Nvidia Announces Financial Results for Fourth Quarter and Fiscal 2024
nvidianews.nvidia.com
nvidianews.nvidia.com
What ever happened?
It shows that the most important move is to simply stay in the game, not just maximize profits
Edit: added ATI which I just remembered!
3dfx stagnated and fell behind due to bad management decisions. They seemed stuck on designing 3d accelerators like it was still 1996, which were just drawing triangles and doing texture mapping, while banking on higher clocks and adding multiple chips on board and even multiple cards in SLI to boost performance, while having the CPU take care of the rest in SW.
But it was 1999 already and Nvidia moved to graphics processing units (GPUs) which could also offload the CPU and do transform, lighting, triangle setup/clipping, MPEG2 motion compensation and a multi-pipeline rendering engine for more polygon thruput at same clocks, along with being fully DX7 compliant, all in a single chip.
The market decided that Nvidia's new direction is the future and not what 3dfx was doing. 3df tried to change course and catch up but it was too late, the GPU race was too cutthroat in those days. Nvidia was launching a new GPU generation every 6 months(!) instead of every 3 years like today, so a 6-12 month setback was a guaranteed death sentence no matter what you did.
Also, one of 3dfx biggest blunders was trying to become a board maker themselves by spending a lot of cash buying a board manufacturer, which emptied their bank accounts and angered their other board partners who now saw them as unfair competitors so they moved to making ATI/Nvidia boards instead.
>It shows that the most important move is to simply stay in the game, not just maximize profits
How do you stay in the game if you go bankrupt because your products aren't competitive?
https://www.acquired.fm/episodes/nvidia-the-gpu-company-1993...
https://www.acquired.fm/episodes/nvidia-the-machine-learning...
CUDA itself is also a non-negligible factor in their success.
If you look at the same landscape from AMD's side, it's lackluster to say the least. And they're only now trying to catch up... And they have to make it CUDA compatible!
Nvidia won on the tools side, the GPUs aren't necessarily faster, but the only way to use them easily is CUDA.
Cg made programmability of shaders widely accessible, since C was (20 years ago) widely known?
For 3dfx, the Voodoo 3 happened.
Whereas previous Voodoo cards were sold by a large variety of OEMs (like Diamond or ELSA), 3dfx decided to make the Voodoo 3 graphic card on their own. It was a perfectly fast and capable card, but also expensive (for its time), not as easy to get as OEM cards were, couldn't do 32-Bit textures (16-Bit only), and didn't support AGP Texturing.
nVidia's TNT2 was a bit slower, but not by much, and thanks to the many OEMs making cards, you could pick one up anywhere and usually pretty cheap. (Just make sure you're not getting the M64 by accident. But even then, that one might've been a good way to get a capable 3D graphics card for cheap).
3dfx was still competitive here and probably the best gaming card, but they had clearly bitten off more than they could chew, and it looks to me that their R&D just didn't produce anything else than "Voodoo 2 but a bit faster and delayed multiple times".
When nVidia released the Geforce 256 (Hardware Transform and Lighting!), that was it. Both the Voodoo 4 and 5 were duds (though seeing 3dfx's complete brute force approach with the never-released Voodoo 5 6000 still brings a smile to my face), they couldn't get other products to the market quick enough, and went bankrupt starting in mid 2000.
I don't even think it was about "GPU" stuff: The first programmable shader card (and thus "GPU") was the Geforce 3 in 2001, at which point 3dfx was already out of money and without perspective.
I'm not sure if they could have been saved if Sega would have chosen them for the Dreamcast (seeing how the Dreamcast sunk Sega's console business) and it might have been possible to save 3dfx still, but realistically, the one-two punch of the Voodoo 3 and Geforce 256 killed them.
That's how my parents got scammed when I asked them for a PC for Christmas. Yes, I got one with a TNT2 M64 because that was on sales in Christmas 2000/2001. Childhood ruined.
Not clear if Meta is included in the "cloud providers", if not that would push the share of revenue from Big Tech even further to 60-70% considering the 300k H100 order from Meta.
Should I be scared of the bubble-ness of Nvidia? I have a lot of faith in the company and its vision, but reading about the bubble always scares me.
That being said, Nvidia will have a good 2-3 years in the future, until the cloud providers start mass replace their chips with in-house ones.
For money, unfortunately, it is as good as anyone's guess, but if you join now, double your share, it will be close to 4T. I personally didn't think that is sustainable amount of money if market/competition does their duty.
How do you imagine that happening on such a rapid timeline?
IMO, the biggest threat to Nvidia is radical innovation on the software side of AI that allows common use cases (training and inference) to run on clusters of commodity non-GPU hardware for cheaper price than GPU.
A similar threat is posed from ASIC chips that are dedicated to LLM tasks, but at least in that case Nvidia can still compete for a specific hardware design rather than overcoming an entire category of commodity hardware.
https://qz.com/intel-microsoft-new-chip-semiconductor-nvidia...
NVDA's margin is too high, it attracts competitors to eat into that.
Second, there is no need to be market leading, it just needs to be cost efficiently to be successful, even if it is slower, as long as it is cheaper. All cloud providers have huge chip design/manufacturing history, AWS/Azure and Google.
I would argue it IS indeed the core competency for those cloud providers, because of price. The inhouse chips exist because they can reduce the cost.
I know for a fact, there is huge momentum in some big clouds to replace Nvidia GPUs with their own chips, at least for their internal compute needs at first.
Also, if GPUs are easier to make than CPUs, then why isn't Intel leading the market in GPUs by now? Surely it's been clear they should compete with Nvidia there for at least the past half a decade. Or do they need more than 2-3 years to catch up?
I'm not disagreeing they can catch up within 10 years. But 2-3 years is an insane estimate IMO.
The beauty of being fabless is that they dont need to worry about that. When theyve got a design they like they just call TSMC and buy out all their capacity, they could pay double what nvidia does and still save a boatload. I agree that 2 years seems impossible and 3 a stretch, but I do think its possible for at least one cloud company to have something by then.
But also consider the reason Nvidia is doing so well in the first place. The reason is because they were already developing hardware that specialized at performing tensor operations. Maybe not at the size or always the same types we see in ML, but not very far either. There's reasons why they aren't filling cards with tensor cores. But they are clearly developing specialized hardware for ML tasks. I'm not sure betting on LLMs sticking around is the best bet though. I'll be very impressed if scale is all you need.
If scale is all you need, then we'll get there. But I'm not remotely convinced and see this as the bitter lesson's bitter lesson (a misinterpretation of the bitter lesson). But if we fund alternative pathways and hedge our bets, it's possible we get there without missing a step. But imo it looks like we've created a railroad and are just going all in on laying more tracks. Doesn't seem like the right move to me, but hey, I don't know the future any more than anyone else.
The share price of nvidia will collapse as soon as anyone else releases a competitive datacenter gpu. End of the day GPU compute is fully commoditized. Massive margins in commodity markets dont exist because competition leads to a race to the bottom in pricing. Hard for me to imagine that doesnt happen eventually, but wouldnt be surprised if it took 5 years.
I don’t what are you on about…
The company itself is great, and is making a ton of money. But the share price of Nvidia is absolutely pure speculative hype !
The math - in the past year, during which Nvidia suddenly found itself in a monopoly situation during a huge spike in AI interest (or / and hype), they made 60B$ in revenue (x2 from 2023), and 30B$ in net profit (x10 ! from 2023). Both numbers are huge. But ... the company is valued at 1800 B$ currently. There is this pesky little ratio that is sometimes useful, called price / earning ratio, and it gives a number around 60 here.
So in short - great company, perfect positioning, but way overpriced.
https://www.astralcodexten.com/p/sam-altman-wants-7-trillion
Just sell shares when they vest to reduce your risk.
If NVDA has a position that fits you and you like the team, go for it! If you're just chasing money.. please don't.
Or, just talk to your hiring manager(s) about the dilemma. For both companies, it'd be better to talk through any concerns before joining. If you accept one offer and then regret it a few months later, everyone loses: you, NVIDIA and [FAANG].
It's a consensus Wall Street favorite. Retail traders who barely know what they do are now being lured in. I think the stock is close to a major top.
Bull markets are born on pessimism, grow on skepticism, mature on optimism and die on euphoria. Judging by general consensus I follow here, on Twitter, and in the financial press, I'd say we are basically at euphoria. Euphoria can last for a bit but not for long. Not only do I not see a ton of upside in the share price here, I think it's a fantastic short opportunity, as when the tide goes out on a stock like this that is totally driven by the idea that it is supply-constrained, if that looks to be in doubt for any reason whatsoever, it is going to fall very far. I'm talking back to the double digits given the amount of supply they are currently producing.
Keep in mind that there are only a handful of companies that can afford these chips at scale and the business model for what they are doing is far from proven. At some point, likely soon given where interest rates are, they are going to need to show a ROI or they are going to be forced to cut spending.
It's a very unique positive feedback loop. Here you have the wealthiest most profitable companies in history that all have a vested interest in not being viewed as waving off the AI hype train. Therefore they basically have to show the street that they're investing which drives up demand for the chips and consequently the price. But you must understand that this process can play out in reverse at 10x the speed. All you need is consensus that spending on AI is suddenly not such a great investment.
I suspect that what you might see happen is earnings for Microsoft, Alphabet, Meta, and Amazon start to suffer. This will require them to cut spending on data center development (unless they are able to show profit growth coming from that investment- unlikely on a short time scale). At that point Nvidia is going to have to cut costs and likely cut back on supply. This is the identical cycle that played out during the last hype cycle, which hilariously, centered on the Metaverse.
They are but it takes a lot of time.
Most of the big players - Google, Meta, OpenAI, Amazon, and Microsoft all are actively developing TPU/NPUs that would be used instead of the H100/A100's everyone is using for machine learning.
Google (tensorflow/jax), Meta(pytorch), Microsoft(onyx), Openai(triton) and Apple (mlx) each have software stacks for optimizing models for multiple platforms.
It takes a lot of time to develop the silicon and software stack. As a result everyone is using H100's in the interum until the hardware/software catches up. Google has been using their TPUs already.
There's other companies like Groq that are also developing NPU/TPU like devices.
And some PoC work: https://vortex.cc.gatech.edu/publications/hotchips-poster.pd...
Neither are specialized TPU/NPUs but they do fast vector operations.
When you include Nvidia's huge margins, this bet is less risky, but it's still only one that the largest companies can make.
Software alone cannot do it, you need to make a bet on hardware.
Joking aside, it is quite expensive and I'm pretty sure you'd run into legal issues due to vertical integration, depending how you perform it. I'll put it this way, a few years ago the headlines were about how China was poaching TSMC workers and offering huge salaries[0,1]. I haven't seen Chinese chips become competitive yet, so it looks like >5 years.
Beyond that, it's more than the chip. AMD has caught up to Nvidia in hardware. But no one is rushing to buy AMD cards because they are still not as good. Nvidia's secret sauce is Cuda (MKL is still an advantage to Intel). The naivity of the Tiny Corp was thinking that everything could be resolved in a few weekends of hacking. But Cuda is deep in a lot of projects. Like a project's backend's backend's backend deep. You got decades of engineers using Cuda and the huge momentum around it. While most programmers will never touch Cuda or GPU code, almost everyone touches things that connect with them (and this is every single day for many people. We could also say the same thing about optimization libraries like MKL). It is something that looks simple because we don't talk about it much or haven't had the experience with. But I assure you, GPU programming is a whole other world. Optimized programming is a very different style of programming and you need a different framework of thinking and problem solving. GPUs add another lay of complexity on top of that. It's why you'll hear so many people complain about writing kernels, but damn, the results speak for themselves.
So you gotta match on the hardware. Then you got to develop great software. Then you got to get that software into the other software that everyone else is using. And you gotta convert people along the way, getting them to turn from a thing they already know and have experience with to a completely new thing.
```edit
The disadvantage of being a first mover is you got to invent everything yourself and page the path, letting others follow. But the disadvantage of being a follower is that to get people to use your road or lane you can't just be equal, you have to be *better*. And you usually have to be significantly so. Momentum is a really powerful force and I think it is highly undervalued.
I think Nvidia is safe for the next few years. They aren't slacking and relying on their momentum. They're still pushing very hard, which only makes it harder for competitors. It's hard to displace a sleeping giant. It's even harder when that giant is fighting back.
```
> And how long would that take?
A long fucking time.
[0] https://news.ycombinator.com/item?id=24129861
[1] https://www.reuters.com/technology/taiwan-raids-chinese-firm...
MI300x is excellent and I'm buying them. =)
It's so challenging, capital intensive, and takes so long to bring these things online (especially if it's not your core competency), that but by the time you got something working in this paradigm, it's possible that a better paradigm/approach will have emerged.
I still would love to see a consumer card with 128GB of VRAM
Imagine this - instead of a motherboard mounted to the back of the case, it is mounted in the center. Behind it would be another motherboard, with a socket for a GPU and ram slots for memory. Kinda like a right/left brain connected with a high-speed interconnect. You open the left side of the case to access CPU and its memory, you open the right side of the case to access the GPU and its memory.
You would be able to upgrade your GPU chip independently of everything else. It wouldn't be mounted sideways in a PCI slot where additional brackets and support arms are needed. You could upgrade the RAM independently. Cooler, etc.
Following on from that though, there are a number of miniitx cases that followed on from that type of layout such as the fractal terra [1] which utilize a pcie riser to pass it rearward to the graphics card behind the motherboard. ATX/ITX wasn't entirely designed for it, but the layout does seem to be quite elegant.
[0] https://www.tomshardware.com/reviews/nuc-11-extreme-kit-beas...
[1] https://www.fractal-design.com/products/cases/terra/terra/te...
The only one doing lots of VRAM are Apple and nobody really care because performance is not that competitive and price is about as bad as entreprise stuff.
I still think modular arch will stay for consumers products there much more benefits to that than chasing a bit more perf for low volume products. Apple can get away wit their shenanigans because they subsidize their high-end chips with the iPhone. If they could only sell Macs it would extremely unsustainable (like the PowerPC venture showed).
The other complication is that we don't have intuitions for AI, so it's easy to underestimate how astronomically large the market will be.
And from an execution standpoint, Jensen (Nvidua CEO) has been preparing for a very long time (albeit unintentially, so nobody could predict AI to this extent). He also tends to go all-in on things, which has caused problems in the past (e.g. crypto crash), but just happened to be incredibly prescient in this moment.
After software have eaten the world, now AI is eating everything.
Even with those materialized earnings, as you say, the market valuation is still at >30 times those earnings (and I'm being very generous by extrapolating this quarter earnings, not the past year).
There is a reason for the saying "the market can stay irrationnal for longer than you can stay solvent". One should not short bubbles. But it's still a bubble !
EDIT: It's a bubble in NVidia valuation. Per se, it's a very good business, generating huge amounts of $.
This suggests huge space for competition to drive earnings growth down. But it's unclear when that will happen.
And more you've heard about.
And yet more you haven't, since they're in stealth.
Theoretically even Apple could start producing AI chips based on their Neural Engine.
The better NVIDIA's results are now the more that encourages the competitors to try and take some of the cake for themselves.
Lisa Su (Jensen's cousin, btw) seems to have turned the ship around.
TIL! Wow, that's wild.
Training is largely done on Nvidia cards, but there's nothing mandating that.
Google trained Gemini on their own in-house TPUs, and according to their published stats it exceeds ChatGPTs performance.
There's a collective delusion right now that somehow CUDA entitles Nvidia to a forever monopoly on GPU compute. There's no way they will maintain ~90% gross margins on their hardware sales. It's far too economically inefficient for purchasers in the long run.
The ones who figure out how to use competitor cards at less than half the cost will have a huge advantage
Right now it's fevered money pouring into what they see as the fastest way to get their feet into the game. Nvidia will do well, but that's already more than priced in right now.
But anyway, good to pull an exact figure!
Totally open to being wrong here!
Given what we currently know, it is hard to say this valuation is irrational or crazy as some are saying in this thread.
What is crazy is Palantir up 4% after hours because of NVDA earnings.
We are certainly selling a ton of pickaxes, pans and shovels here. Not sure about how much gold we are really finding.
The market can remain irrational far longer than you can remain solvent.