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

Google has issued its 2024 Environmental Report, a more than 80-page document describing all of the massive company’s efforts to apply tech to environmental issues and to mitigate its own contributions. But it totally dodges the question of how much energy is AI using — perhaps because the answer is “way more than we’d care to say.”

You can read the full report here (PDF), and honestly it’s got a lot of interesting stuff in it. It’s easy to forget how many plates a company as big as Google keeps spinning, and there is some really noteworthy work in here.

For instance, it’s been working on a water replenishment program, whereby it hopes to offset the water used in its facilities and operations, eventually creating a net positive. This is done by identifying and funding watershed restoration, irrigation management and other work in that area, with dozens of such projects around the world being at least partially bankrolled by Google. It’s gotten to 18% of its water usage replenished (by whatever definition of that word is used here) that way and improving every year.

The company also takes great care to frontload the potential benefits of AI in climate, things like optimizing watering systems, creating more fuel-efficient routes for cars and boats, and predicting floods. We’ve highlighted a few of these already in our AI coverage, and they actually could be quite helpful in many areas. Google doesn’t have to do this stuff, and many large companies don’t. So credit where credit’s due.

But then we reach the section “Responsibly managing the resource consumption of AI.” Here Google, so sure of every statistic and estimate until now, suddenly spreads its hands and shrugs. How much energy does AI use? Can anyone really be sure?

Yet it must be bad because the first thing the company does is downplay the entire data center energy market, saying it’s onlly 1.3% of global energy usage, and the amount of energy Google uses is only at most 10% of that — so only 0.1% of all the energy in the world is powering its servers, according to the report. A trifle!

Notably, in 2021, it decided it wanted to reach net-zero emissions by 2030, though the company admits there is a lot of “uncertainty,” as it likes to call it, in how that will actually happen. Especially because its emissions have increased every year since 2020.

In 2023, our total GHG [greenhouse gas] emissions were 14.3 million tCO₂e, representing a 13% year-over-year increase and a 48% increase compared to our 2019 target base year. This result was primarily due to increases in data center energy consumption and supply chain emissions. As we further integrate AI into our products, reducing emissions may be challenging due to increasing energy demands from the greater intensity of AI compute, and the emissions associated with the expected increases in our technical infrastructure investment.

(Emphasis mine in this and the quote below.)

Yet the growth of AI is lost among the aforementioned uncertainties. Google has the following excuse for why the company is not being specific about the contribution of AI workloads to its general data center energy bill:

Predicting the future environmental impact of AI is complex and evolving, and our historical trends likely don’t fully capture AI’s future trajectory. As we deeply integrate AI across our product portfolio, the distinction between AI and other workloads will not be meaningful. So, we’re focusing on data center-wide metrics since they include the overall resource consumption (and hence, the environmental impact) of AI.

“Complex and evolving”; “the trends don’t likely fully capture”; “the distinction … will not be meaningful”: This is the kind of language used when someone knows something but would really, really prefer not to tell you.

Does anyone actually believe Google doesn’t know, down to the penny, how much AI training and inference have added to its energy costs? Isn’t being able to break down those figures so precisely part of the company’s core competency in cloud computing and data center management? It has all these other statements about how efficient its custom AI server units are, how it’s doing all this work to reduce the energy required to train an AI model by 100x, and so on.

I have no doubt there are a lot of great green efforts going on at Google, and you can read all about them in the report. But it’s important to highlight what it seemingly refuses to: the enormous and growing energy cost of AI systems. The company may not be the primary driver of global warming, but despite its potential, Google doesn’t seem to be at a net positive just yet.

Google has every incentive to downplay and obfuscate these figures, which even in its reduced, highly efficient state, can hardly be good. We’ll be sure to ask Google to get more specific before we find out whether they get even worse in the 2025 report.