI is the majority of data center electricity use, it’s still only a small part of worldwide use.
Of course, AI not the only tech going through massive growth right now and power hungry 5G cellular may soon catch AI. Electric vehicles are one of the fastest growing new users of electricity and will easily top out at 100x as much power compared to what AI uses today.
What about water and CO2 emissions?
Depending on which end of the scale you look at it from, broader AI resource use is both tiny, yet huge.
The full flush once you are done reading this article could provide cooling for 10 prompts a day for almost 5 years. One 500 ml disposable bottle of water is cooling for 2,000 prompts, while the 100GL of water used for cooling AI in datacenters each year is about the same as what’s spent on watering golf courses while it’s raining.
We shouldn’t be wasting water, but data centers (and AI) are pretty far down the list of usage we should be cutting back on.
Then there’s the carbon footprint. 0.03 grams of carbon dioxide equivalent per prompt, which is about the same as you exhale each breath. 10 prompts a day are equal to the CO2 emissions from the candle at a one-year-old’s birthday party or idling your car for less than a second.
Or, for a more universally understood unit of measurement, I calculated 100 million prompts per Zamboni year, with a quarter of that from Canada.
All up, the yearly emissions equivalent from AI power use is around 0.07% of the global total. Looking at the big picture, that’s a tiny fraction – about the same as the ridesharing industry.
But 0.07% of a very very large number is still a lot of CO2 and puts AI on par with the entire country of Denmark. That’s not great.
Disproportionate effects
The real environmental story of AI isn’t the tiny sip of power your individual prompt uses; it’s the massive, concentrated impact of new data centers on the specific towns and ecosystems they’re built in.
But more on that in part three of Keep calm and count the kilowatts series.
Not convinced Zamboni year is a valid measurement of CO2 output? Drop the mitts and let’s dance in the comments!
AI skeptics might also like
AI enthusiasts might also like
How we use AI
Here at TechRadar, our coverage is author-driven. AI helps with searching sources, research, fact-checking, plus spelling and grammar suggestions. A human still checks every figure, source and word before anything goes live. Occasionally we use it for important work like adding dinosaurs to colleagues’ photos. For the full rundown, see our Future and AI page.
https://cdn.mos.cms.futurecdn.net/PgBRvmiooysaY9uTwi5UkK-970-80.jpg
Source link
lindsay.handmer@futurenet.com (Lindsay Handmer)
