@manton Isn't Apple preparing for a more limited use case? And isn't it outsourcing some of this?

@manton Might simply be too early to talk about Apple in this space. I doubt anyone’s been able to test or measure much of substance yet.

@markstoneman It’s more limited but it’s going to touch so many devices, I expect it to be pretty massive in scale after a couple years.

@chartier That’s true, but the authors interviewed so many people, I think Apple should’ve been on that list.

Feel like this article doesn’t actually say much about AI energy usage, I don’t get any sense at all of what proportion of new data centers are dedicated to AI vs streaming video or whatever.

I’ve been checking out every article I can find on AI energy usage, because it seems important, but every article is like this. I have a really hard time assessing them because they don’t make any concrete reports or claims about how much energy AI is using or is projected to use.

If the power demands of AI result only in the successful development of a viable fusion reactor then the endeavor will be worthwhile. But remember the reason this moonshot-style effort is being discussed at all is that the power demands of AI are projected to be truly staggering. From the article:

One large data center complex in Iowa owned by Meta burns the annual equivalent amount of power as 7 million laptops running eight hours every day, based on data shared publicly by the company.

and

A ChatGPT-powered search, according to the International Energy Agency, consumes almost 10 times the amount of electricity as a search on Google.

Those are today's figures. If the field expands as expected the problem gets worse. There isn't a way to cache AI query results (as Google can for search) if everyone supplies a unique prompt tailored to their individual interests.

@sharkjacobs From the article:

In the Salt Lake City region, utility executives and lawmakers scaled back plans for big investments in clean energy and doubled down on coal. The retirement of a large coal plant has been pushed back a decade, to 2042, and the closure of another has been delayed to 2036.

Energy use in Salt Lake City is driven by Meta and Google. True, there's no hard evidence given that Meta and Google are building these data centers for AI, but it's the best conclusion we can make without a mole inside the companies sharing the receipts.

I wonder if Apple was asked but, as usual, declined to comment.

> there's no hard evidence given that Meta and Google are building these data centers for AI, but it's the best conclusion we can make

This is exactly what I’m talking about, it's a reasonable conclusion to draw based on the headline, but I'm not sure if the article actually supports it. The article has information about data centers, information about the power sources supplying data centers, but no information at all about how much power is used by AI

It’s just misleading to equate data center power usage with AI power usage as this, and almost every article I’ve found, does. There is no reasonable reason to believe that AI currently makes up more than 10% of data center usage by Meta, Microsoft, Amazon, or Google (or even Apple)

Generative AI is very computationally expensive, I think being concerned about its power usage is reasonable. I am concerned about its power usage! But articles like this don’t help me understand how concerned I should be, because they elide the massive difference between “data centers” and “data centers used for AI”

I wish I hadn’t said 10% in that last comment. My point is that if any of these companies were currently dedicating 10% of their global data center compute to AI I would be astonished. I didn’t want to say 1% because it seems aggressively dismissive because 1% is “the smallest” percent (rereading it, my tone was maybe already too aggressive) but by saying 10% instead it makes it seem like I think Amazon could reasonably be dedicating 5% of AWS to generating pictures of dogs on surfboards and customer service responses