What Are AI’s Climate Costs Really Paying For?
What Are AI’s Climate Costs Really Paying For?
Computing technologies like crypto and AI run on data centers, which are like the power plants of the internet.
Last week, AI was at the center of the Nobel Prizes. First, the Nobel Prize in Physics went to John Hopfield and Geoffrey Hinton for their invention of methods that can store and find patterns in data and which underlie the machine learning boom and modern neural networks. Then, the Nobel Prize in Chemistry went to Demis Hassabis, John Jumper, and David Baker for their use of these kinds of machine learning tools to predict the structures of the biological molecules called proteins and design ones that are new to nature.
I gotta give the bio team at Quanta, and especially writer Yasemin Saplakoglu, a pat on the back for predicting this: In June, we published a 10,000-word feature, the biggest story I’ve worked on, unfolding the full story of the development of protein folding AI, including the chemistry Nobel winners. Through working on this story, I learned how machine learning was developed and how it works (see Part 2), an increasingly useful understanding. What I love about the story is that the full unpacking of this specific case study reveals what kinds of problems machine learning is well suited to solve (hint: ones where we have tons of data already collected that describe both the problem and the solution), the kind of disruption high-quality AI causes, and what mysteries it conceals.
So I thought I’d take a beat and for the newsletter look into the climate costs of AI — and try to assess what those costs are getting us.
Big Tech Takes Big Electricity
Computing technologies run on data centers, out of sight and out of mind, that are the power plants of the internet. A data center “is a giant building with a lot of power, a lot of cooling, and a lot of computers,” said one Google employee in a 2015 promotional video — sounds like a power plant to me. These server farms take in power from the grid and convert it into computerized experience, whether that’s computing internet traffic, crypto coins, or neural networks.
As long as the grid is powered by fossil fuels, data centers will emit carbon dioxide and produce heat that requires more energy to cool — and that makes them not-insignificant global warmers. According to the International Energy Agency, data centers and transmission networks each accounted for 1-1.5% of global electricity use in 2022. That might not seem like a lot, but AI requires a lot of power, and ChatGPT didn’t debut until the end of November 2022. The Guardian:
AI is far more energy-intensive on data centers than typical cloud-based applications. According to Goldman Sachs, a ChatGPT query needs nearly 10 times as much electricity to process as a Google search, and data center power demand will grow 160% by 2030. Goldman competitor Morgan Stanley’s research has made similar findings, projecting data center emissions globally to accumulate to 2.5 billion metric tons of CO2 equivalent by 2030.
It’s hard for me to judge how big a problem that really is, relative to, say, electric power plants or gas-powered cars. But this summer, Google announced that its emissions had risen 13% in 2023 over the previous year, citing AI and the demand it puts on data centers. The Associated Press:
Some experts say the rapidly expanding data centers needed to power AI threaten the entire transition to clean electricity, an important part of addressing climate change. That’s because a new data center can delay the closure of a power plant that burns fossil fuel or prompt a new one to be built. Data centers are not only energy-intensive, they require high voltage transmission lines and need significant amounts of water to stay cool. They are also noisy.
Meanwhile:
According to a Guardian analysis, from 2020 to 2022 the real emissions from the “in-house” or company-owned data centers of Google, Microsoft, Meta and Apple are probably about 662% — or 7.62 times — higher than officially reported.
Protect the Data Centers at All Costs
My recommended read of the week is “Silicon Valley’s Influence Game” by Charles Duhigg in The New Yorker about the growing political power of the crypto and AI industries. He reports that pro-crypto donors account for almost half of all corporate donations to PACs in the 2024 election cycle and that “the tech industry has become one of the largest corporate donors in the nation.” Following the story of one of these PACs, called Fairshake, Duhigg holds crypto donors responsible for taking out Rep. Katie Porter (D-CA) in her Senate primary, and more:
In total, Fairshake and affiliated super PACs have already spent more than $100 million on political races in 2024, including $43 million on Senate races in Ohio and West Virginia, and $7 million on four congressional races, in North Carolina, Colorado, Alaska, and Iowa. $3.5 million was used to help vanquish two left-wing representatives who were members of the so-called Squad: Jamaal Bowman, of New York, and Cori Bush, of Missouri. Of the 42 primaries that Fairshake has been involved in this year, its preferred candidate has won 85% of the time. … Its donations to political candidates are on par with those of the oil-and-gas industry, the pharmaceutical industry, and labor unions.
OpenAI recently hired Fairshake’s consultant Chris Lehane (whose story is recounted in Duhigg’s feature) as its vice-president of global affairs. Its political campaign is “in its early stages.”
All of this money and now political power is invested to protect a couple of cloud-based ideas: crypto and now AI. Which means that really what they are protecting in in the physical plane are data centers.
What Is the AI Industry, Really?
AI and machine learning can help solve some difficult problems, those for which we have great descriptive data. It can find patterns in large complex datasets that outscale the human mind’s ability. Predicting the structure of a protein is one such application; surely there are many more applications in science and medicine. (One favorite tool: Cornell Lab of Ornithology’s Sound ID tool in the Merlin app, which uses machine learning to ID birdsong to species in real time.) But right now the main way most of us are affected by AI is through generative AI tools: Using neural networks to produce text or images, bypassing the creativity, effort, and choices of the beautiful human mind.
The vast bulk of AI-generated results are completely silly, ugly, garbage — what Max Read called “AI slop” in New York magazine. All over the web you can find great human minds fretting about the AI “mind” (e.g. Ted Chiang, Emily Bender). At the bar we trade tip on how we are or are not using AI tools. We will continue to fret about this for some time. In the meantime, what is generative AI really? It’s a tool for making a few guys a lot of money. Tech researcher Jane Ruffino:
Sam Altman is making money, Elon Musk is making money, and Marc Andreessen and Peter Thiel (whose name I really hope you know by now) are doing better than ever. There’s a small group of people, mostly men, who are connected to each other, and to people and plans to do things like ethnically cleanse San Francisco, the bad and deceptively named Long Termism and Effective Altruism movements, along with the extremely creepy natalist movement, and many of them are openly backing fascism. Those are the people getting rich off of AI, not because it’s revolutionary, but because they’ve figured out how to make money off of anything and everything, at the expense of anyone and everyone.
…
And as the companies start to fall, they’ll consolidate into fewer and fewer hands, which, again, will be disproportionately in Silicon Valley, all of which is bad for the global Internet — moving us further from the distributed network of networks that it was supposed to be — and that’s bad for human beings everywhere.
Palate Cleanser: Ed Yong Gets Personal
Journalist Ed Yong covered the pandemic rigorously and empathetically; he seemed prescient, able to predict how it would play out. At XOXO Fest he shared his personal experience and toll of this work and how he has recovered. Definitely worth your time.
When I started this edition I wasn’t really sure where it would end up, and I think it got somewhere. Be safe out there and don’t forget to take care of your mind. It’s way better than a computer.
Warmly yours,
Hannah