Five months later, Nvidia’s $100 billion OpenAI investment plan has fizzled out
- Thread starter JournalBot
- Start date
Also known as the MGKSA: the Musk Giant Kessler Syndrom Accellerator
Upvote
7
(9
/
-2)
I believe you are correct in the assumption about reusing starlink chassis. A good way to use a lot of power in space for compute is to have a lot of small units with a big shaded surface area. But I don’t either understand how it can be economically viable outside the hype growth of atocka. And it is for sure not an environmental solution.
Musk talking about hourly launches of starship is just crazy.
Musk talking about hourly launches of starship is just crazy.
I still don't think the financial calculus is going to actually work, but I can see the appeal of free real estate. But SpaceX already has a business that involves replacing satellites every 5 years, so. I think these'll start as a V3 starlink chasis, with the ground radio and antennas removed, and replaced by an equal power draw worth of compute power. Laser link to tap into the existing starlink down channel.
Upvote
7
(7
/
0)
The Meta Answer is that Nothing in this Timeline is supposed to make sense in the way we understood it in the Before Times.
Words like allies, truth, consequences, promises, freedom, peace, trust, terrorist, fair, agreement and color of the sky no longer mean anything in particular. You can use them as you wish
Upvote
23
(23
/
0)
I wish that if (I believe it's more of a "when" but I am giving it the benefit of a doubt) the AI Bubble pops and Nvidia has to pivot back to regular consumers (because the datacenter market has an oversupply) for their GPU revenue then the consumers will blue ball them by refusing to buy GPUs. Which will force Nvidia to lower prices and their stock price to take a dive. It's the only way consumer can give a FU to Nvidia I think.
Highly unlikely but one can dream.
Highly unlikely but one can dream.
Upvote
13
(13
/
0)
I have a plan for a 100GW of datacenters, gonna get the investors, then fill the buildings with heaters, and a single GPU, going to be such a powerful DC, cos look at that power draw. Just don't come anywhere close to it as you may set on fire.
Upvote
18
(18
/
0)
Upvote
18
(18
/
0)
It’s also a good way to assess GenAI data centres since no actual useful work is being done; we might as well measure their performance purely in terms of the waste heat warming the environment.
Upvote
19
(19
/
0)
Long after OpenAI is gone. Maybe even after the post mortem is finished. Although there is the possibility the post mortem on what went wrong at OpenAI will last longer than OpenAI did.
Upvote
7
(7
/
0)
The idea that the whole TulipAI hype may actually kill Oracle fills me with much more glee than it should. In reality this will probably not happen even though Oracle is pretty much overexposed to AI risks already but a man can dream though.
Upvote
21
(21
/
0)
There's no law in space to shut your child porn generators down!
Upvote
6
(8
/
-2)
I wouldn't be surprised if Sam Altman disappeared one day with all the money he’s raising. If Zelenskyy were that good at fundraising, he’d probably have defeated Russia by now!
Upvote
8
(8
/
0)
You'll probably like this article but Ed Zitron, even though it's a paywalled newsletter (no relation to the author although I subscribe)
Search for Ed Zitron haters-guide-oracle
Upvote
6
(6
/
0)
Over-valuation in 2001 was weighted towards tech, but not limited there. Much like today. We also had Enron, Sept 11, and other "unique" events spooking confidence. Much like today. Criticism of the Fed from the administration about its reluctance to move aggressively on rate cuts was rampant. Much like today.
To your point though, it was more a pullback on investment than a broader liquidity issue, and the details do matter.
Upvote
10
(10
/
0)
It's actually stupider than that, because it creates a perverse incentive for the supplier to make their products less efficient and for the service providers/data centre builders to buy less efficient hardware. It's almost as if the entities involved are interested in something other than actually delivering the goods and services they're telling everyone are "the future" profitably.
To be clear, it makes sense in certain contexts, like for relatively fixed and stable platforms such as conventional cloud computing, where you're mainly using it to do things like plan loads and consider things like incremental upgrades over time.
But in the context of, for example, the chips used in high performance compute for a supposedly emergent technology early in the S-curve? It's just asking for trouble. The GPUs at least (and the ASICs are probably in a similar situation) are being iterated on annual basis, what you get for your given 100MW isn't rally fixed which means in a lot of the contexts where they use it the watts are not really that interchangeable, but they're effectively comparing them like for like anyway (one datacentre running at, say, 400MW might be using GPUs from a couple of years ago while another 400MW DC is using the latest - their functional capabilities are not the same but if all you talk about is the IT load they look equivalent).
It's like a lot of things with these GenAI companies, they take something that isn't necessarily problematic and quite standard but use it in ways that are inappropriate to induce the irrational exuberance which is at the heart of how they're operating. They do the same thing with financials like annualised revenue. These are legitimate metrics which can tell you things which are useful, but they use them misleading to imply other things.
This is one reason why the Deepseek moment scared them all so much. It threatened to expose that the entire "hyperscaler/more dakka" approach taken by the primarily US-led wing of the GenAI industry is fundamentally wrong and destined to fail. But "luckily" the greed combined with latent Sinophobia and privacy/security concerns (though arguably this is also Sinophobia in as much as none of the Western foundational model companies is particularly trustworthy in that regard either - OpenAI is literally run by a guy simultaneously running an AI/biometrics tinged crypto grift) quickly re-clouded all the money people's judgment.
Google is their main competitor and they've got the advantage of being an actual going concern with other businesses to underwrite the loss-making inference services for some time.
OpenAI also competes with Microsoft who are admittedly nowhere close to Google's level in this field but also benefit from being an established company with other cash cows to keep them afloat in all this. And who, oh yeah, own a huge chunk of OpenAI and have access to all their relevant IP, leaving them perfectly primed to simply absorb the shattered remnants of clammy Sammy's outfit when it collapses.
OpenAI is vulnerable for those reasons and others. Anthropic is vulnerable for some of the same reasons as OpenAI - the most fundamental one being that foundational models are relatively fungible. These companies have no moat. Your Microsoft and your Google have ecosystems they can wrap the models up in to sell as products. The models are just components. Businesses can go shopping for them. Apple just did, for example. They licensed Google's. Most users don't care about models, so once models exist that platform holders can use, for the most part people will just use the one their chosen platform has if they want to use GenAI. So if basically all your company does is supply foundational models and foundational models with wrappers around them?
You're adrift at sea on a plank of wood trying to race against ocean liners.
Upvote
21
(22
/
-1)
MilanKraft
Ars Tribunus Angusticlavius
You're not wrong, but one key difference between this scenario and the usual "perception is reality" attitude that many marketing dickheads have, is that Ars has reported and helped to explain a core truth about the LLM frenzy — that it is mostly smoke and mirrors, including bullshit deals like this one. (Pump and dump should absolutely be everyone's #1 suspicion.)
Whereas with Marketing Brain, many of them literally believe it's OK to lie about a product or service as long as the customer perception created by that lie, drives sales for a while before customers begin to figure out the truth and the cascade of bad reviews kick in. At which point you pivot / apologize / PR your way out of it and then start a new cycle of half-truths and lies to "serve the customers." The worst part is they will say shit like this openly in some cases, like "hey it's just the world, man.... stop acting like I personally am doing something wrong."
In my experience, Marketers and Advertisers are only slightly more honest and respectable as people than politicians. Always exceptions of course, but on balance the entire "profession" is a festering wound on the ass of humanity. I'm all for driving negative perception here — by almost any means necessary — but it would be great one day if the "perception is reality" attitude becomes a relic and symbol of everything that is wrong with how companies operated in the first half of the 21st century.
Last edited:
Upvote
27
(27
/
0)
I do not believe I am the only one to have a visceral, immediate, negative impression of anyone self-describing as an "Influencer."
[turns head and spits]
Upvote
21
(21
/
0)
Yes, but the first company to break the chain wins 100 billion dollars.
Upvote
8
(8
/
0)
nvidia is probably realizing there's more gold to be made selling to everybody else than dumping it into altman's pocket.
Upvote
5
(5
/
0)
Good point, if the seal cracks it won't matter.
But I specified 6L6 because I had practicality in mind1 and the steel envelope would provide protection against straight photoelectric effect as well as charged particles from solar storms.
1Not really.
Upvote
8
(9
/
-1)
Bit small. I would be looking at things with a bit more ceramic. A Ctk 25-4 perhaps.
You could even probably design an autoloader to replace them remotely when they burn out.
Upvote
1
(2
/
-1)
RickRoyLeonPrisZhoraRacha
Ars Scholae Palatinae
Whoa. Slow down. Let's make some money on it, then make it on its slow death or absorption.
/s
However, the truth of it all lies with TSMC. You see, the biggest stick isn't nVidia in this AI game, its the chip maker. nVidia isn't truthful about scaling and its power needs (aka 4x the consumption of normal cloud server centers). And that is getting out.
Upvote
3
(3
/
0)
Sadly, intelligence has nothing to do with it.
A lot of people makes a lot of money in a stock market bubble so most of them goes along with it even if they think it's stupid.
Kind of the same thing with Donald Trump. I imagine that most CEO's and people in the administration think his the worst president ever. However they make a lot of money so they don't object.
Upvote
16
(16
/
0)
Nvidia recently released their new chip less than 2 years after the previous version. Meanwhile AI companies are increasing the depreciation timelines on their hardware to 5+ years. Seems to me somebody is trying to prop up earnings anyway possible. There will be a crash but it’s a little hard to see how it will affect the broader economy.
Upvote
7
(7
/
0)
Wall Street is a casino now. Literally a gambling product for too many investors since COVID.
I remember when Tesla's valuation was based on being the biggest automaker and replacing all ICE cars and now lolwut? There might be stock shenanigans and foreign intrigue with Elon's companies, but there's clearly millions of investors who have no idea how an economy works, or what they're investing in. They're extremely online and driven by memes and hooked on feelings more than results, like the shortsighted gamblers that they actually are.
Upvote
15
(15
/
0)
$100 billion the same as the GDP of the following countries.
Luxembourg a high development country of 700,000 people.
Serbia a middle income country of 6.5 million people.
Côte d'Ivoire a low income country of 33.5 million people.
So from a small US city to a population greater than any US state apart from California.
Luxembourg a high development country of 700,000 people.
Serbia a middle income country of 6.5 million people.
Côte d'Ivoire a low income country of 33.5 million people.
So from a small US city to a population greater than any US state apart from California.
Upvote
10
(10
/
0)
His claim isn't as much about launching space data centers as it is about pumping the upcoming IPO for SpaceX.
Elon lies. Minions buy.
Upvote
14
(14
/
0)
The rare plants market has recently been demolished by a YouTuber who started a series of videos about how to clone plants in a home lab with a couple hundred bucks of supplies.
Upvote
5
(5
/
0)
If "invest" means getting paid in shares, and OpenAI was thinking chips at September-prices while Nvidia expects current prices, a few misunderstandings are quite possible even if both sides say the same words.
Upvote
3
(3
/
0)
I've heard some fairly credible reports that these super-expensive, high cost chips from Nvidia are having reliability and durability issues (i.e. a not-insignificant number of those chips just aren't going to last 2 years, let alone 5+ years--if they even work out of the box). I can't imagine that's improving the real numbers any (which of course we probably won't see until after the bubble pops).
Upvote
10
(10
/
0)
Certainly. I didn't mean to imply one specific company would have a bad outcome. The whole damned economy and the whole universe of finance is wrapping itself around this crap. That's the story, not the earnings line of one company.
Upvote
2
(2
/
0)
Upvote
12
(12
/
0)
The dotcom bubble burst a year and a half before 9/11. By early 2001, traffic on the 101 was light since so many startups had gone bust. That’s before even the Enron energy crisis, which was Enron’s death rattle.
9/11 merely extended a downturn that was already well under way.
Upvote
11
(11
/
0)
Upvote
2
(2
/
0)
And how much "compute power" would that be, exactly?When I ran physical servers, I rented racks in a data center. That included space, power, cooling, and 100Mbps. I had no reason to deprecate old hardware, because I was paying for all of that, whether I used it or not. If I had free space in a rack, keep the old hardware running. I had 12 year old 1U servers doing video transcoding, so I could keep the faster machines free for something that was more interactive.
Cloud providers don't have that same math. Space, power, and cooling are at a premium. If there is obsolete hardware that's generating little to no income, it's not worth the power and cooling it takes to keep it online. As soon as a replacement will earn more money, it gets replaced. The calculus is a little complex, but not very.
Once in space, the only cost is internet bandwidth and the salary of the people managing them. It's virtually free to keep them running until they deorbit or burn out. And you don't have to replace an old one, you can launch a new one next to it.
I still don't think the financial calculus is going to actually work, but I can see the appeal of free real estate. But SpaceX already has a business that involves replacing satellites every 5 years, so
Upvote
2
(2
/
0)
„I accidentally crashed the rare plants market“
View: https://youtu.be/7dk3uSLxMI4?si=FWe3aUPU2GHIN8Qg
Upvote
8
(8
/
0)