I don’t get it, why do we need so fucking many? Reading the thread about temps downwind from data centers and it’s describing all the multiple HUGE data centers just around phoenix and it’s just, like, why are they building so fucking many???
It’s one thing to be like “yeah we’re building a facility that uses more power than human civilization to power the Lying Machine” but it’s like they’re building fucking thousands of them. What the fuck is that?
Let’s be incredibly generous and assume The Nerds invent intelligence, even then what is the fucking point of all this shit
Remember how people accused China’s economy of being fake and propped up by mass construction of housing that they did not need?
It’s that, but it’s data centers.
They always accuse their enemies of doing what they themselves are talking about doing behind closed doors.
a bit of a tangent but bear with me. ‘traditional’ data centers were either for content delivery networks (mainly storage) or web servers (general computing, CPUs). their focus was usually to optimize networking as they were built for the internet. they were and still are extremely profitable. their electricity demand was elastic (did not use 100% of their full capacity all the time) and generally low compared to their service, extremely lean crew (a massive data center can run with 25-100 people), low maintenance cost (most of the spending is in capital expenditures), and always in demand. Amazon is basically a data center business (AWS) with a retail component on the side. because of this over the past three decades politicians created multiple incentive programs in order to attract the data centers. granted these data centers really didn’t contribute that much to the local economy and cooling them was still an issue, but it was basically seen as a free money glitch through taxes.
then crypto happened. after the first mainstream bitcoin rally around 2017, crypto farms became a thing. bitcoin has a ‘proof-of-work’ scheme where in order to validate a bunch of transactions you have to find a random value that combined with the transaction data, produces a hash with specific restrictions, and if you do you’ll get rewarded with bitcoin. what that means in practice is that you need to have a bunch of GPUs running 24/7 guessing random numbers and checking if they work or not. extremely demanding and wasteful with most of its work being useless, generating shitcoins for gamblers that insisted this was the future of money.
basically up until the AI boom, these fuckers both ruined grid stability and GPU prices as well as the second hand GPU market. they went everywhere where there was cheap electricity and set up shop, ruining the local grid, buying out and scalping GPUs, and finally dumping the worn out ones as “basically brand new” on the second hand market for their “after scheme scam”. even with their industrial scale use of electricity their impact was limited worldwide because it was just shitcoins. illegal crypto farms were partially responsible for the iranian (and following that iraqi) grid failures for multiple years because iran had (and has) one of if not the cheapest electricity anywhere in the world (which will change this summer as the government is planning to use the war as cover to ‘liberalize’ most of the electricity sales through energy exchanges).
getting to the point, most of these ‘GPU as a service’ companies started from crypto, and pivoted to AI once NFTs crashed and permanently turned most people away from crypto. so you have these companies starved for cash which already have some experience in the field, silicon valley flush with cash looking for the next big thing after crypto and metaverse flopped, and a business model that has traditionally been very profitable with massive incentives behind it. so you get these new data centers popping all over the place.
without explaining how current AI training methods work, these new data centers have extremely inelastic (basically 100% usage all the time) and insanely high power demands. every negative consequence of data centers comes from the fact that there is an ungodly amount of electricity being turned into heat and sound due to resistance in semiconductors, in order to crunch numbers. this kind of use is absolutely killing the local grids which were not designed to handle this much and this kind of load, not to mention the sound pollution, the heat pollution and the water used for cooling.
another thing that i barely mentioned was CAPEX or capital expenditure. CAPEX is basically how much you spend on building long term infrastructure, which is not recorded as a loss until there is depreciation or asset impairment. how this should be calculated is a subject of debate but from what i see most of these calculations for data centers are bullshit. they are using 5 to 6 year constant depreciation for GPUs with basically no asset impairment. first the number itself is bullshit, most GPUs will probably not last beyond 1 or 2 years. besides that this number used to be 3 to 4 years which magically got bumped into 5 to 6 years while the load has become increasingly more destructive. second the workload is massive but it is not stable, the GPU is not wearing out evenly and after you are done with it it’s completely worthless, and every new generation of GPUs and AI models tanks the value of existing GPUs. what i am saying is that GPU spending is extremely front-loaded while companies are pretending it’s not, and are spreading the massive losses over multiple years. even with that the numbers do not look good at all.
will these data centers be profitable? not all of them but some will. despite all the jokes and insults (which i also do don’t get me wrong it’s fun to shit on venture capital dudes), the investors are not dumb. if we can see AI is getting more efficient, that this amount of capacity is not needed and most of these will go bankrupt, they do so as well even if they don’t admit it publicly. they just think that after the dust has settled their business will be the one left standing. but in order for that to happen, you need to spend a lot right now otherwise the competition will eat your lunch. competition (as much as it is worshiped in the civil capitalist religion) is inherently wasteful and excessive. this bubble has still not reached the levels of train mania of the 19th century britain, and the logic is that “sure a lot of people were miserable for a decade or two following the bubble, but after that UK was left with state of the art infrastructure which allowed it to become the biggest economy in the world, and our current economy is much more dynamic so if it worked out well before in a worse situation, it should work now”.
TL;DR: why do we need so many fucking data centers? we don’t need this many, but we’ll need some, and the ruling class have nothing “better” to spend their insane amount of wealth on. they think that based on past trends this will work out for them no matter how it turns out, so they are just doing it, the rest of us be damned.
It’s pretty impressive how much electricity they can turn into heat without everything turning into soup.
Contractors grifting, tech bros buying in on hype, politicians getting lobbied.
I hate it.
What i want to say is very
God all this waste and I’ve yet to see literally anything nice come of it all. When I use ai to make fun stuff I do local model or piggyback off shared consumer gpus. I try to make it a point to not actually use the big llms, like chatpgt or grok, because of how wasteful they are and half the fucking time they’re wrong anyway.
I will say that (at least the google search ai, my primary interaction) it isn’t bad at summarizing search results, by which I mean half the time I google anything i’m looking for reddit discussions where people who know what they’re talking about are discussing shit, and it does an okay job at finding and summarizing those. It’s consistently better at what i expect it to do than people here act like it is. But it shouldn’t take 10000 data centers and enough energy to raise global temps by 10 degrees C
I used Gemini the other day for the second or third time ever. I asked it to reference the 2021 IPC (internat’l plumbing code) and lay out a mocked up bathroom that I am remodeling in the next two weeks. I hate to say it, but it nailed the layout. It followed ADA rules to a tee. It laid out my underground work, and gave me a take off list that was a couple fittings short of my written list. It priced out fittings, pipe, and fixtures. The only thing it couldn’t do for me was estimate my time to work everything.
I won’t use it except in very rare circumstances but anecdotally it seemed to handle taking a known ruleset (plumbing code) and producing something off of it decently well. It’s even drew up my layout in ASCII. I attempted this a couple years back with the 2018 IPC and it fell on its own face trying to layout then.
I’m not sure what my point is here, at any rate.
I think there’s a lot of utility for these things and people on here are frequently piss drunk on their own ideology regarding how actually useful they are, but at the same time I wouldn’t say it’s worth the massive resource cost needed to sustain it
that’s my take on the current state of LLMs, obviously they used to be even worse tho
I think in their current state they could definitely save a lot of people a lot of work i.e. your example with plumbing, as long as they’re either somehow coded to prevent fucked up bullshit that goes against building regulations, or as long as people are trained/aware enough to look over and fix the results
Despite my opinion of AI there are some use cases for it. I used to be one of those aforementioned piss drunk ideology guys. Im certainly not a convert, but im sort of coming around to the idea that there can be use cases for it, but like you said the resource abuse it requires to make Will Smith eat spaghetti is a travesty.
I think it’s pretty obvious there are at least some valid uses for it, like even if it just does simple customer service stuff (before escalating to an actual person) I’m sure it’d save a lot of labor and also people actually being abused by mad hogs
but a lot of the “uses for it” it’s also pretty obvious (to me, a total expert who definitely is not a layman when it comes to computers) were also solved forever ago and without token nonsense. Like my only interaction with the Spotify AI stuff so far has been to try to make playlists where it doesn’t play songs I’ve heard before. But I am almost 100% certain “make a playlist with only new plays” was something we could have done with one one millionth the energy in like, 1990
It’s consistently better at what i expect it to do than people here act like it is.
It’s basically either copying an already extant text nearly verbatim instead of just providing the original source, or it’s latched onto something categorically different and is completely wrong. And in both of those cases it can also just introduce random errors into the answer too.
I’ve used the duckduckgo interface to chatgpt for researching things where accuracy doesn’t matter and I just need something that’s vaguely in the right ballpark to build a character generator or the like, and in my experience that’s been basically just a shittier version of what google used to be but a less shitty version of what google is now, and it helpfully summarizes complex topics it cannot possibly summarize correctly into usable chunks of garbage.
yeah you say that though but the “completely wrong” seems to be really rare for the things I search for, and when I see people talk about how wrong it is I can’t help but think their last interaction with it was literally years ago
When I say “it’s better than I expect it to be” though like I mean I expect it to just distill me a bunch of reddit posts with errors throughout, but it’s usually generally accurate (distilling those reddit posts, and yes, redditors don’t know shit) and the time it takes me to suss out the Wrong Ass Shit it says is about equivalent to the time it’d take me to browse whatever Reddit thread it’s stealing from to suss out who knows wtf they’re saying
It’s not enough for some to hate the technology they also have to convince themselves that it doesn’t even work.
While I am loathe to speak fondly of AI, I did run into a very useful case the other day. It helped me fix my UPS.
Had to replace the battery, but the brand new one was also showing expired. I went back and forth with support for weeks, they ended up sending me a replacement battery. It too showed as expired. Support told me I would need to replace the unit.
So I popped the issue into Deepseek and it managed to find the exceedingly simple solution buried in a years-old forum thread. Apparently there’s a configuration mode and I just had to change one value from 0 to 1 to indicate to the unit that the battery had been replaced.
Their own (outsourced) support department didn’t even seem to know about this, but it saved me like $1000 having to buy a new UPS.
It’s getting consistently worse, at least Gemini is, somewhere around December it stoped translating some words, so there would be a random Chinese character in an English paragraph, and sometimes it was a made up character, not in the original, but most often it translates characters as numbers, what the fuck is a “question mark 17” supposed to be? It didn’t do that last year; this is particularly egregious in chemistry text, if you don’t already know what is supposed to be happening, it’s just gibberish about question mark 13 and 17 and 11s.
If you look at the listings for data centre locations on places like datacentres.com for the big tech companies (Amazon, Google, Meta, Microsoft, etc), two trends jump out.
- The majority of these data centres globally are in the US.
- The bulk of those data centres in the US are in Virginia, usually not far from the Pentagon.
It’s all about asking ChatGPT where to bomb next.
A question answered by point 2. of your comment
Bubble is gonna pop and they will all be conveniently repurposed into surveillance infrastructure.
Before 2020 they already dual-purposed as surveillance infrastructure, it’s just a matter of quantity/scale.
Ironically, they may end up with so much generated data that it becomes harder to respond to threats that the machine learning model doesn’t account for.
Reductively, it’s a simple math problem. Nvidia sold something like 10GW of GPUs in the last 3 years. That means people spent money on them. They spent money on them expecting to make money by using them. To use them, they need to be in data centers. By some estimates, the industry has only added 1GW of data center capacity in the past 3 years. So they need to keep building to house the hardware they bought to make money from the hardware they already paid for.
Less of an ouroboros, more of a human centipede
I visited my sister last month and we went to Costco to buy some supplies for a BBQ we were gonna have with her husband and some of his family.
In the parking lot I heard an annoying dull humming sound and asked her wtf that was. She pointed to some warehouse looking building across the highway and said “data center”. God fucking forbid you have an apartment near one of those things!
It’s not even that it’s impossible to to avoid the noise pollution. It would cost more and they just don’t care.
It’s the rate of profit. Hogs have to invest in something even if there is no return, there may be, the line says it’s data centers. Economically the goal is for the big hogs to cannibalize the smaller ones, leave them holding worthless stock once it becomes apparent that there was no profit to be made on those investments.
The batch of tech billionaires that rose up through the dotcom and mobile booms were all young nerds in the 90s (except Elon Musk, who was never young). They read all the same tech singularity wank that every other 90s nerd read.
They all grew up assuming the singularity will happen during their lifetimes. But the last AI boom was 50 years ago, and those tech billionaires aren’t getting any younger. So this one must be it. “It’s not happening” is unthinkable.
So we know this round of AI is going to cause the tech singularity, and thus make owning lots of computers the most important thing in the world, the crux that all power flows through, the lever of the world. Thus it is very important to make more buildings full of computers than any of our rivals, so we’re the most powerful and special in the new era.
And if any of the tech CEOs haven’t fallen for this, enough investors believe having the most data centers is the key to winning the future, that the CEOs still need to build for that future to keep their stock prices high.
i wanna blow up everything
I mean, it’s cringe, and it matters to some extent, but it doesn’t matter to nearly the extent that the news, or the stock prices, say it does. Add together the handful of people still being crypto grifters, smart home enthusiasts, and VR gaming fans, to get the sum total result of the last several tech hype cycles, and get not enough to worry about.
People are gonna blame a lot of stuff on AI, but at the end of the day it’s just bosses trying to overwork and underpay workers, with new rhetoric attached.
So the AI bubble can keep growing bigger
I don’t even know how it does that!
is the promise that it’ll make the computer Overlord smarter? faster?
I just don’t get it especially when i can just type random shit into the google search bar, knowing this costs them money, and it seems to be entirely infinite and not rate limited, so I just don’t understand what all these additional facilities are even for if it’s already got enough juice to “work” consistently
The money in the stock market has to go somewhere and the rest of the economy is already in a recession.
You don’t get it for the same reason you did not get NFTs.
WHAT DO APES HAVE TO DO WITH THIS
mods help mods help hlep modddddssss i spil my slurp jice
EXACTLY
but everything you search with google just adds another data point about you which they sell to the thousands of greedy little advertisers out there
so if I just sit around goonin gand googling all day I’m adding to the GDP??
Gooning domestic product
Life must goon
ChatGDP
Palestine is the test run. We’re next in line
Once they finally build god they gotta store him somewhere
put his ass in the world trade center 😂😂😂
There’s some assumptions the techbros are making:
- AI will continue to become more effective at a steady rate
- AI will become so prolific to the economy that you must have it and lost of it
- We have the energy to support it and/or we will build that
- AI’s need for computation will continue to grow
In reality
- AI’s growth in complexity and effectiveness is starting to stall and we are seeing huge diminishing returns the bigger and more complex the models are under the current design paradigm
- While AI is a great tool for a few things it is starting to prove it isn’t actually a replacement for labor (that’s another discussion)
- We don’t have the energy to support it and we won’t be getting it within a decade here in America
- AI’s need for resources is actually shrinking and every month we get new methods and models that reduce the amount of resources we need in both storage and compute to achieve the same or similar results
The false assumptions are driven by companies like NVIDIA to drive up their sales and stocks. It’s a bubble and it will burst.
For the record I think AI is here to stay as a tool. I also think it’s way overhyped and going to crash. I think the future is in more efficient, smaller models that companies will be running on their own servers. Even now, the ability for the average gamer with a mid-range rig to run an AI girlfriend or boyfriend right on their home PC without ever dialing out to a datacenter is already pretty impressive. The models you can run locally to help you with coding tasks are even better. At least, good enough for small-bean developers.
it takes tens of thousands of gpus to run these massive llms for even a small subset of the users out there.
in order to run them for millions and eventually billions of people, they’ll need hundreds and thousands of data centers hosting tens of thousands of gpus.
it also takes a fair amount of hardware to train these models in the first place and to iterate on them and all that crap.
the good news is they won’t be able to build these fast enough to avoid the bubble pop
