OpenAI's new $400 billion announcement plan reveals both growing AI demand and circular investments.
See full article...
See full article...
Why does OpenAI need six giant data centers?
- Thread starter JournalBot
- Start date
I won't pretend to assume Texas does anything sensibly, but following the Enron-fomented Gray Davis mis-managed energy fiasco in California, various industrial users in our area were required to develop plans to reduce draw by XXX amount within a set time frame, or face being totally blacked out. We did so, but to the best of my knowledge were never obliged to execute on it.
Upvote
4
(4
/
0)
This has been the case for some time and it’s because this is the best way to measure the capacity of a facility. Computing power advances and changes fast but the energy supply capacity for a facility does not. So, yes, computers themselves are measured in compute capability but the DC is measure in energy capacity to host those computers.
Upvote
1
(1
/
0)
I'm going to echo several others by just expressing my skepticism that this kind of massive investment can ever be recouped with these products. The sheer scale of it is absolutely jaw-dropping, and the number of companies who are in on it is absolutely shocking. Literally nobody has any sense at all.
Upvote
7
(7
/
0)
Upvote
0
(0
/
0)
Singularity Sky and Iron Sunrise by Charles Stross are also pretty interesting explorations of what happens to a society when too much of the locally available mass gets turned into compute
Upvote
5
(5
/
0)
There's a fine line between mass delusion and nefarious intent, and I don't know where to draw it. I wonder if OpenAI is simply banking on becoming too big to fail by burrowing itself deep into investment funds and government services before the inevitable collapse of the bubble. The unfortunate lesson of the 2008 crisis was that holding the economy hostage pays off big time if you play it right, and that personal accountability at such a scale doesn't seem to exist.
Upvote
5
(5
/
0)
No they can't.
The majority of the capital is being spend on the IC's inside the data centres. Those IC's will become obsolete (or at least uneconomical to use at scale) in only a couple of years.
Yes, the buildings could be reused, but there's no telling how standards will evolve and whether those buildings will be appropriate for futures tech.
This is so different from the dot com bubble. at least fibre could be used in the future. This is generating waste!
Upvote
6
(6
/
0)
For a bit more context, Zuck was saying "surveys show most people want and are happiest with about 15 friends. But Americans only have 3.". His solution was not that we should all stop using his divisive products, go throw smartphones into a pit, touch grass, and meet each other; it was that you should have 12 AI friends on his platforms.
Upvote
7
(7
/
0)
It's going to make the 2008 financial crisis look like being overdrawn on your current account.
It's a debt pile based on a service that doesn't turn a profit and still hasn't demonstrated fundamental utility.
The comparisons with Uber should come with a health warning: it took them 15 years to show profit on essentially running a taxi firm - scarcely reaching outer limits of high techology.
If nothing else you're shortly going to get data centre capacity for pennies on the dollar. The value of investments in cloud computing is going to nose dive when there'll be this much excess capacity.
It's a debt pile based on a service that doesn't turn a profit and still hasn't demonstrated fundamental utility.
The comparisons with Uber should come with a health warning: it took them 15 years to show profit on essentially running a taxi firm - scarcely reaching outer limits of high techology.
If nothing else you're shortly going to get data centre capacity for pennies on the dollar. The value of investments in cloud computing is going to nose dive when there'll be this much excess capacity.
Upvote
1
(2
/
-1)
ColdWetDog
Ars Legatus Legionis
Scamming people is a viable business model. You can rightfully say that a significant portion of our economy rests on this concept.
Upvote
3
(3
/
0)
Translation: “let’s subsidise the construction of this money hole by taking from taxpayers. Then when the bubble collapses we’ll double dip and make them pay for the cost of filling the hole in again.”
Upvote
1
(1
/
0)
Only if you've got a viable product that's just dying for access to high-end GPUs. Like others have said, buying up dark fiber for pennies was a useful outcome of the dot-com bubble. If you had the cash, buying distressed houses from a bank for cheap in 2009/2010 was a very small silver lining. But fiber and houses are useful things. It's not clear if gargantuan datacenters full of LLM-optimized GPUs are the same thing.
Maybe if the bubble genuinely pops after this capacity is built, Boeing and Airbus are going to have effectively-infinite CFD capacity to design their new planes. Maybe the PS7 and Xbox (random confusing suffix) will be virtual consoles capable of generating 8k visuals.
But I'm not in the industry and I don't know how adaptable all this hardware actually is.
Upvote
1
(1
/
0)
Not really, simply because BTC mining is all about hashes per joule (or hashes/sec per watt). This place will have a tons of watts coming in, but even state of the art GPUs don't do well on hashes/sec per watt compared to the dedicated ASICs that are used to actually mine BTC.
Upvote
2
(2
/
0)
Upvote
0
(0
/
0)
Upvote
0
(0
/
0)
At this scale, will there be any services or business models that can make huge systems like this possible?
Upvote
0
(0
/
0)
Sam Altman writing about this on his personal blog - "Maybe with 10 gigawatts of compute, AI can figure out how to cure cancer. Or [...] how to provide customized tutoring to every student on earth. If we are limited by compute, we’ll have to choose which one to prioritize; no one wants to make that choice, so let’s go build."
My jaw is just on the floor. I mean:
My jaw is just on the floor. I mean:
- This man basically has the audacity to say "if you oppose building this data center, you are forcing a choice between curing cancer vs. customized tutoring for every child on earth".
- Its a completely false choice because:
- ML for cancer research is not Transformer LLMs
- Tutoring solutions will be paid SaaS services (not free handouts), never reaching the global south
- In the global north those services will be the poor quality slop fed to students who cannot afford private schools with real human teachers
- AI usage results in worse student outcomes anyway, due to short-circuiting the key process of giving students guidance and then letting them learn by going through the exercises.
Upvote
9
(9
/
0)
As someone familiar with neural nets, I know your statements above to be true. However, I think they miss the point. What is Open AI building these DC's for? Are they offering compute while you BYONN? No - these are specifically to train, and run inference, on their LLMs.
Sure, they could in theory be repurposed if OpenAI goes belly-up. But they may be obsolete by then (i.e. their operational cost may not be worth the effort). That's what people are referring to, as the difference between fiber and GPUs.
Upvote
1
(1
/
0)
Friendly reminder -- every financial company offers a selection of funds.
If you have a 401k, chances are you are very heavily invested in AI, including companies like Facebook and Tesla, because it's the default option, but you can just move out of it.
Last edited:
Upvote
1
(1
/
0)
The data centers will be fine far into the future. Datacenters are not built out to only be used for 5 years and then burned down.
This likely would lead to a significant slow-down in new datacenters being opened for a very long time.
Upvote
-1
(0
/
-1)
My thoughts are 'instead of waiting for the bubble to burst how do we force the bubble to burst faster without triggering freedom of speech and freedom of commerce laws from a government policy perspective, or being technically liable for fraud from a civil perspective'
I'm aware if people can cause bubbles and bubble bursts you're officially in the realm of securities fraud and no one in the financial sectors should be allowed to make livable or investment values off short positions in a healthy economy, but being able to force bubbles to pop and financial sectors back to reality on the basis of tech over-hype and credit-leveraging without throwing people in prison would be a very handy trick for the middle class to avoid more 'credit-default-swaps' and exotic financial instruments to cover against lending risks, and one the billionaire classes and VC markets will be hell-bent on applying civil or criminal penalties to.
Upvote
-1
(0
/
-1)
"there is still actual unmet demand for generative AI capacity today"
Well...there is only demonstrated unmet demand for free generative AI capacity, I think.
Well...there is only demonstrated unmet demand for free generative AI capacity, I think.
Upvote
5
(5
/
0)
Those are two separate (though intertwined) issues. The first is about technology. In its infancy the computer was for hobbyists; IMHO, AI is likely to go the same way, just at a greater cost and complexity ... because it's expensive to run something orders of magnitude more complex than a computer. But the technical leap is undeniable and irresistable.
Then there's the question of business. These kinds of investment certainly involve risk, but big leaps usually do. We already see AI - new, somewhat inept, "Commodore-level" when we look back in 10-20 years - being integrated into everything. Don't think that's going back in Pandora's Box. It feels like a slow motion video of parkour.
Upvote
0
(0
/
0)
The buildings might be fine, but the buildings are only a tiny fraction of what the money is being spent on. It's the stuff inside the buildings that'll be obsolete.
On the other hand, I can imagine the largest ever Spirit Halloween now being setup in those abandoned data centres.
Upvote
2
(2
/
0)
All sectors will be hit when the bubble pops. NASDAQ fell hard in 2008, despite the issue being in the housing and financial sector; similarly DJIA and S&P 500 back in 2001/2002 with the doctom crash.
Upvote
0
(0
/
0)
All sectors were hit, but some sectors were hit harder.
The 2008 crash broader because most things depend on finance, but some sectors like real estate were especially wiped and others like retail and utilities did okay.
The dotcom bubble was more selective. It hit tech stocks way harder than everything else. And it was hard on the S&P in large part because the S&P had tech in it.
And it's the same way today. Most people are passively in the S&P. Which is "diversified". Except it's like 30% AI right now. If the AI bubble pops, the S&P will get slammed, and other ETFs with less Musk and Zuckerberg in them will get hit way less.
Pick a safer industry, like retail, or an ETF focused on Value or Dividends instead of Growth, and you won't be part of the bubble. You might be affected by it, sure, but that's not the same thing.
Last edited:
Upvote
1
(1
/
0)
Utilities were indeed one of the less hard-hit industries in 2008; they "only" declined in value by 30%, as opposed to over 50% for financials. That's still really bad.
Betting on what industry is going to make out better or worse is going to be pretty difficult. Anyone who's good at it is quietly already placing their bets, and they're playing with real money, not just a little 401(k).
Upvote
0
(0
/
0)
Yes, it is bad, though in both cases the market recovered eventually. If you can't afford to wait that long then you shouldn't be all in stocks.
It isn't that complicated. I can assure you AI stocks will do worse in an AI bubble. Yes you could also short AI or buy gold or something but simply not owning AI is a reasonable step.
Upvote
0
(0
/
0)
I guess when decision about building mega datacenters stopped being about "securing investment" and became "securing electrical supply".
Upvote
0
(0
/
0)