Google’s environmental report pointedly avoids AI’s actual energy cost

Google has issued its 2024 Environmental Report, a greater than 80-page doc describing all the large firm’s efforts to use tech to environmental points and to mitigate its personal contributions. However it completely dodges the query of how a lot power is AI utilizing — maybe as a result of the reply is “far more than we’d care to say.”

You can read the full report here (PDF), and truthfully it’s received lots of fascinating stuff in it. It’s straightforward to overlook what number of plates an organization as massive as Google retains spinning, and there may be some actually noteworthy work in right here.

For example, it’s been engaged on a water replenishment program, whereby it hopes to offset the water utilized in its services and operations, ultimately making a web optimistic. That is finished by figuring out and funding watershed restoration, irrigation administration and different work in that space, with dozens of such initiatives world wide being not less than partially bankrolled by Google. It’s gotten to 18% of its water utilization replenished (by no matter definition of that phrase is used right here) that approach and enhancing yearly.

The corporate additionally takes nice care to frontload the potential advantages of AI in local weather, issues like optimizing watering methods, creating extra fuel-efficient routes for vehicles and boats, and predicting floods. We’ve highlighted a number of of those already in our AI protection, they usually really could possibly be fairly useful in lots of areas. Google doesn’t have to do that stuff, and lots of massive corporations don’t. So credit score the place credit score’s due.

However then we attain the part “Responsibly managing the useful resource consumption of AI.” Right here Google, so certain of each statistic and estimate till now, all of the sudden spreads its fingers and shrugs. How a lot power does AI use? Can anybody actually be certain?

But it should be dangerous as a result of the very first thing the corporate does is downplay your entire knowledge middle power market, saying it’s onlly 1.3% of world power utilization, and the quantity of power Google makes use of is barely at most 10% of that — so solely 0.1% of all of the power on the earth is powering its servers, in response to the report. A trifle!

Notably, in 2021, it determined it needed to succeed in net-zero emissions by 2030, although the corporate admits there may be lots of “uncertainty,” because it likes to name it, in how that may really occur. Particularly as a result of its emissions have elevated yearly since 2020.

In 2023, our complete GHG [greenhouse gas] emissions had been 14.3 million tCO₂e, representing a 13% year-over-year enhance and a 48% enhance in comparison with our 2019 goal base 12 months. This consequence was primarily resulting from will increase in knowledge middle power consumption and provide chain emissions. As we additional combine AI into our merchandise, decreasing emissions could also be difficult resulting from growing power calls for from the better depth of AI compute, and the emissions related to the anticipated will increase in our technical infrastructure funding.

(Emphasis mine on this and the quote beneath.)

But the expansion of AI is misplaced among the many aforementioned uncertainties. Google has the next excuse for why the corporate shouldn’t be being particular in regards to the contribution of AI workloads to its basic knowledge middle power invoice:

Predicting the longer term environmental affect of AI is complicated and evolving, and our historic traits possible don’t absolutely seize AI’s future trajectory. As we deeply combine AI throughout our product portfolio, the excellence between AI and different workloads is not going to be significant. So, we’re specializing in knowledge center-wide metrics since they embody the general useful resource consumption (and therefore, the environmental affect) of AI.

“Complicated and evolving”; “the traits don’t possible absolutely seize”; “the excellence … is not going to be significant”: That is the sort of language used when somebody is aware of one thing however would actually, actually choose to not inform you.

Does anybody really imagine Google doesn’t know, right down to the penny, how a lot AI coaching and inference have added to its power prices? Isn’t having the ability to break down these figures so exactly a part of the corporate’s core competency in cloud computing and knowledge middle administration? It has all these different statements about how environment friendly its customized AI server models are, the way it’s doing all this work to cut back the power required to coach an AI mannequin by 100x, and so forth.

I’ve little question there are lots of nice inexperienced efforts occurring at Google, and you may learn all about them within the report. However it’s necessary to focus on what it seemingly refuses to: the large and rising power value of AI methods. The corporate might not be the first driver of world warming, however regardless of its potential, Google doesn’t appear to be at a web optimistic simply but.

Google has each incentive to downplay and obfuscate these figures, which even in its lowered, extremely environment friendly state, can hardly be good. We’ll remember to ask Google to get extra particular earlier than we discover out whether or not they get even worse within the 2025 report.