Does the math check out, though? For reference, Nvidia added $17B of non-marketable securities in 2025, and its total revenue was $215B. So, assuming all these equity investments went to GPU purchases, they only amount to 8% of its annual revenue.
Since these firms (neoclouds and AI startups) will get consolidated if Nvidia puts too much money into them, I assume the majority of their funds are coming from elsewhere.
GPU scarcity is a really poor proxy for the health of the AI market. I don’t know if I should be surprised that someone so bullish about it doesn’t understand this.
Why is Nvidia financing neoclouds to buy Nvidia chips to rent to AI customers? At least it was my understanding that run their own DCs for GeForce Now for example
We did but people simply keep pretending it’s not happening.
Between the circular financing and behaviour some analysts call “channel-stuffing”, it does seem that with Jensen Huang gone, the music is perhaps about to stop.
No one seriously denies the actual activity was happening, they just dispute its implications, and for good reason. Vendors frequently extend financing with the hope of making more sales, the most obvious example being auto loans and BNPL
Vendor financing is already a sign of an unhealthy market. When widespread, almost always. If the only easy way you can borrow enough to buy a thing is from the vendor, that’s bad. And if vendor financing is an increasing, not decreasing, proportion of the finance in a market, that’s worse.
Channel-stuffing, if it is happening, is usually plain illegal.
The implications are obvious: if people are using GPU scarcity and prices as a proxy for AI demand (and they absolutely have and still are, at every level), and use it as an indication that they must secure GPUs while they can, then it matters a lot if in fact that demand is being padded out and front-loaded in various ways.
>Vendor financing is already a sign of an unhealthy market. When widespread, almost always. If the only easy way you can borrow enough to buy a thing is from the vendor, that’s bad. And if vendor financing is an increasing, not decreasing, proportion of the finance in a market, that’s worse.
I just am unable to understand whether it's a bad thing or not, you could frame it as being circular, but you could also see it as Nvidia simply giving hardware in exchange for startup equity, is that really fundamentally problematic, is it the scale at which it was done, or is it a valid but risky tactic?
As a result they never seriously pursued frontier research for open models - which could have easily absorbed any excess compute they planned for.
On the other side their customers are rapidly becoming their competitors and will happily eat up spare capacity at TSMC if nvidia can’t fulfill its obligations.
Edit: to be fair there are several counterfactuals that could have played out negatively if nvidia decided to go into frontier research or go upstream like Elon and build their own fab. Jensen himself said they were completely focused on building general purpose compute and his bets had turned out to be right so far.
To be fair their Nemotron models are pretty good. The Parakeet speech-to-text model is also pretty good. That being said these pale in comparison to the frontier labs. They should've invested more into it, but I understand that they don't want to piss off their biggest competitors.
Worse, I think companies that warehouse gpu's will buy newer more efficient gpu's by the time they get their data centers running but won't resell the older but unused gpu's afraid to let their competitors to make inference.
Isn't that straightforwardly explained by typical hoarding behavior, not " afraid to let their competitors to make inference"? During the pandemic people were hoarding TP that they couldn't possibly use within a reasonable amount of time, but it'd be crazy to accuse people of doing that because they were afraid of their neighbors having clean asses.
Even if NVidia crashed and there ends up being an oversupply of GPUs or TSMC fab slots, this doesn't change that OpenAi and Google still need memory for their own hardware accelerators, so the price of boards wouldn't go down by much
Part of me wonders (maybe naively) if with falling cost of AI inference and improvement of models means there is an oversupply of inference infrastructure. since we may reach a point where no one needs more and everything you could want can be done by the machine and adding more compute for most users is not any kind of benefit (like it was before LLMs there was an obvious ceiling where the majority of people did not need more compute).
I know Jevons paradox is a thing and I could see it going the other way as well but the first possibility seems to be ignored and I am not sure we can. Eventually demand for training might dry up as well if are models become largely very good and only a bit of fine tuning is required for most applications.
I don't see how it helps to think so narrowly about the computation needed solely for training when we are clearly headed to a future where tons of realtime applications (think robots) need to do insane amounts of processing ASAP while doing time-sensitive tasks.
IMO its copium to think that chips which have 10x more demand than supply are suddenly going to be easy to get as "AI is eating the world" like software did over the past decades.
The author is a developer working on software that scales to 1,024 cores and greater -- which is within spitting distance now thanks to 256-core AMD EPYC processors. If you count hyperthreads as a "virtual core" then a dual-socket system gets you there!
That blog made me realise that almost all software can't utilise 1,024 cores, not even close! Microsoft SQL Server, as a random example, tops out at 64 cores for any single query.
Similarly, people couldn't utilise a machine that big, not with pre-AI software, except for certain embarrassingly parallel codes such as simulations and 3D graphics rendering.
Now... NOW... there's no upper limit. None. One person(!) can cheerfully utilise millions of cores, billions even, and there's no end in sight.
If you think one conversation thread in "ChatGPT" is all we need, you're not seeing the big picture of what's possible here.
Those math proofs OpenAI produced were created with tens of thousands of agents running in parallel, orchestrated by thousands more agents, and overseen by hundreds of agents, all triggered by someone typing as little as a sentence or two to start that avalanche of compute!
With that kind of AI power at your fingertips, huge, unthinkable projects become a fun exercise for the weekend:
- Translate the entirety of Wikipedia to every human language.
- Rewrite the entire Linux kernel to Rust in one go
Same. I replaced the pads and paste on my now ancient RX 580 in the hopes of surviving to the next time prices are affordable.
For anyone wanting to do so - you can! You have to make sure you nail the pad thickness though. That can be tough because it can have changed between revisions of the same card. Thankfully, AI is pretty good at combing through past data to figure out what you need. Also it voids your warranty.
Surely this is Nvidia deploying excess capital by acquiring equity in high growth companies at significant discounts (factor in that a major portion of the investment comes back to them against their high margin products). This isn't a new business concept and is generally accepted as smart business? Not some breaking news conspiracy?
> (factor in that a major portion of the investment comes back to them against their high margin products)
One problem is that there isn't one number that indicates what percentage of Nvidia's equity investments come back to it in the form of chip purchases. In some cases, it appears it's 100%; in others, it has been suggested that every $1 invested produces >$3 in revenue. $1 invested for $1 in revenue versus $1 invested for $3 in revenue are really different outcomes.
There are also the questions of the revenue and revenue quality of the companies Nvidia has invested in, how much debt they've taken on, what sort of backstops/guarantees Nvidia has given them, etc.
From WSJ below. Seems clear they went too far. But not clear Kress can navigate this.
Over the course of the meeting, he proceeded to tear into the woman who is now taking over for him as Nvidia CEO.
Colette Kress was "a deer in the headlights," he said, interrupting her presentation, according to a person familiar with the comments.
He had summoned Kress and his senior finance team to diagnose a ballooning problem: Nvidia was in talks to provide billions of dollars of guarantees to companies that couldn't afford its chips. That included deals led by Huang's son-in-law, Nico Caprez, who had been offering upstart cloud providers and data-center developers financing in the hopes of turning them into long-term customers.
Huang had launched a dealmaking spree and was now starting to question whether it had gone too far, according to people familiar with the matter.
Kress was a "reporting CFO," not someone who could help him define a strategy or make decisions, Huang said during the August meeting.
You’ve now edited the comment I replied to (originally it was just the quote), but it is still not responsive to the OP; nor is your new comment responsive to my earlier criticism.
Well that's actually a hopeful sign for the future. I was a bit worried hearing she'd been conservative about investments in the past... but if the conservatism was about vaguely fraudulent circular financing deals we'll all be better off.
What scares me is that anybody could believe this garbage. It costs me $1 to make, I sell it for $5. Good deal! In fact, such a fabulous deal, I'm going to do this: Make 2 for $2, sell one for $5, then loan these other dudes the $3 I made in profit plus go $2 in the whole for a total of $5 so that they can give me the $5 back.
The more often you do it, and the greater the margin becomes due to squeeze, the dumber it is.
I mean, preventing bubbles is literally the reason short sellers are allowed to exist. When they see a stock that is overpriced, they help it correct. Innocent people lost a lot of money on Enron.
Depends, you buy a used server off of eBay or 3d print your own fan shroud. Waiting for prices to drop some and then I’m looking to build my own rack in the basement, so I can use open weight models and not directly support horrible data center practices.
Not for gaming they don't have monitor outputs and most of them aren't even in a standard PCI-E format. Some of them might be useful as a home AI accelerator with adapters and jerry rigged cooling.
Some of them, sure. You can invest in a server mb, and put a rack in your shed or basement. If someone was selling B100 like that, I'd happily buy a few for 15k. I somehow doubt its going to happen though.
I am not sure how inference constrained we are, btw. As an enterprise customer, I have zero problems with 429s and capacity from google or aws. All large volume inference jobs - so customer journey bots that help our company make money - run on dirt cheap models kept on short leashes in custom harnesses. Message delivery costs us 10x more than inference (with meta price hikes of October). All long-horizon jobs such as coding are probably somewhat inference constrained, but only to the point that sometimes the devs and corporate users with their own agents wait a bit for something to be completed, but nobody is noticing it. Sure, one notices a difference between /fast mode and normal while building software, but nobody really has a reason to care much, /fast modes are a pure luxury thing outside of incident remediation. I don't see what would Anthropic actually do if they had 10x the compute or 100x the compute they are planning for. Surely, some new business models, trolling mathematicians at a larger scale maybe?
https://www.theguardian.com/technology/2025/dec/28/nvidia-in...
Since these firms (neoclouds and AI startups) will get consolidated if Nvidia puts too much money into them, I assume the majority of their funds are coming from elsewhere.
Between the circular financing and behaviour some analysts call “channel-stuffing”, it does seem that with Jensen Huang gone, the music is perhaps about to stop.
Channel-stuffing, if it is happening, is usually plain illegal.
The implications are obvious: if people are using GPU scarcity and prices as a proxy for AI demand (and they absolutely have and still are, at every level), and use it as an indication that they must secure GPUs while they can, then it matters a lot if in fact that demand is being padded out and front-loaded in various ways.
How much % of car loans are vendor financed?
As a result they never seriously pursued frontier research for open models - which could have easily absorbed any excess compute they planned for.
On the other side their customers are rapidly becoming their competitors and will happily eat up spare capacity at TSMC if nvidia can’t fulfill its obligations.
Edit: to be fair there are several counterfactuals that could have played out negatively if nvidia decided to go into frontier research or go upstream like Elon and build their own fab. Jensen himself said they were completely focused on building general purpose compute and his bets had turned out to be right so far.
:D
"Banks allegedly engineered inflation by funding[1] consumer's spending"
[1] aka. credit cards
I know Jevons paradox is a thing and I could see it going the other way as well but the first possibility seems to be ignored and I am not sure we can. Eventually demand for training might dry up as well if are models become largely very good and only a bit of fine tuning is required for most applications.
IMO its copium to think that chips which have 10x more demand than supply are suddenly going to be easy to get as "AI is eating the world" like software did over the past decades.
The author is a developer working on software that scales to 1,024 cores and greater -- which is within spitting distance now thanks to 256-core AMD EPYC processors. If you count hyperthreads as a "virtual core" then a dual-socket system gets you there!
That blog made me realise that almost all software can't utilise 1,024 cores, not even close! Microsoft SQL Server, as a random example, tops out at 64 cores for any single query.
Similarly, people couldn't utilise a machine that big, not with pre-AI software, except for certain embarrassingly parallel codes such as simulations and 3D graphics rendering.
Now... NOW... there's no upper limit. None. One person(!) can cheerfully utilise millions of cores, billions even, and there's no end in sight.
If you think one conversation thread in "ChatGPT" is all we need, you're not seeing the big picture of what's possible here.
Those math proofs OpenAI produced were created with tens of thousands of agents running in parallel, orchestrated by thousands more agents, and overseen by hundreds of agents, all triggered by someone typing as little as a sentence or two to start that avalanche of compute!
With that kind of AI power at your fingertips, huge, unthinkable projects become a fun exercise for the weekend:
- Translate the entirety of Wikipedia to every human language.
- Rewrite the entire Linux kernel to Rust in one go
- Rewrite all open source C code to Rust.
- Formalise every maths paper ever written.
Etc, etc, ... without end, without bound.
Similar to the fact that supply of consumer RAM is currently low.
Plus some people are buying consumer gear and trying to use it commercially.
For anyone wanting to do so - you can! You have to make sure you nail the pad thickness though. That can be tough because it can have changed between revisions of the same card. Thankfully, AI is pretty good at combing through past data to figure out what you need. Also it voids your warranty.
One problem is that there isn't one number that indicates what percentage of Nvidia's equity investments come back to it in the form of chip purchases. In some cases, it appears it's 100%; in others, it has been suggested that every $1 invested produces >$3 in revenue. $1 invested for $1 in revenue versus $1 invested for $3 in revenue are really different outcomes.
There are also the questions of the revenue and revenue quality of the companies Nvidia has invested in, how much debt they've taken on, what sort of backstops/guarantees Nvidia has given them, etc.
Reducing circular financing agreements seems like ... a good idea?
> in response to which Huang allegedly lay into Kress, calling her a "deer in the headlights,"
Over the course of the meeting, he proceeded to tear into the woman who is now taking over for him as Nvidia CEO.
Colette Kress was "a deer in the headlights," he said, interrupting her presentation, according to a person familiar with the comments.
He had summoned Kress and his senior finance team to diagnose a ballooning problem: Nvidia was in talks to provide billions of dollars of guarantees to companies that couldn't afford its chips. That included deals led by Huang's son-in-law, Nico Caprez, who had been offering upstart cloud providers and data-center developers financing in the hopes of turning them into long-term customers.
Huang had launched a dealmaking spree and was now starting to question whether it had gone too far, according to people familiar with the matter.
Kress was a "reporting CFO," not someone who could help him define a strategy or make decisions, Huang said during the August meeting.
I’m so ready this time, CFO at the helm
1. circular financing agreements
2. see 1
I also feel this is a good thing. Looks like the elephants in the room are being addressed.
The more often you do it, and the greater the margin becomes due to squeeze, the dumber it is.
Here’s the original article: https://www.wsj.com/tech/jensen-huang-colette-kress-nvidia-c...?
I imagine the shortsellers and put option holders like that.
Seems like a great way to cash out at the top
I am not sure how inference constrained we are, btw. As an enterprise customer, I have zero problems with 429s and capacity from google or aws. All large volume inference jobs - so customer journey bots that help our company make money - run on dirt cheap models kept on short leashes in custom harnesses. Message delivery costs us 10x more than inference (with meta price hikes of October). All long-horizon jobs such as coding are probably somewhat inference constrained, but only to the point that sometimes the devs and corporate users with their own agents wait a bit for something to be completed, but nobody is noticing it. Sure, one notices a difference between /fast mode and normal while building software, but nobody really has a reason to care much, /fast modes are a pure luxury thing outside of incident remediation. I don't see what would Anthropic actually do if they had 10x the compute or 100x the compute they are planning for. Surely, some new business models, trolling mathematicians at a larger scale maybe?