I have to give it to NVIDIA, they are great at taking advantage of various industry trends and riding them. I see it over and over again: NVIDIA out in front on a new industry trend.
It is the biggest constant of NVIDIA I've witnessed over the last 2 decades. Each of these trends inevitably comes to an end, but NVIDIA internally is clearly enabling itself for the next possible trends to take advantage of.
The constant is that NVIDIA has been investing in CUDA for the last 2 decades, and has built up an enormous lead in terms of mindshare and existing CUDA code out there.
Specific applications may vary over time, and it's NVIDIA's investment in GPGPU and CUDA that allows them to be a leader in each of these. But it's also worth noting that one of these, neural nets, have been big on NVIDIA hardware for a decade now, so it's not exactly a "trend".
Arguably AI and machine learning IS a trend. Just because it is ten years old does not mean it isn't a trend. A trend is a consistent increase or decrease, with no limits on duration.
AI/ML is far older than a decade, more like 6 decades at this point. However, I agree that interest in it is also wildly going up and down, so maybe more correct is to say that the attention AI/ML receives is what's trending, rather than the area itself.
I was using NVIDIA before CUDA and they were always on trend even back then. First they had the fastest 3D GPUs, or at least the best advertised ones. Then they had the programmable shaders, and then NVIDIA Cg, a great GPU programming language, which was pre-HLSL/GLSL. Then they acquired AGIA, which brought them PhysX + CUDA. But even recently they have innovated with fast networking, tensor cores and RTX. They are a consistent innovation machine.
Shorting NVDA today feels to me like shorting AMZN in the aughts. The secular tailwinds are simply too great to bet against. GPT-4 is only the beginning.
> The secular tailwinds are simply too great to bet against
You can take up both positions if you want. I've got long exposure through various ETFs. I also maintain a small hedge in the form of put options. Overall, most of my NVDA money pile says "it would be fantastic if this all works out". But, I have a small % of the pile dedicated to "Just in case...".
The only situation I don't want is one wherein things stay exactly at this level and don't move anywhere for a long time (or do so incredibly slowly).
Would you bet against volatility in the AI/GPU market over the next 12 months?
> Would you bet against volatility in the AI/GPU market over the next 12 months?
The problem with betting on volatility (via options at least) is that it's already priced into the market, which means you pay a large premium for that bet, which in turn means that things need to make big moves for it to pay out. So it's not enough to be confident that there will be some movement in either direction—it has to be a relatively large movement for it to be profitable.
So to answer your question, maybe. I don't think it's unreasonable to think that Nvidia stock won't drastically change over the next 12 months.
Disclaimer: I'm long Nvidia because I'm willing to wait and see how high this ride goes.
HFTs and hedge funds are profitable because stuff isn't actually priced in. Nvidia has a lot of historic volatility which makes the options expensive, but if you believe the fall will be as fast as the rise, put options rarely price that in.
> HFTs and hedge funds are profitable because stuff isn't actually priced in.
HFT firms and hedge funds often have data or other information that retail investors don't have access to, so that's not an apples to apples comparison.
> if you believe the fall will be as fast as the rise, put options rarely price that in.
That's not what "priced in" means. Unless you have information that others don't, there is no reason to believe that your bet is any more likely than anyone else's. So in that sense, it is priced in based on the probability (according to the broader market) of it actually happening.
Let me rephrase in terminology you like: If your investment thesis based on all the information you have is that the fall of Nvidia will likely be as rapid as its rise, options are rarely priced for that possibility, so you may find the price of an option attractive (in my words, the possibility of that is not priced in).
The term "priced in" doesn't just apply to information - it applies to any sort of investment thesis. The common investment theses (based on a conventional understanding of public information) are usually priced in. The uncommon investment theses are not, but that doesn't mean that they are necessarily wrong (this one probably is, though).
There is certainly no reason to believe that your ideas are better than anyone else's, even if you are a sophisticated hedge fund guy, but the way hedge fund people make money is basically by having contrarian ideas, which aren't priced into securities prices, and being right. Trading that (perceived) mispricing is how a hedge fund generates uncorrelated alpha.
That makes sense, but I guess what I'm saying is that a rapid fall is priced in, in the sense that the market doesn't find it to be very likely. So it's only mispriced if you think the odds of that happening are higher than what the market believes as a whole, and again, it's unlikely (but not impossible) that you have good reasons for believing that as a retail investor.
With that said, I just want to make clear that I don't believe in the random walk theory, and I believe it's perfectly possible to be contrarian and be correct if you have information and/or insight that most other investors don't. I just don't think that the average person on HN (myself included) fits that bill, and I wouldn't hang your life savings on a whim like this unless you're comfortable treating it as what it is: gambling. :)
I never tried to put on the trade (thank goodness) but there was a period of years where Tesla looked like the equity was going to $0. I think as a matter of traditional corporate finance, that was indisputably true, but obviously it did not work out that way in practice.
God bless the short sellers, they’re made of sterner stuff than I am.
>The question, though, is what will be the analogy to mobile (and the cloud), which exploded demand and led to one of the most profitable decades tech has ever seen? The answer may be an already discarded fad: the metaverse.
People haven't really noticed that building a general-purpose (humanoid) robot has become 1000x more feasible now even compared to 1.5 years ago.
Yep, there's a giant pile of training (and inference) that has yet to be done because we've barely touched real-world data (audio, visual and tactile). GPT-4 already demo'd a multimodal version, and I don't see a reason why GPT-N won't be able to seamlessly work on video and tactile input and output actuator commands. I also don't see a reason why humanoid bodies would cost more than a car.
Maybe people don't realize how low the bar for mass-adoption is? We're kinda mesmerized by Boston Dynamics' Parkourbot, but the real-world is made for people of many different abilities, you don't need to do summersaults, or run, or lift a lot to be able to do useful work. A $30k robot with a lifespan of 8 years would cost $10/day. $10 is very little money for in-your-house work, being able to do basic cleanup will be worth it for a vast number of households.
> you don't need to do summersaults, or run, or lift a lot to be able to do useful work
I agree that it's a gimmick, but every time I see it I'm reminded of how far we are from having an in-home robot that can do useful work like wash the bed sheets
and clean the toilet.
Raking leaves, very basic yard upkeep and mowing, is worth around $10 / day to most people (if it's close to basic human performance at those tasks). One can debate the year it will arrive (less than 30 years, more than 10 years), the utility value of a robot that can do that well is extremely high though.
And bring on the snow shoveling bots. Nearly every 50+ year old in cold climates will want one. Will we have to outfit them in specialized winter gear?
I'm pretty convinced we can do snow removal robots now for a few thousand dollars plus a few hundred a year of maintenance (snow and electronics don't mix well), but you are competing with a 25-year-old with a truck and a snowplow blade who comes by and clears it for $100/month.
Lawnmower bots are relatively cheap, so the cost for snow ones will probably go down as well over time (then again the market for them is probably much smaller and yeah, snow removal is cheaper and easier for a human to do compared to lawn mowing)
With all due respect to that company, it also doesn't look like it can stand up to removing very thick snow or operating on a large surface area. My house (and a lot of others who would pay for this) regularly gets more than 12 inches of snow in one event every winter, so the robot would have to operate in the middle of a snowstorm to keep the driveway and sidewalk clear on the day that matters.
Snow removal is also lot more physically strenuous than most other yard tasks, meaning much bigger motors and batteries, almost approaching car-sized.
Robots which do this are already cheap and widely available, I see them everywhere.
Snow blower robots seem to be a thing too, not sure how useful are they. Racking leaves is probably also solvable problem.
It just seems that it might be more effective/cheap/viable to have a bunch of way more simple specialized 'robots' than a very complex/expensive general purpose one.
Not particularly good though. I did a fair bit of looking at them and tried a few on loan a couple years ago. At the time at least, you have to set up an electronic fence to set boundaries, which is very annoying if you have scattered garden beds or other obstacles it cannot bump into.
Also like Roombas (the brand, not robot vacuums in general) they have poor sensors and no real coverages algorithms, they just semi randomly roam around.
Battery life is also terrible, outdoor charging stations are harder to maintain etc.
All I’m saying is, there is a long way to go before they become defacto better than a person with a mower, they are a fair bit of work and fuss right now to set up and keep running.
> they have poor sensors and no real coverages algorithms, they just semi randomly roam around
Husqvarna supports that, but yeah it might struggle on very complex areas (I've only really have any experience with it on fairly simple lawns). It's certainly not just randomly roaming around:
Moravec's paradox is the usual counterargument given to this line of reasoning. We've had far less progress in embodied robotics, where a robot has to interact with the real world in any kind of generalized, tactile way, compared to visual, audio, and language processing tasks. The history of AI is littered with predictions that <a reasoning or computation AI breakthrough> will lead to a humanoid robot, and the predictions always end up in the regime of ~real world data collection and integration is harder than we thought.
Maybe this time it's different, and maybe it's not, but that's why most recent robotics predictions fail to convince the ML industry broadly.
I don't think that's true. I recall when Amazon acquired Kiva robotics for their warehouse operations one of the people involved said something like "If you want something that can pick assorted objects out of a box and put them in another box, I'll need a NASA research team and 5 years, but moving the boxes around, we can do that with our Kiva robots." Here we are 10 years later and Amazon does in fact have the picker arms, though I'm not sure how production-ready they are, Amazon has demoed them.
Honestly I think there's been very dramatic improvements in robotics alongside ChatGPT but ChatGPT is easy to demo with nothing but an internet connection so it's just a lot less visible.
That first quote is right on-the-money, in my experience with competitive FRC and automation. It's extremely easy to work with a limited set of parameters like a cardboard box; you can easily estimate object volume and bounding-box collision in software. Making a robot that manipulates millions of Amazon products is a suicide mission by comparison; especially if you expect it to behave consistently.
Computer vision, AI and inverse kinematics have all come a long ways in the past few years. That being said, it's still easier by an order of magnitude to design the box-pushing robot.
To me, humanoid robotics seems to be the 'flying cars' of our era.
They seem cool and futuristic at a glance, but upon close inspection, they are kinda not pragmatic at all, and there are simpler, cheaper solutions that will beat it in the market.
For example, we could build a humanoid butler robot to do chores around the house, or we could build a dishwashers + roombas + washing machine. These simple appliances do 90% of the work for 10% of the headache (including the consumer side headache of performing maintenance on a humanoid machine!).
You clearly need both? You'd never task a humanoid robot to wash clothes by hand. But you would task it to remember to check whether its time to do laundry, load the laundry, run it, move it to the dryer, fold the clean clothes and put it away.
It's not the case that existing robots (dishwasher, washing machine, etc) do 90% of the work when you consider the full task in its entirety. Humanoid robots would just replace the work that humans are doing assuming they have access to the specialized robots.
I don't think you "need" either. It's just a conversation about what the market will reward.
> Humanoid robots would just replace the work that humans are doing assuming they have access to the specialized robots.
Or, there won't be any replacement and the majority of people in domestic households just keep doing that little remainder part because it's not really a big deal.
This all holds perfectly with the flying car analogy -> ground cars are pretty much fine. We could try way harder to make them flying-grade at a consumer level, or we can just keep driving road cars (or taking public transit) and be happy. That's the fundamental thing that 60s futurists missed when envisioning the future.
edit: btw, I'd like to add that from a design perspective, humans kind of have a crappy form, which is perhaps one of the reason that rarely does one design something which ends up humanoid. E.g. something with wheels can usually move farther with less energy (or even something like tank treads if rough terrain)
a humanoid robot is a clunky solution to all problems.
You'll always have more specialized equipment that just does the job better. The "all purpose very complex" mediocre solution is not something that will stick in my opinion.
> building a general-purpose (humanoid) robot has become 1000x
Why? I'm not even sure how ChatGPT is that relevant for this? Sure if you want a scifi style "intelligent" robot, which is capable of basic reasoning, you can talk to it's useful.
However I don't see how does it help with it being able to pick up fragile stuff without crushing it, operate in dynamic 3d environments etc. LLM don't seem to be that useful for these kind of problems.
I think the key reason why the world will move a lot of stuff that “already exists on CPUs” to GPUs is not that we’re porting existing code, it’s that GPU acceleration is a meta-trend that is disrupting all highly technical fields. It started much earlier decades ago in gaming and video. It then moved to grid and scientific computing. It has now fully revolutionized AI. The returns on better CPU code are much less than the returns on GPU code in the medium term as many of these trends in graphics, scientific computing, AI are mutually reinforcing. When 90% of your pipeline runs of GPU then the last 10% might as well too. There are obviously tons of small things that require CPU but the most important and crucial systems code should be moved to massive parallel architectures.
CUDA being feature complete and having libraries & frameworks optimized and immediately deploy-able for it is all the difference. I don't like much of their business/marketing practices for their gaming GPU division, but their engineering and HPC teams are world class and have been awesome to work with / use as an someone who writes scientific computing code on the GPU.
I'm all for the post, but are we really attempting to add logic to stock prices? The price of NVDA stock does not correlate with any sort of logical math, this plus this = stock price type of equation.
Stock prices are entirely based on what people think they are worth.
there is also a secondary market for options - where investors trade volatility, and use stock as instrument to trade volatility and express their view.
if volatility is priced at a bargain, investors doesnt care much about the stock, as he will be hedging his delta exposure anyways by means of dynamic hedging/etc
Does it not? Given their current profits, we would expect 25b in rolling year profits. Their market cap is 1.2t. Apple, with about 80b in rolling year, has a market cap of 2.8t
Given the kind of revenue and profit growth forecast NVDA has given, these numbers don't seem terribly far off.
Worth noting that NVDA's profits 8xed in a single year. A lot of people are basing their NVDA valuation based on the past year's revenue, which does make them look extremely overvalued. However, if NVDA has enough sales and enough of a moat to double their profits again, suddenly they'll be biting at Apple's heels profit wise with a much lower valuation.
It doesn't seem like a totally crazy valuation to me, despite the uncertainties involved.
> Stock prices are entirely based on what people think they are worth.
Are are priced on what people think they might be work in the future current financial figures don't really matter that much compared to growth.
And the market is pricing in extremely high revenue growth (and also very high margins). So far Nvidia seems to be delivering. Their forward PE went down several times and it's now "only" 30. A few more quarters like that and the current stock price might seem cheap even by traditional standards.
Obviously it's a risky bet but I don't see what's illogical about it?
>Model training, though, is an up-front cost, not dissimilar to the cost needed to buy those Sun servers and Cisco routers in the dot-com era...
This is only partially true even now, and I expect it will be less true in the future as continuous training becomes the norm. As Hericlitus said, you cannot step in the same river twice, and data drift is a real issue when modeling complex processes. It will become even more of an issue as the widespread use of models starts to impact the data streams they rely on for training: acting on a prediction can change the underlying assumptions that produced the prediction in the first place.
Not that I disagree but I think this is what's going to pop the AI bubble. It doesn't actually work that great for a lot of stuff, anything that requires true generalization as opposed to just fitting a particular distribution. That doesn't make it useless, but it makes it way less useful than people believe in waves of optimism. It's what killed the last hype cycle and it will again.
It didn’t pop the cloud bubble that you have to do constant maintenance. In the end, a new class of experts arises to handle these challenges, paid for by the layoffs of these made redundant by AI
There's a lot being said about the moat nVidia has with CUDA, but when companies like Tesla are forking $300M on 10,000 GPUs [1], it can't be that difficult to port whatever code to whatever platform if 100s of millions are in question, or is it?
Can Google/Tesla/etc create an H100-like GPU in a couple of years at a fraction of the cost? And is CUDA really necessary if this much money is at stake?
> it can't be that difficult to port whatever code to whatever platform if 100s of millions are in question, or is it?
> in a couple of years at a fraction of the cost? And is CUDA really necessary if this much money is at stake?
I mean, that's the thing; is resisting CUDA really worth the cost when this much money is at stake?
It's not like you have to pay to license CUDA. The two big drawbacks are that it's proprietary and locks you into their ecosystem; neither of which really matter when shipping stuff at that scale. These companies could spend a few million dollars to accelerate their specific codepath, but it's frankly wasted money unless you have a specific reason to avoid Nvidia.
The eventual Nvidia-killer will probably be platform-agnostic tooling like ONNX, unless hardware manufacturers revive a sort of OpenCL-style acceleration library.
Yeah, but a lot of other companies have similar libraries with varying levels of commitment/coverage. ARM has ARMnn, Apple has CoreML, Intel has both Vino and OneAPI (plus SSE/AVX implementations), Google has NNAPI for Android and Google Cloud tensor accelerators, AMD has... Vitis/Xilinix, ROCm, OpenCL and MIGraphX, Windows has DirectML/Olive alongside Microsoft's Azure Execution Providers, Rockchip has RKNPU and Huawei has CANN.
Suffice to say, there are a lot of "competing standards" a-la the XKCD : https://xkcd.com/927/
I don't think any of those will really topple CUDA, though. The best way to unseat Nvidia's dominance would be to target their two weaknesses; the closed nature of CUDA and the lock-in to Nvidia hardware. It would be difficult to overturn them both, but an open and fast CUDA alternative is really all people actually want. That's why I think libraries like ONNX have the right idea; instead of relying on chip manufacturers to not rip each other's throats out (they won't), they unify everyone's proprietary APIs. Barring some ground-up GPGPU library like OpenCL, this seems like the smartest path to me.
CUDA's moat is not built around inference, as far as I can tell. It's a very handy tool for deployments, and often gives them the edge when Nvidia pulls ahead, but it isn't a be-all-end-all. The inferencing situation has been pasted over by countless vendor-specific libraries and stuff like Pytorch and ONNX.
The real reason for the moat comes down to a number of things, like:
- The flexibility of generic GPU and ML acceleration primitives
- The availability of systems with hundreds of terabytes of GPU memory for you to scale to
- The struggle of trying to use commodity hardware for actual ML acceleration
I deploy models to freely-provisioned ARM servers, I don't think I'm a choosing beggar in the slightest. When I deploy to Nvidia hardware though, the experience is much nicer on-the-whole. This stuff is definitely possible with consumer hardware, ROCm acceleration or CoreML optimization, no doubt. It's not hard to see why Nvidia is at the mountaintop right now though, and unless the industry agrees to stop building CUDA-style moats then this is the history we're damned to repeat.
Dunno if this is a rhetorical question because a lot of people here seems to think that yes, these companies can make their own silicon for much cheaper. The problem is that they need time, and right now they still have work to do, so the best way to solve this problem is to buy from nvidia. Or am I saying too much?
> it can't be that difficult to port whatever code to whatever platform if 100s of millions are in question, or is it?
300M hires a lot of engineers. Maybe the fact that they are buying instead of building tells you how hard it is to build and how strong a moat Nvidia has.
Musk has publicly said that the true test of their own training hardware efforts is if the ML team switches to it. I.e. it’s a software, not hardware problem.
Maybe flexibility? Google's TPUs are comparable in performance for pure ML, but at the same time if you're building a huge expensive cluster, you might want to be able to have the flexibility to accelerate other tasks (eg scientific compute).
I’d be surprised if we don’t see more specialized chips in the future- kind of like Bitcoin isn’t mined on GPUs anymore. There just hasn’t been much of a need for specialized AI chips so far but GPUs have been useful and profitable in rendering etc for a long time. But I think Google made “TPU”s already (tensor processing unit)
Seems odd not to mention competitors such as AMD. Not as if Nvidia has any secret sauce. "Analysts" like this also seem to misunderstand Cuda as that such a sauce, when in fact AI wants higher-level toolkits (which are not at all tied to Nvidia).
Nvidia does have secret sauce, though. That's kinda the point of CUDA cores and Tensor accelerators and RT cores, even. They are all secret (read: proprietary) and provide relevant "sauce" via performance improvements Nvidia monopolizes. It's arguably unfair, but just the way the cookie crumbles. Nvidia funded GPGPU drivers when AMD and Apple abandoned their cross-platform efforts. Nvidia funded early GAN and transformer-architecture papers, and incorporated their findings into the software layer.
Analysts certainly misunderstand CUDA, but that doesn't change it's status as a non-negligible moat. It is a by-product of an overly hostile computer industry, spearheaded by a company that's more than happy to profit from bad blood.
CUDA is the secret sauce. AMD has no equivalent, nor does it have a high level alternative. People have begged AMD for this for years and they’ve only started to take it seriously now.
AMD's equivalent is HIP [1], for sufficiently flexible definitions of "equivalent". I can't speak to how complete/correct/performant it is (I'm just a guy running tutorial/toy-level ML stuff on an RDNA1 card), but part of AMD's problem is that it might not practically matter how well they do this because the broader ecosystem support specifically for the CUDA stack is so entrenched.
I’m not super familiar with HIP, but historically AMD’s past responses to CUDA only supported a small set of GPUs, was poorly documented, and was poorly maintained. It seemed like AMD’s attention span on these projects felt similar to Google’s attention span for new projects ie it’s a good bet that it will die in favor of a new SDK which will ultimately suffer the same fate
I don't disagree with your general impression, but the flip-side of it (combined with the fact that the stack is open-source) is that they're not in a position to execute some exquisitely bullshit market segmentation strategy, and there are at least some wins to be had if you put in a bit of legwork (like making Stable Diffusion run on the GPU I've owned for years instead of spending an indefensible amount of money on the rough Nvidia equivalent).
HIP is a good start since it has CUDA compatibility and open source from the beginning. The issue is that developers still might not invest in it until AMD shows that they’re willing to commit to it unlike their past projects. A lot of them have been burned on this front by AMD, but the open source designation should help a lot.
It's not even AMD's unwillingness to commit. They were among the first to implement Vulkan and OpenCL, and have a decent legacy of GPU hardware support. Their open-source kernel drivers on Linux are cream-of-the-crop, and their raster APIs are so good that they truly rival Nvidia.
The biggest issue is frankly that other stakeholders are doing their own thing, and pretending it's okay. Long-term, I think this will be Nvidia's real coup-de-grace; letting local devices do inferencing, while selling their GPUs to larger deployments and training applications. Stuff like CoreML and ROCm are nice, but don't directly threaten CUDA. If their competitors want to be rid of CUDA, then they have to gang up on them and offer a better alternative. Until then, Nvidia will have a market.
It's true what they say, nice guys really do finish last in this economy.
> they're not in a position to execute some exquisitely bullshit market segmentation strategy
nevertheless that is exactly what they spent the last decade trying, with ROCm not even officially supporting the consumer-card variants of the handful of radeon pro cards that actually got support. AMD's GPGPU software support was almost entirely segmented to the CDNA cards for the HPC market until literally a month ago, and still completely ignores any sort of binary compatibility story (they really want you to distribute as source and compile in-situ, which really makes it a non-starter for end-user software or commercial distribution).
I think the point estimating GPU utilization (the terminology is transcending ), and by extension GPU demand misses the point. Of course they will buy them, they have to because everybody else is.
The questions is whether the GPUs makes money for these companies. When the market sniffs the answer to that question, that's when the bottom falls out or not.
Is Google going to make more money from me than it already does by embedding LLMs into its search product? Am I going to be spending any meaningfully more amount of time using any of these products than I already do by the addition of AI? Because if not where does the money come from?
As a PR gamer, I'm concerned the biggest graphics card maker makes most of its money selling to data centers bulking up AI capabilities.
Like, if there's a shortage of GPUs in 2026 and Nvidia has to choose between selling limited inventory to AWS or individual gamers, it's obvious who wins (AWS and scalpers).
At first I thought I was missing something and had to look at the plots a second time to realize that only 1 of the graphs was correctly labelled and the others are all missing legends :( which means you have to guess what each series corresponds to by reading the surrounding paragraphs.
I'm surprised here that he doesn't investigate: "Is NVIDIA the minicomputer of their era?"
ie. Will an Intel (or indeed one of the cloud vendors) undercut their margin with a better priced GPU in a similar manner as Compaq did the IBM Mainframe?
Or, and hear me out here, Ben's previous analysis was just wrong and he failed to see what Nvidia's gameplan was. He writes as if they did some serious soul-searching when their stock price was low (perhaps even reading some brilliant analytical blogs!), but the reality of product development in the chip/peripheral space is that nothing happens that fast. Nvidia is riding high now because so many big bets that they made years ago are paying off at once.
I am baffled when people say Nvidia got lucky with the recent AI wave.
Alexnet, one of the first papers in 2012 to kickstart the current AI revolution was using Nvidia GPUs. Nvidia has been prioritizing AI research (both hardware and software) for the last decade.
Yeah, Ben’s primary talent seems to be weaving a narrative about how he’s successfully predicted 18 of the last 11 trends to reshape the industry, and to make it sound believable. Nvidia’s executing on all cylinders, and has been for a while (some missteps in the consumer marketing space notwithstanding). I’m still kicking myself a little for not buying some of their stock a year ago when it was obvious even to a simpleton like me that they were very well positioned to capitalize on current trends.
Actual monetization of AI is likely to be a large hurdle. Right now lot of companies are throwing shit at the wall to see what sticks in a panic to not get left behind. The downstream business models which will generate sustainable revenue to keep AI R&D running is somewhat lacking right now. That does smell like the 2001-era internet bubble. There's probably one Google-Search-level idea out there, but I'm not sure exactly which one it is. And remember it has to not only be a good idea, but it needs to be monetizable, with a substantial moat around it.
So over the near term, 4-5 years I suspect AI hype comes back down closer to Earth again. Sure the secular trend is all up and to the right, but the secular trend of the internet was still up and to the right in late-2000 as well.
96 comments
[ 3.1 ms ] story [ 173 ms ] threadIt is the biggest constant of NVIDIA I've witnessed over the last 2 decades. Each of these trends inevitably comes to an end, but NVIDIA internally is clearly enabling itself for the next possible trends to take advantage of.
Specific applications may vary over time, and it's NVIDIA's investment in GPGPU and CUDA that allows them to be a leader in each of these. But it's also worth noting that one of these, neural nets, have been big on NVIDIA hardware for a decade now, so it's not exactly a "trend".
You can take up both positions if you want. I've got long exposure through various ETFs. I also maintain a small hedge in the form of put options. Overall, most of my NVDA money pile says "it would be fantastic if this all works out". But, I have a small % of the pile dedicated to "Just in case...".
The only situation I don't want is one wherein things stay exactly at this level and don't move anywhere for a long time (or do so incredibly slowly).
Would you bet against volatility in the AI/GPU market over the next 12 months?
The problem with betting on volatility (via options at least) is that it's already priced into the market, which means you pay a large premium for that bet, which in turn means that things need to make big moves for it to pay out. So it's not enough to be confident that there will be some movement in either direction—it has to be a relatively large movement for it to be profitable.
So to answer your question, maybe. I don't think it's unreasonable to think that Nvidia stock won't drastically change over the next 12 months.
Disclaimer: I'm long Nvidia because I'm willing to wait and see how high this ride goes.
HFT firms and hedge funds often have data or other information that retail investors don't have access to, so that's not an apples to apples comparison.
> if you believe the fall will be as fast as the rise, put options rarely price that in.
That's not what "priced in" means. Unless you have information that others don't, there is no reason to believe that your bet is any more likely than anyone else's. So in that sense, it is priced in based on the probability (according to the broader market) of it actually happening.
The term "priced in" doesn't just apply to information - it applies to any sort of investment thesis. The common investment theses (based on a conventional understanding of public information) are usually priced in. The uncommon investment theses are not, but that doesn't mean that they are necessarily wrong (this one probably is, though).
There is certainly no reason to believe that your ideas are better than anyone else's, even if you are a sophisticated hedge fund guy, but the way hedge fund people make money is basically by having contrarian ideas, which aren't priced into securities prices, and being right. Trading that (perceived) mispricing is how a hedge fund generates uncorrelated alpha.
With that said, I just want to make clear that I don't believe in the random walk theory, and I believe it's perfectly possible to be contrarian and be correct if you have information and/or insight that most other investors don't. I just don't think that the average person on HN (myself included) fits that bill, and I wouldn't hang your life savings on a whim like this unless you're comfortable treating it as what it is: gambling. :)
God bless the short sellers, they’re made of sterner stuff than I am.
People haven't really noticed that building a general-purpose (humanoid) robot has become 1000x more feasible now even compared to 1.5 years ago.
Check out https://www.lerf.io/ for a taste of the future
Maybe people don't realize how low the bar for mass-adoption is? We're kinda mesmerized by Boston Dynamics' Parkourbot, but the real-world is made for people of many different abilities, you don't need to do summersaults, or run, or lift a lot to be able to do useful work. A $30k robot with a lifespan of 8 years would cost $10/day. $10 is very little money for in-your-house work, being able to do basic cleanup will be worth it for a vast number of households.
> you don't need to do summersaults, or run, or lift a lot to be able to do useful work
I agree that it's a gimmick, but every time I see it I'm reminded of how far we are from having an in-home robot that can do useful work like wash the bed sheets and clean the toilet.
And bring on the snow shoveling bots. Nearly every 50+ year old in cold climates will want one. Will we have to outfit them in specialized winter gear?
https://www.yarbo.com/shop
Lawnmower bots are relatively cheap, so the cost for snow ones will probably go down as well over time (then again the market for them is probably much smaller and yeah, snow removal is cheaper and easier for a human to do compared to lawn mowing)
Snow removal is also lot more physically strenuous than most other yard tasks, meaning much bigger motors and batteries, almost approaching car-sized.
Robots which do this are already cheap and widely available, I see them everywhere.
Snow blower robots seem to be a thing too, not sure how useful are they. Racking leaves is probably also solvable problem.
It just seems that it might be more effective/cheap/viable to have a bunch of way more simple specialized 'robots' than a very complex/expensive general purpose one.
Also like Roombas (the brand, not robot vacuums in general) they have poor sensors and no real coverages algorithms, they just semi randomly roam around.
Battery life is also terrible, outdoor charging stations are harder to maintain etc.
All I’m saying is, there is a long way to go before they become defacto better than a person with a mower, they are a fair bit of work and fuss right now to set up and keep running.
Husqvarna supports that, but yeah it might struggle on very complex areas (I've only really have any experience with it on fairly simple lawns). It's certainly not just randomly roaming around:
https://www.youtube.com/watch?v=KLdSLJu8acg&ab_channel=Husqv...
I'm not even sure what other brands exist. They seem to completely dominate the market where I am.
Maybe this time it's different, and maybe it's not, but that's why most recent robotics predictions fail to convince the ML industry broadly.
Honestly I think there's been very dramatic improvements in robotics alongside ChatGPT but ChatGPT is easy to demo with nothing but an internet connection so it's just a lot less visible.
Computer vision, AI and inverse kinematics have all come a long ways in the past few years. That being said, it's still easier by an order of magnitude to design the box-pushing robot.
They seem cool and futuristic at a glance, but upon close inspection, they are kinda not pragmatic at all, and there are simpler, cheaper solutions that will beat it in the market.
For example, we could build a humanoid butler robot to do chores around the house, or we could build a dishwashers + roombas + washing machine. These simple appliances do 90% of the work for 10% of the headache (including the consumer side headache of performing maintenance on a humanoid machine!).
It's not the case that existing robots (dishwasher, washing machine, etc) do 90% of the work when you consider the full task in its entirety. Humanoid robots would just replace the work that humans are doing assuming they have access to the specialized robots.
I don't think you "need" either. It's just a conversation about what the market will reward.
> Humanoid robots would just replace the work that humans are doing assuming they have access to the specialized robots.
Or, there won't be any replacement and the majority of people in domestic households just keep doing that little remainder part because it's not really a big deal.
This all holds perfectly with the flying car analogy -> ground cars are pretty much fine. We could try way harder to make them flying-grade at a consumer level, or we can just keep driving road cars (or taking public transit) and be happy. That's the fundamental thing that 60s futurists missed when envisioning the future.
Are you down that there's no flying cars?
I'm pretty happy with it tbh.
edit: btw, I'd like to add that from a design perspective, humans kind of have a crappy form, which is perhaps one of the reason that rarely does one design something which ends up humanoid. E.g. something with wheels can usually move farther with less energy (or even something like tank treads if rough terrain)
a humanoid robot, i think, is a good solution for a bunch of problems.
the comparison is kind of useless and doesn't really help us.
My laundry machine is certainly better than by hand, but folding laundry and putting it away is significant.
Why? I'm not even sure how ChatGPT is that relevant for this? Sure if you want a scifi style "intelligent" robot, which is capable of basic reasoning, you can talk to it's useful.
However I don't see how does it help with it being able to pick up fragile stuff without crushing it, operate in dynamic 3d environments etc. LLM don't seem to be that useful for these kind of problems.
Maybe on the data center, but in the audio world this is already happening.
https://www.gpu.audio/
VSL (a huge company in the audio/production world) added it recently to their convolution reverb product.
https://www.gpu.audio/partners/vsl
https://www.youtube.com/watch?v=L3xcq_m3Llw
Stock prices are entirely based on what people think they are worth.
So yes, that makes it possible for the price to become disconnected from any reasonable prediction.
if volatility is priced at a bargain, investors doesnt care much about the stock, as he will be hedging his delta exposure anyways by means of dynamic hedging/etc
Given the kind of revenue and profit growth forecast NVDA has given, these numbers don't seem terribly far off.
Worth noting that NVDA's profits 8xed in a single year. A lot of people are basing their NVDA valuation based on the past year's revenue, which does make them look extremely overvalued. However, if NVDA has enough sales and enough of a moat to double their profits again, suddenly they'll be biting at Apple's heels profit wise with a much lower valuation.
It doesn't seem like a totally crazy valuation to me, despite the uncertainties involved.
> Stock prices are entirely based on what people think they are worth.
Are are priced on what people think they might be work in the future current financial figures don't really matter that much compared to growth.
And the market is pricing in extremely high revenue growth (and also very high margins). So far Nvidia seems to be delivering. Their forward PE went down several times and it's now "only" 30. A few more quarters like that and the current stock price might seem cheap even by traditional standards.
Obviously it's a risky bet but I don't see what's illogical about it?
This is only partially true even now, and I expect it will be less true in the future as continuous training becomes the norm. As Hericlitus said, you cannot step in the same river twice, and data drift is a real issue when modeling complex processes. It will become even more of an issue as the widespread use of models starts to impact the data streams they rely on for training: acting on a prediction can change the underlying assumptions that produced the prediction in the first place.
Can Google/Tesla/etc create an H100-like GPU in a couple of years at a fraction of the cost? And is CUDA really necessary if this much money is at stake?
[1] https://www.tomshardware.com/news/teslas-dollar300-million-a...
> in a couple of years at a fraction of the cost? And is CUDA really necessary if this much money is at stake?
I mean, that's the thing; is resisting CUDA really worth the cost when this much money is at stake?
It's not like you have to pay to license CUDA. The two big drawbacks are that it's proprietary and locks you into their ecosystem; neither of which really matter when shipping stuff at that scale. These companies could spend a few million dollars to accelerate their specific codepath, but it's frankly wasted money unless you have a specific reason to avoid Nvidia.
The eventual Nvidia-killer will probably be platform-agnostic tooling like ONNX, unless hardware manufacturers revive a sort of OpenCL-style acceleration library.
Suffice to say, there are a lot of "competing standards" a-la the XKCD : https://xkcd.com/927/
I don't think any of those will really topple CUDA, though. The best way to unseat Nvidia's dominance would be to target their two weaknesses; the closed nature of CUDA and the lock-in to Nvidia hardware. It would be difficult to overturn them both, but an open and fast CUDA alternative is really all people actually want. That's why I think libraries like ONNX have the right idea; instead of relying on chip manufacturers to not rip each other's throats out (they won't), they unify everyone's proprietary APIs. Barring some ground-up GPGPU library like OpenCL, this seems like the smartest path to me.
We get equivalent performance on non-NVIDIA chips without it and most of the stuff is abstracted away these days
We do write CUDA when needed but really only a handful of folk will actually train models as it is a pain
The real reason for the moat comes down to a number of things, like:
- The flexibility of generic GPU and ML acceleration primitives
- The availability of systems with hundreds of terabytes of GPU memory for you to scale to
- The struggle of trying to use commodity hardware for actual ML acceleration
I deploy models to freely-provisioned ARM servers, I don't think I'm a choosing beggar in the slightest. When I deploy to Nvidia hardware though, the experience is much nicer on-the-whole. This stuff is definitely possible with consumer hardware, ROCm acceleration or CoreML optimization, no doubt. It's not hard to see why Nvidia is at the mountaintop right now though, and unless the industry agrees to stop building CUDA-style moats then this is the history we're damned to repeat.
300M hires a lot of engineers. Maybe the fact that they are buying instead of building tells you how hard it is to build and how strong a moat Nvidia has.
Musk has publicly said that the true test of their own training hardware efforts is if the ML team switches to it. I.e. it’s a software, not hardware problem.
Analysts certainly misunderstand CUDA, but that doesn't change it's status as a non-negligible moat. It is a by-product of an overly hostile computer industry, spearheaded by a company that's more than happy to profit from bad blood.
[1] https://github.com/ROCm-Developer-Tools/HIP
The biggest issue is frankly that other stakeholders are doing their own thing, and pretending it's okay. Long-term, I think this will be Nvidia's real coup-de-grace; letting local devices do inferencing, while selling their GPUs to larger deployments and training applications. Stuff like CoreML and ROCm are nice, but don't directly threaten CUDA. If their competitors want to be rid of CUDA, then they have to gang up on them and offer a better alternative. Until then, Nvidia will have a market.
It's true what they say, nice guys really do finish last in this economy.
nevertheless that is exactly what they spent the last decade trying, with ROCm not even officially supporting the consumer-card variants of the handful of radeon pro cards that actually got support. AMD's GPGPU software support was almost entirely segmented to the CDNA cards for the HPC market until literally a month ago, and still completely ignores any sort of binary compatibility story (they really want you to distribute as source and compile in-situ, which really makes it a non-starter for end-user software or commercial distribution).
The questions is whether the GPUs makes money for these companies. When the market sniffs the answer to that question, that's when the bottom falls out or not.
Is Google going to make more money from me than it already does by embedding LLMs into its search product? Am I going to be spending any meaningfully more amount of time using any of these products than I already do by the addition of AI? Because if not where does the money come from?
Google Search might not earn more money it does now by buying Nvidia devices, but might loose big to a competitor if it does not.
Like, if there's a shortage of GPUs in 2026 and Nvidia has to choose between selling limited inventory to AWS or individual gamers, it's obvious who wins (AWS and scalpers).
More impressive to me is that he didn’t let folk go because of the many downturns on the way
ie. Will an Intel (or indeed one of the cloud vendors) undercut their margin with a better priced GPU in a similar manner as Compaq did the IBM Mainframe?
I mean someone will build an AI to do that right ^_^
Video and 3D are pretty much in the next year, all the components see there to dramatically shift cost structure and they need GPUs.
Alexnet, one of the first papers in 2012 to kickstart the current AI revolution was using Nvidia GPUs. Nvidia has been prioritizing AI research (both hardware and software) for the last decade.
So over the near term, 4-5 years I suspect AI hype comes back down closer to Earth again. Sure the secular trend is all up and to the right, but the secular trend of the internet was still up and to the right in late-2000 as well.