Nvidia Stock: Be Fearful When Others Are Greedy (NASDAQ:NVDA)

Funding Thesis

Even after a unprecedented rally to affix the membership of trillion-dollar market cap firms, it appears investor sentiment concerning Nvidia (NASDAQ:NVDA) stays very bullish, with some even claiming that the inventory has solely turn out to be cheaper.

Nonetheless, it has not. In absolute phrases, the inventory has solely turn out to be dearer (clearly), which suggests the doable draw back has solely turn out to be higher, arguably extra so than the upside going ahead, even regardless of the valuation at first sight not being extreme.

One can’t predict how a lot demand Nvidia will see sooner or later for its chips. Whereas, positive, the explosion of demand resulting from GAI has accelerated what was already a multiyear uptrend, buyers do should weigh the expansion of AI (on one hand) towards the aggressive dynamics of a participant with very excessive market share and pricing, which it has the danger of shedding (on the different hand).

Background

To make certain, the thesis right here is not that Nvidia is essentially overvalued or in a bubble. Given the explosion in gross sales and earnings, since that’s after all what the inventory worth displays (with a sure a number of of gross sales and/or earnings), it has certainly just about grown in step with these outcomes. So already in earlier protection of the inventory, I had conceded to being incorrect about Nvidia: Nvidia: All Bets Are Off.

Historical past and Aggressive Panorama

Some time in the past, Intel (INTC) CEO Pat Gelsinger said that, whereas he acknowledged Nvidia’s success on this market, it was however the results of luck. To recap, AI is a extremely parallelizable workload. As such, when it began gaining prevalence (since 2012), it was discovered to run sooner on GPUs (initially created for gaming the place many pixels should be rendered concurrently) than on the extra general-purpose CPU. That is why Nvidia principally benefited from the rise of AI quite than Intel.

It wasn’t till 2017, nonetheless, that Nvidia launched the primary deep learning-specific acceleration into its GPUs with its Tensor Cores. By that point, Intel too had already launched the VNNI instruction into its Xeon Phi line-up (which was its many-core CPU on the time that competed towards Nvidia in HPC) and had additionally acquired Nervana in 2016, the predecessor to its 2019 Habana acquisition that it presently positions as its aggressive resolution to Nvidia.

In that regard, my very own view can be that it has been a mixture of luck and talent. Whereas Nvidia clearly has strongly executed on its roadmap, Intel then again stumbled (each with its personal 10nm course of, which brought on the delay and diminished the competitiveness of virtually all of its merchandise, in addition to by axing all Nervana growth in favor of restarting with Habana) proper because the demand for AI ({hardware}) first began meaningfully rising round 2017.

In addition to Intel and AMD (AMD), which like Intel has additionally been late to launch a aggressive product, there had additionally been a growth in AI chip start-ups. None of these has been very profitable both. Lastly, many larger firms additionally began creating their very own chips with extra blended outcomes, such because the Google (GOOG) TPU and Amazon (AMZN) Trainium and Inferentia.

Implications

Nvidia’s place within the AI market appears harking back to Intel within the knowledge heart as much as about half a decade in the past. Intel had the market utterly for itself, as AMD was nonetheless within the midst of creating Epyc and getting some preliminary adoption. Whereas this resulted in some criticism as Intel rolled out its highest-specced Xeons for $10k or much more on the time, arguably these dwarf compared to the pricing of Nvidia’s newest H100 (and upcoming H200) collection, which throughout this time of shortages has ballooned to tens of 1000’s of {dollars} (and possibly was by no means lower than $20k or so, to start with). For comparability, TSMC (TSM) expenses lower than $20k for a N5-class wafer, which accommodates on the order of 70 chips.

For my part, because of this Nvidia is leaving the market extensive open for competitors to enter with aggressive (equally performing) merchandise at a probably considerably diminished worth. That is beginning to occur from each AMD and Intel. Though many bulls have argued that Nvidia has a moat with its CUDA software program, as Pat Gelsinger has said this can be a shallow moat at greatest as software program growth principally happens in higher-level languages reminiscent of Python.

Since for any rational enterprise there isn’t any basic purpose to purchase an Nvidia chip over a equally performing (however less expensive) Intel or AMD chip, because of this conventional aggressive market dynamics sooner or later ought to kick in. This may most probably lead to each market share and pricing energy loss for Nvidia.

From that view, buyers ought to ask themselves if they might quite spend money on a ~$200B or ~$1.35T firm. Observe that each firms, so additionally the smaller one, have greater than sufficient assets to develop an AI chip. And for the reason that energy and efficiency traits of those chips are largely decided/restricted by course of expertise, the ensuing chips ought to certainly be fairly comparable (apart from worth). That’s precisely what it seen out there. The truth is, since Intel’s Gaudi collection lacks conventional GPU performance, the Gaudi chips are literally noticeably superior to their Nvidia competitor on the identical course of expertise (see MLPerf).

To make use of some concrete math, if Nvidia would lose half of its market share and needed to cut back costs by half, then its income would drop by 75%. Observe that Nvidia’s margin is so excessive that decreasing the value by 2x remains to be a really conservative situation. If one treats these chips as commodities reminiscent of within the reminiscence (DRAM/NAND) area (which arguably is not all that far-fetched), then Nvidia’s costs would actually should drop by on the order of 10x. (Observe that this estimate has been obtained by solely contemplating the price of the silicon chip, another elements such because the HBM reminiscence and CoWoS packaging are uncared for.)

Finally, that is the type of threat, which is much from unrealistic, that buyers a minimum of should take into account, after which weigh this in comparison with any doable additional upside that has not but been baked into the inventory worth. Since being valued as a progress inventory means there’s already some appreciable additional improve in demand baked into the value, demand must surpass even these ranges for the share worth to extend something in any respect.

Extra Issues

As said, it’s not doable to foretell the longer term demand for AI {hardware}. Even when the danger thesis as described up to now precisely performs out, there’s nonetheless the likelihood that the market would possibly develop sooner and/or turn out to be bigger than any theoretical loss in income from decrease costs and market share, as, for instance, AMD has predicted at its occasion late final yr. Additionally, much like Intel, inertia would already go a great distance in the direction of stopping too sudden market share shifts.

So, to recap some doable drivers of demand for buyers to think about for themselves. One is the distinction between coaching and inference. Coaching a mannequin should be finished solely as soon as, so the demand for coaching {hardware} doubtless has a fairly finite restrict. Nonetheless, the coaching workload is such that it may fairly actually use an limitless quantity of computing energy, since coaching the most important fashions on monumental quantities of knowledge may take many months, and even then it might nonetheless be doable to additional improve the quantity of knowledge and/or parameters.

Therefore, this doubtless has some implications with regard to cost elasticity. For instance, as a substitute of firms shopping for a sure variety of chips, as a substitute, they could have a sure funds in {dollars}, they usually would possibly simply purchase as a lot chips as they’ll inside that funds. So, if Nvidia would scale back the value by 2x, they could simply purchase 2x extra GPUs.

Observe, although, that on this case, since even the most important clouds and massive tech firms solely have a finite funds, there ought to certainly be some higher restrict on demand, and buyers must reply why, moreover the shortages, this restrict wouldn’t have been reached but within the wake of the ChatGPT hype, with Nvidia’s run fee approaching $80B (which when it comes to the whole value of a knowledge heart solely represents spending on GPUs and networking), and as a substitute would proceed to develop significantly for years to return.

Alternatively, and maybe that is the reply to the earlier query, whereas inference (utilizing the skilled mannequin) is usually thought of to be the primary AI software within the long-term, it however has considerably completely different compute necessities, because the computation value of a single inference is usually a lot decrease. As a substitute of coaching a mannequin for months, in inference one requires ends in seconds or milliseconds.

This in flip implies that it could possibly be extra appropriate for different chips than GPUs, reminiscent of CPUs with on-chip accelerators. Intel has such {hardware} with its newest Xeon and Core Extremely server and consumer/edge CPUs. Certainly, inference in some/many circumstances/functions may merely occur on-device as a substitute of on an costly Nvidia GPU within the cloud.

Monetary Dialogue

As a reminder, the Nvidia rally began because of the steering for $11B income within the (calendar) Q2 2023 quarter, up from round $7B. Since then, the income has surged to over $18B in essentially the most just lately reported quarter, with steering for $20B in This autumn, implying a tripling in income in only a few quarters. Given its extraordinarily excessive margins, EPS has equally surged from ~$1 to ~$4 (which has allowed the inventory to quadruple with out altering the valuation a number of).

The truth that principally all this (further) income and revenue is generated in/from the info heart additionally reveals each the info heart income TAM growth in addition to the rise in income market share, as only a few years in the past principally the overwhelming majority of the info heart silicon logic market was captured by Intel.

The ahead estimate is for over $90B in income in 2024. Whereas that is very massive in greenback phrases, not like something even Intel reported when it nonetheless had a CPU monopoly (Intel’s highest was on the order of $30B knowledge heart income, which included its telco networking enterprise), when it comes to models this may correspond to roughly 2 million H100/H200s at a worth of $40k per unit. That is an ASP (common promoting worth) that’s on the order of 40x bigger than both AMD or Intel. This implies Intel will really nonetheless promote much more silicon into the info heart (tens of millions of Xeons every quarter), simply at a a lot cheaper price.

Whereas bulls might even see this disparity in ASP between each as Nvidia’s crowning achievement, for extra sceptical buyers this could current one massive purple flag. As a monopolist, Nvidia enjoys an absolute pricing energy, which has resulted in costs very removed from having any relation to the precise value of products offered of the {hardware}. As talked about, Nvidia may doubtless drop its costs by 10x and nonetheless find yourself with margins in line or above the business. Clearly, promoting 2 million H100/H200s at $4k per unit would offer a really completely different monetary image. Though, as additionally mentioned there could possibly be some worth elasticity concerns in that case, as Nvidia would then in all probability promote much more of these $4k chips.

Valuation

The valuation is a $1.35T market cap. Based mostly on This autumn steering of $20B income and analyst expectation of $4.5 EPS, this represents a valuation of 17x P/S and 30x P/E. This isn’t excessively costly given the supposed additional progress via 2024, catching up from the shortages. Whereas 17x P/S usually needs to be thought of costly, the truth that EPS is just a 30x a number of once more reveals its extraordinarily excessive margins.

Within the image of the general thesis, the primary level right here hasn’t been that the inventory is overvalued, it’s quite the danger that Nvidia is presently capitalizing as principally the one scaled provider of chips powering the LLM/AI explosion. In order extra suppliers (learn: Intel and AMD) ramp their aggressive merchandise, extra wholesome market dynamics ought to kick in, forcing costs downwards and decreasing Nvidia’s market share. Each (for now hypothetical) developments would drastically cut back income and earnings going ahead. It is just then that the inventory would turn out to be costly.

So even when demand for these AI chips/workloads will increase additional within the years forward, in precept confirming the thesis that AI is and stays a progress alternative, this potential situation may however imply that Nvidia’s income could be close to an all-time/long-term excessive. That is much like what has been discussed in the All-In podcast.

In essence, the query is methods to deliver Nvidia’s monopolistic place into the valuation, with a few of the situations starting from sustaining market share and pricing energy (which given the billions of {dollars} Nvidia’s present clients may probably save by switching suppliers appears not possible), to seeing a dramatic decline in each, so one thing in between.

Dangers

In addition to the dangers (for each bull and bear circumstances) already mentioned, on condition that Nvidia has been in provide constraint, a minimum of a few of the future progress is already just about locked in. Nonetheless, this nonetheless doesn’t rule out the potential for a correction in demand later.

Secondly, one of many foremost arguments is that Nvidia doesn’t have a moat since/and chips for AI are principally commodities. Whereas this could be the case to approximation, components reminiscent of inertia may nonetheless stop fast market share swings.

Why Would Chips Be Commodities?

In extension of the earlier threat part, a significant a part of this thesis hinges on the assertion that AI chips ought to to a minimum of some extent be thought of commodities, the place wholesome competitors between a number of distributors is feasible, which clearly has main implications on the doable margins that may be obtained on this market.

This could be a little bit of a extra controversial assertion, since in any case chip design usually takes a few years and tens of millions as much as a whole bunch of tens of millions of {dollars}. As well as, elements ought to as programmability additionally require a growth burden. Certainly, these observations alone clearly point out that it will possibly’t be a pure commodity.

The principle argument is that AI is a extremely parallel workload. It’s principally simply pure (repetitive) math, for which {hardware} functionality is measured in operations per second (OPS). Because of this the bottom unit (the equal of a single core in a multi-core CPU) is a quite simple core, which is then repeated a gazillion instances with the intention to create a big, highly effective chip. So whereas the main points could also be a bit extra sophisticated, there actually is not a lot room for differentiation for the reason that purpose is solely to acquire as many TOPS (Tera-OPS) as doable. As mentioned, each AMD and Intel by now have launched very equally performing chips.

Additionally be aware that this concept of a small core (“cell”) is kind of much like the primary commodity market in semiconductors, which is reminiscence and storage. In reminiscence, a small reminiscence cell is certainly additionally repeated a gazillion instances. Whereas, once more, the main points could also be a bit extra sophisticated (for instance, in NAND there are issues like controllers), there actually is not a lot degree for differentiation, a minimum of not past the method expertise degree.

Total, with reported worth tags starting from $20-40k per chip, even when this consists of issues like HBM and a PCB, within the CPU area not too way back $4-8k (a couple of 5-10x cheaper price) already would have been thought of very costly even for essentially the most high-end chips. Once more, there isn’t any basic expertise in an AI chip that warrants such an extreme premium (in comparison with already premium-priced CPUs). The truth is, as simply argued CPUs are literally extra superior than GPUs/NPUs (since a single CPU core is orders bigger than one GPU/NPU core).

Investor Takeaway

Be fearful when others are grasping. Varied anecdotal experiences preserve surfacing from folks contemplating Nvidia to be cheaper than it was earlier than its rally, and therefore, preserve shopping for Nvidia on the present worth. Nonetheless, with a $1.35T market cap, Nvidia is much from low cost in absolute phrases.

The principle warning is that utilizing ahead estimates as gospel entails severe dangers, and as mentioned concerning Nvidia and the LLM hype, these dangers are very authentic. Observe that threat may merely imply an unknown, and certainly, even earlier than another concerns, there are numerous unknowns concerning the evolution of AI {hardware} demand. Whereas, granted, within the bull case the potential demand for AI computation/{hardware} is just about limitless, there are however numerous financial constraints and realities. With income already within the tens of billions of {dollars}, that is charting into the unknown of what has ever been seen earlier than within the knowledge heart market, by no means thoughts any additional upside.

Nonetheless, that’s precisely what’s required to seize investor-grade returns going ahead. In order mentioned, with authentic competitors from each Intel and AMD significantly ramping up, and no moat (having debunked the CUDA fantasy) that will essentially stop rivals from taking market share, nor any actual doable differentiation that will in precept stop costs from going to the underside, Nvidia’s margins on the finish of the day are just too excessive for this to not be an actual and enormous concern. There is solely nothing in any respect in Nvidia’s chips that warrants the extreme costs/margins that it asks its clients. The one logical expectation is for pure aggressive market dynamics to kick in, to Nvidia’s shareholders’ massive detriment, with the primary indicators of this certainly changing into seen from AMD and Intel.

Buyers who really feel fortunate (as Pat Gelsinger has argued) would possibly purchase Nvidia primarily based on the case of ever-growing demand and miraculously sustaining market share and pricing energy. However any rational and risk-aware investor ought to in all probability keep away from the inventory primarily based on too many unknowns, uncertainties, and dangers of a inventory with a really elevated valuation.