TODAY'S RAMBLINGS
4 Minute Read
First, thanks to my dear friend and America’s favorite granny, Elizabeth “Polly” Michaels, for being a muse of sorts for today’s post.
Next: but it’s not what I thought it was going to be.
I began with the intent of writing one of my anodyne “Let me tell you about . . .” posts. This time the subject was to be - and still is - the giant data centers that provide the computing horsepower for all things Internet, but especially social media (Instagram, TikTok, Facebook, X, Reddit, etc.) and now, Artificial Intelligence (ChatGPT, Claude, Perplexity, Gemini, etc.).
In industry parlance, these are known as Hyperscale Data Centers. For obvious reasons, as you’ll see, if you don’t know already.
Yes, I was going to thrill you with a rehash of my glory days at TEECOM, leading a team of 25 engineers and CAD people here and in London designing the fiber optic networks at 9 of Facebook’s (now Meta’s) data centers.
“What do those look like, Portico?”
No problem.
This is Meta’s Eagle Mountain facility, in Utah; my team worked on this one, and I visited one just like it in Iowa.
Those vertical “legs” of the H-shaped buildings above are a quarter-mile long. So yeah, these are m-f’ing hyper facilities.
Yes, I was going to first wow you with how cool my job used to be. I was then - and still am - going to tie that into some facts and figures about where this is all headed, in terms of . . . the global environmental apocalypse they might cause if we don’t do something, and now.
And then I did my research for this post. And that’s when I found the information in the lead photo.
First, my heart rate soared. Then, I thought I might throw up.
Finally, I calmed down, and decided to feature today the sheer madness of Meta getting even a penny in tax subsidies. But almost a billion? What the fuck is the matter with us? We’ve not just lost the script, we’ve poured gas on it and then hit it with Dear Leader’s breath to light it aflame.
My God, Meta does not need any help! They can afford to build their own data centers, with their very own money!
10 Seconds of Seriousness: in 2024, Meta’s net income was $62.36 billion, up 59.5% from its 2023 net income of $39.1 billion. A reminder for you non-accounting kids out there that “net income” means “profit.” So after everything Meta/Facebook spends money on - salaries, lobbyists, and yes, building data centers - they still had over $60 billion left over. In one year.
And we’re giving them money to help build the facilities they need to conduct business?
Great deal, if you can get it!
Sadly, it’s not all likes and thumbs up.
These things use an amount of electricity and water that is difficult to put in perspective. Naturally, I will try; and it’s not just a smear job.
If you’ve been paying attention here, you know I am a staunch advocate of AI (if not social media), and these facilities are providing the computing power. That compute power is also difficult to get your head around.
Electricity
More industry jargon, but it’s important. The size of data centers is typically classified by the power they consume. When I was in the game (2019), 35 megawatts was considered a large hyperscale data center.
Now? The largest, AI-purposed facilities are 10 times that size. It’s true, and that is a LOT of power:
35 megawatts → power ~7,000 homes
350 megawatts → power ~ 70,000 homes
Flagstaff, Arizona and Woodland Hills, California both have approximately 70,000 homes.
Remember, that’s one single data center; it is estimated there were 1,150 hyperscale data centers worldwide by the end of 2024. Wait, there’s more: by 2030, there will of course be a lot more of them, and they are projected to consume 150,000 megawatts - 150 gigawatts. Speak of the devil; this is from perplexity.ai:
Thus, 150 GW annual consumption is more aligned with the scale of a large country’s total power usage rather than any city or state globally.
Water
Water usage ranges from 5.3-25.5 million gallons per year per MW of capacity.
Thus, a typical 100-150 MW facility consumes 530-3,825 million gallons annually.
For perspective, a single 15 MW data center uses as much water as three hospitals or two 18-hole golf courses.
But that’s a proverbial chip shot, because global water usage at hyperscale data centers is approaching one trillion gallons annually:
The average American uses about 100 gallons of water per day. One trillion gallons could supply approximately 10 billion people with a day’s worth of water.
One trillion gallons is about 1.6 day’s worth of flow from the Mississippi River into the Gulf of Mexico, and also the amount of water that flows over Niagara Falls in about 38 days.
In terms of rainfall, it would be like a 1-inch rain falling over an area roughly the size of Connecticut.
I’m getting thirsty!
Compute Power
Again - crazy town.
A typical facility has 50,000-200,000 servers delivering 1-2 exaflops of computing performance, while AI-optimized centers achieve 10+ exaflops. More? Google's latest facilities are designed for 26 exaflops.
WTF is an exaflop?
An exaflop represents 1 quintillion (10 to the 18th power) floating-point operations per second.
The human brain, for comparison, is estimated to perform roughly 1 exaflop worth of operations per second when you account for all neural activity.
Historical perspective: the first electronic computer, the ENIAC from 1946, could perform about 5,000 operations per second. An exaflop machine is roughly 200 trillion times faster than ENIAC. If ENIAC were a person walking at 3 mph, an exaflop computer would be traveling at 600 billion mph - fast enough to cross the entire Milky Way galaxy in about 20 minutes.
An exaflop is roughly 5-10 million times more powerful than modern consumer computers.
And exaflop is 30 billion times the computing power of the Cray-1 supercomputer from the 1970s.
At Least I Can Have It Analyze My Novel
I am no fool: the benefits (dangers, too) of all of this are themselves difficult to put in perspective. But there’s only one direction it’s all going, and that’s more and more and more.
For better or worse.
It’s major days here in SF: 3 Dead & Company shows in Golden Gate Park in celebration of their founding here 60 years ago. We’ll be streaming it live in 4K here with Ol’ Purple Label and André Aurich, and a shout-out to Friends and Family Suite returning champion Professor Howard Blum, Esq. and first-timer Penny Zamboni. They’ll be taking it all in from the front row, or something awfully fucking amazing.
And I’m not forgetting Byron Browne IV and Louise Lederhosen, also attending live, although they’ll be in the “cheap” seats.
Have a great show, guys!
FROM THE UNWASHED MASSES
Thank you for reading this newsletter.
KLUF
Sure, I could play a Grateful Dead or Dead and Company concert. But isn’t this more accurate: