FR version is available. Content is displayed in original English for accuracy.
Advertisement
Advertisement
⚡ Community Insights
Discussion Sentiment
40% Positive
Analyzed from 2046 words in the discussion.
Trending Topics
#nvidia#more#money#circular#don#gpus#prices#big#bubble#deal
Discussion (131 Comments)Read Original on HackerNews
Nvidia invested $2b into CoreWeave for 9% equity stake. CoreWeave is spending $35b in CapEx in 2026. Therefore, Nvidia's investment is only 5.7% of CoreWeave's single year CapEx. The other $32b is coming from other sources that isn't Nvidia. This is hardly circular.
Nvidia invests in Neoclouds because it's a hedge against hyperscalers having too much power, ie designing and prioritizing their own chips, and not fully using Nvidia's rack design. Neoclouds give hyperscalers competition. Neoclouds accept Nvidia investments because it allows them to secure Nvidia chips first, which is a competitive advantage since new Nvidia chips have been as much as ~5-20x more efficient than old Nvidia chips.
Nvidia was planning to directly compete against hyperscalers through DGX Cloud. They cancelled public DGX Cloud access when they found that investing in Neoclouds would accomplish the same goals without having to compete against their biggest customers.
If you're Nvidia, it's smart because Neoclouds that you have a large stake in will deploy your full stack from GPUs to networking to storage racks. They will share valuable usage data back to you so you can design a better next generation. Hyperscalers are likely a lot less cooperative, prefer to use their own designs if possible, and will guard their usage data.
Depends if they actually got the $2b in real money. There's a difference.
It's a big deal if no money was involved. Nothing even entered the company directly. Some deals have structured with Special Purpose Vehicles where money goes to the SPV. The SPV buys GPUs with it (from Nvidia). GPUs is loaned back to the company involved. So this company is stuck with this GPU rental, which may or may not be what they want and not $2b.
This sounds like a bad deal? So Nvidia had to sweeten the deal and promise min utilization on those GPUs by renting it themselves even if they don't need it.
So what's income and what's expense here?
That's the problem. It's inflated and messed up.
One aspect of the profitability might be the utilization and the pricing a few years down the line for slightly older hardware. Already now it seems like the increased processing you get from newer devices versus the cost difference makes something like an H100 or even A100 significantly less desirable than newer more powerful ones. As an individual, I am happy to be able to get an H200 on demand, but the B200 or B300 can do so much more work with optimized software and models for only modestly more cost that if those become available then from a business perspective you really have to prefer that if you can keep it occupied.
Then with Vera Rubin being like 3 times more effective or whatever, that adds a new layer of gradual obsolescence. So the question is can they keep the pricing up on the older ones a few years down the line enough to fill out the end of those expected payback periods.
The real boogeyman for a neocloud that has heavily invested in expensive Nvidia hardware might be a variation of that beyond Nvidia with startups that have even more dramatic efficiency increases pushing the leading edge even further. For example, if companies like Mythic AI and d-Matrix could somehow rapidly rapidly scale, that would push prices down for all of Nvidia hardware that is significantly less efficient.
I guess so far it doesn't look like any startups with really big efficiency breakthroughs are even close to being able to scale like Nvidia though, especially with the manufacturing and power crunch. But I suspect some of that is because of favoritism and strong arming protecting investments rather than a free and fair ecosystem.
They don't expect to keep the prices flat over time, and everyone involved will have planned for this. Prices are highest when they're the newest and greatest (part of why it's valuable for neoclouds to be first in line for new models), and drop year by year as newer GPU models can do equivalent work at lower cost.
You can see a pretty cool dataset of this at [1]; H100 prices where $3/hr in 2023, and dropped linear-ish to $1.75/hr by 2025. And also the notable exception that prices are up this year due to shortage.
[1] https://semianalysis.com/gpu-pricing-index/
Micheal berry doesn’t know shit about GPU pricing or depreciation schedules. A100 demand is very high and easily 2 dollars an hour for reserved right now.
B200 and Vera reubin don’t help much if you don’t benefit from quantization, and that’s exactly my situation and many other AI research orgs situation.
A100s are going to continue making money per hour until 2030. Mark my words.
CoreWeave feels very YC-ish. I thought I had an in as a referral for a position there and got interviewed by someone who knew a lot of my peers where I worked. Dude seemed to ask very textbook style questions that you would only learn if you went to a school system for this particular position/subject. I guess I didn't answer to their satisfaction despite knowing more than them on almost everything else. I suppose I'm still bitter seeing as I interviewed with them three times for two different roles. Absolutely wild.
FWIW, I’ve referred someone 3 times to the same position because I’m very sure he would be a good fit, and I’ve seen his work.
But for trivial reasons (“He doesn’t seem enthusiastic enough” and “the other candidates are better at promoting/selling themselves”) , a couple of managers that are above me in seniority (and directly in the hiring loop) just refuse to pass said person.
In the end he’s stopped applying, and I feel shitty for referring him.
That puts a cap on surplus (potentially unused?) datacenter capacity that's around by the time the AI bubble pops.
Any surplus after a pop will be sold for market value and lead to more new cloud provider startups and co-location options.
[0]: https://en.wikipedia.org/wiki/California_gold_rush
Manyfacturers aren't artificially restricting supply, they're running fabs full-tilt. You could want them to build more fabs to meet demand. Which they are, but at a more modest rate than what you would want, because those manufacturers have been burned in previous boom-bust cycles. Never mind that fab-construction lead times are measured in years.
And what's stopping you from fabricating & selling RAM? I've read it's very profitable! Oh yeah, it takes many $B to pull a SOTA fab out of the ground.
Vendors price-gouging? Probably. Wouldn't you?
TLDR; it's not a monopoly issue. This is a high-tech specialized market where a ridiculous spike in demand is near-impossible to cater for. You want some new RAM-heavy gadgets? Shell out $, adjust your RAM 'wants', or be patient.
It is. The GPUs go on to be used to get loans to then get more GPUs.
Now I've got the feeling they don't have huge amounts of GPUs sitting in their DCs, but rented for Opex. In case the bubble pops they might get it at discount as CapEx (like Amazon did with dark fiber after the dotcom bubble).
Nobody lives in GPUs and what was the ratio of equity/debt for the toxic assets in 2007?
Yup. Add to that the decade worth of ZIRP following the 2007ff crash and Covid... all that money has to exit the system again eventually.
https://www.currentmarketvaluation.com/models/s&p500-mean-re...
Financing is circular because creating a liability for one party (debt) creates an asset for another (the bank) off of which more debt can be secured
A bank / financier sells trust and reassurance. They otherwise invent most money from thin air.