What to Do About AI's Environmental Impact
Hey folks,
Eva sent an email to DTNS this morning asking if we could address "the ecological impact of AI going forward." A story from MIT Technology Review came out today about that very topic. Perfect timing! We’ll address it on DTNS too, Eva, but here’s a start.
Free users, don’t forget this newsletter comes out Monday-Friday for a few cents a day if you’d like to upgrade. Here are the topics from the past week:
Copilot+ Recall Is a Privacy Threat https://techtom.substack.com/p/copilot-recall-is-a-privacy-threat
Is OpenAI Ripping off Scarlett Johansson? https://techtom.substack.com/p/is-openai-ripping-off-scarlett-johansson
Microsoft's AI Laptop Play https://techtom.substack.com/p/microsofts-ai-laptop-play
AI Is Now Big Business https://techtom.substack.com/p/ai-is-now-big-business
Cheers,
Tom
Big Story
"AI is an energy hog. This is what it means for climate change. | MIT Technology Review"
A while back I tried to find out what the energy use of cryptocurrency was like compared to the energy used to process non-crypto transactions. I wanted to see if it really was an energy hog or if it was pretty much the same as other financial transactions. Comparing the energy usage to the Netherlands wasn’t very useful in my opinion. Since then, Ether has moved to proof of stake and Bitcoin mining has begun to decline.
In its place is a fear over the use of power by data centers running energy-hungry AI services. There’s no debate that these services use a lot of computing power and a reasonable expectation that this demand will continue to increase. Some countries, including Ireland, have already started to place limits on building new data centers simply to protect the viability of the local power grid. With any large energy consumption, there is a reasonable concern that the consumption will increase emissions and CO2 levels in the air.
MIT Technology Review's Casey Crownhart took a look at the situation so we can make an educated decision about what to do about it. Read the full article to get a full understanding, but here are some of my takeaways:
Don’t fall for the “one image is as much as driving a car four miles” comparisons. They are as useful as comparing it to the usage of the Netherlands. There are much better and more useful data.
Electricity from all data centers—including cloud computing, crypto, and of course AI—made up 2% of electricity demand in 2022. It is expected to double by 2026, so around 4% of electricity demand.
Another way to look at it: The total demand for data centers (not just AI) could be between 160 and 590 TWh by 2026. Total worldwide electricity demand from all uses by people is expected to increase 3,500 TWh by 2026. Electric vehicles and industry growth will be larger sources of demand than all data centers and therefore than AI.
AI services will challenge companies to meet climate goals. Microsoft has a zero-emissions goal by 2029, but its emissions have been rising because of AI use. Interestingly, it’s not the electricity use that is the problem but the emissions from construction of new data centers—steel, concrete, etc.
Crownhart notes a study in 1999 that warned that personal computer use would eat up half of US electricity demand by 2009. It didn’t. The computers got more power efficient. It’s not the rise in demand that’s important, it’s what you do about it that matters.
I’ll quote from Crownhart here: “If we build more fossil-fuel plants to meet our growing electricity demand, it’ll come with negative consequences for the climate. But if we use rising electricity demand as a catalyst to lean harder into renewable energy and other low-carbon power sources, and push AI to get more efficient, doing more with less energy, then we can continue to slowly clean up the grid, even as AI continues to expand its reach in our lives.”
This is excellent reporting and I hope it gets more attention. All power use is rising and AI is not the largest part of it. But that’s no excuse not to act responsibly. In my opinion, if we focus on AI as the problem we might make the wrong choices, like restricting AI use. That would deprive us of the benefits of AI without solving the problem since electricity use would continue to rise by other means. And it’s worth noting that solutions to more efficient and zero-emission energy use might come from someone aided by an AI tool or two.
So personally, I’m not panicking about AI’s power use. I am a little nervous about whether we’re doing enough to make sure our power sources are not making the CO2 problem worse.
More Stories
"Truecaller and Microsoft will let users make an AI voice to answer calls - The Verge"
Caller ID company Truecaller is partnering with Microsoft to let you answer the phone without being there. The Truecaller app will let you record your voice to train Azure AI Speech to generate animation of you. It’s the same thing you can do from 11Labs and a number of other companies. Truecaller will do its normal thing of screening calls and giving you the option of letting the app answer the call to find out more from the caller. But now instead of a preset voice, the voice can sound like you. Microsoft says it adds watermarks to the generated voices and requires full consent of whoever is being recorded.
There’s only so much you can do to prevent misuse of something like this, though. Requiring consent means you can punish people who abuse the privilege when they’re caught, but it won’t stop them from trying. That may be too much for some people to feel comfortable with. But this is one small example of the many issues we are going to have to confront with generative models.
"Google introduces password sharing for family members"
Google Password Manager has added family password sharing in the May Google Play Services update v24.20. If you’ve set up a family group in Google you can choose to share a password with it. Each family member then receives a copy of it in their password manager. This is the feature it announced in February. It should roll out to all instances of Password Manager but does not appear to have made it to the desktop version of Chrome yet.
"Nvidia (NVDA) earnings report Q1 2025"
"Nvidia’s Business Is Booming. Here’s What Could Slow It Down. - WSJ"
Nvidia beat all the expectations for profit and sales even though those expectations were high. Revenue from data centers rose 427% from last year and is firmly the company's most important division now. Gaming revenue was up 18%. And its smaller networking hardware business rose as well. Nvidia will also split its stock.
The Wall Street Journal has a good article on the threats to Nvidia’s continued boom. Obviously, competition from chipmakers, as well as chips and data center building from Google and Microsoft, are a threat. A bigger threat is that the buildout slows as companies finally reach the level of demand at some point. That may happen this year, but it might not happen for a few years. There’s also a push for more efficient AI, which could mean lower hardware needs. Nvidia’s push into networking gear is a good hedge against any slowing in the AI chip-making demand.
"Arc Search's new Call Arc feature lets you ask questions by 'making a phone call' | TechCrunch"
Arc users can use the app to make a “call” and ask a question. The model will understand the voice prompt and respond with its own answer spoken thanks to text-to-speech. The model will stay on the call as long as you want to help you with things like preparing dinner or building Ikea furniture. It’s a cute and familiar interface for voice search. Probably not as silly as it sounds to be honest.
"OpenAI, WSJ Owner News Corp Strike Content Deal Valued at Over $250 Million - WSJ"
"OpenAI didn’t copy Scarlett Johansson’s voice for ChatGPT, records show - The Washington Post"
Two big wins here for OpenAI. OpenAI reached a deal with NewsCorp to pay for access to its content. Details weren't made public but a source told the Wall Street Journal that NewsCorp content would not appear immediately after publication. And OpenAI would also use NewsCorp journalists as consultants. NewsCorp owns Dow Jones, which publishes the Wall Street Journal. The deal also includes MarketWatch, the Times of London, the Australian, the New York Post, and more.
Meanwhile, the Washington Post, which does not have a deal with OpenAI, found internal OpenAI records that show the actress whose voice was used to train the "Sky" model was hired before OpenAI reached out to Scarlett Johansson about a deal. A casting call was put out in May last year that said actors should be “warm, engaging [and] charismatic.” No mention was made of Johansson or the movie Her. Johansson was contacted later that year in September after the Sky voice had been created. The Post got access to documents, recordings, casting directors, and the Sky actress’s agent. The Sky voice remains paused while OpenAI tries to reach an agreement with Johansson.
For Context
"All-screen M5 MacBook with foldable display now rumored for 2026 - 9to5Mac"
An all-screen foldable laptop with an M5 chip is now hoped for release in 2026. That's a year earlier than Kuo previously said. Screen size may have been reduced from 20.3 inches unfolded to 18.8 inches. That would make it the equivalent of a 13-14 inch MacBook Air. Apple wants to have no crease.
The investigation was conducted by Vancouver-based law firm Roper Greyell.
"Google to build first subsea fibre optic cable connecting Africa with Australia | TechCrunch"
It will be the second cable to connect that part of the world to Perth, Australia. The first goes from Oman in the Middle East. It's more redundancy and keeping up with the Joneses for Google.
"Google Pay will now display card perks, BNPL options and more | TechCrunch"
I like the idea of seeing if you have any relevant perks before selecting a card.
"Southwest fares now displaying in Google Flights - The Points Guy"
This is only a big deal in the US, but it’s a big deal. Southwest flights were listed, but you had to click through to see how much, which slowed down price comparisons quite a bit.
"LastPass is now encrypting URLs in password vaults for better security"
They said they would do it, so I suppose we should acknowledge that they did.
Interesting Reads
"Meta AI chief says large language models will not reach human intelligence"
"After almost 28 years, Super Mario 64 has been beaten without using the A button | Eurogamer.net"
"Advertisers react to 'underwhelming' Google Marketing Live 2024"
"This ex-Googler designed a heat pump you'll actually want in your home - Fast Company"