Nvidia - Part 4: Valuation and Returns
Dec 22: Updated to include my financial model at the end of the article
Over the last three articles I explored Nvidia’s moat and the risks it faces in its key segments of gaming and data center. This article is a valuation piece where I present some thoughts on future returns. The folly of trying to forecast growth businesses like Nvidia is not lost on me, however I think it’s important to try support any investment thesis with a view on possible forecast returns. And to quote Jensen himself, “those who guess the most, learn the most”. So in the spirit of being wrong and learning, here we go.
The below chart shows a breakdown of Nvidia’s revenues by its segments. Gaming and Data Center together account for c.90% of Nvidia’s c. $26bn of group revenues. Data Center has recently overtaken Gaming to become the largest segment (over 50% of revenue), and given its double digit growth rate its contribution will only increase going forward. The other two key segments which I haven’t covered are:
Professional Visualization: Provides hardware for graphics rendering, design and digital content creation to enterprises. This segment also includes the much publicized Nvidia Omniverse, its virtual world simulation and collaboration platform for 3D workflows, which is sold as a software subscription for enterprise use and free for individual use. This segment contributes about $2bn of revenues and been growing at a CAGR of about 8% over the last few years
Automotive: Solutions for vehicle cockpits, infotainment and end-to-end hardware/software for autonomous vehicles. This segment is currently a small ($0.6bn of revenue) but a potential growth option for the business. AV is a large opportunity that Nvidia is trying to tap into
As shown on the chart below, not only is Data Center the fastest growing segment, it is also a higher gross margin segment, and this mix shift impact has been the key reason for the growth of Nvidia’s group gross margins from 59% in FY2017 to 65% in FY22.
At the same time operating (EBIT) margins have migrated upwards from the 20-25% range before FY17 to over 30% in FY22.
Returns on Invested Capital
Nvidia’s high profitability has led to high ROICs of over 20-30%, which is indicative of the moat of the business and its asset-light nature. It must be noted that Nvidia has significant cash on its balance sheet ($17bn as of Jul-22), a lot of which is intended for non-operational purposes such as buybacks. As per Michael Mauboussin’s guidance on ROIC, it would make sense to separate out the cash that is intended for operational purposes vs. capital allocation decisions, which is what I’ve done below with a few assumptions. As can be seen, excluding excess cash (and goodwill for that matter) has a material impact on the ROIC metric.
It would be reasonable to assume that ROIC may decline somewhat over time as competition chips away at their business, but even with a steady decline I project for it to stay at above 30% (on an ex-excess cash, ex-goodwill basis) over the forecast period. As I’ll come to in the Valuation discussion, Nvidia’s high ROIC should help drive a more premium multiple.
Below is a summary of my base case forecasts. I will step through each of the key segments in turn.
FY22 (Jan-YE) was a banner year where Gaming revenues grew 61% to reach $12.4bn, driven by the stay-at-home boom in gaming and Ethereum mining. This year in Q2 however, the company announced a sharp 44% decline in gaming revenues, off-the-back of both reopening headwinds and the crash in crypto prices (plus Ethereum moving away from the mining-intensive proof of work to proof of stake). This was coincidentally almost exactly in line with the fall in gaming revenues in the last crypto crash in 2018. I expect that the overall decline for the year will be 30%, before returning to growth with an 8% CAGR from FY23-28. This is lower than the 14% CAGR from 17-23, a period during Nvidia grew slightly faster than the market as it gained incremental market share from AMD (from c. 70% to 80%). Going forward however, I believe there’s a risk that Nvidia may concede some of this market share back to AMD.
Nvidia’s recently announced next gen Ada Lovelace 4080-4090 cards shows a massive performance boost over Ampere especially for the more advanced ray-tracing (RTX) and deep learning super sampling (DLSS) functions. However, this has been receiving quite a lot of pushback in the community due to their high prices and potentially misleading marketing. As Ben Thompson describes here, while Ada’s performance looks impressive, the problem is that gamers are being asked to pay for advanced RTX and DLSS functionality which a lot of games currently just do not use. Pushback against Nvidia’s pricing and marketing strategies are nothing new, but this time around it seems to be more pronounced, and as a result there is a lot anticipation for AMD’s launch of its next gen RDNA 3 cards in a few weeks. More troubling for Nvidia, is that according to semi experts like Dylan Patel (SemiAnalysis), for the first time in a decade Nvidia may have a cost disadvantage vs. AMD. At a high level, this is due to AMD’s RDNA 3 cards having an advanced packaging chiplet design, while Nvidia is still using monolithic chip designs which are much larger in size and more expensive. On top of AMD’s cards being potentially cheaper, they are also likely to have lower running costs due to higher energy efficiency, but we won’t know the exact specs and their relative performance until the official launch on 3 November.
In the past, the strength of the GeForce brand, its performance superiority, along with Nvidia’s more developed software stack, has allowed them to maintain their pricing power and dominant lead over AMD. However, if there is now a structural architectural difference between the two that may cause gamers to drift to AMD, this may not be something that will be a quick fix for Nvidia and as such I believe it is prudent to incorporate this risk into the forecasts. In addition, the weaker macro environment could warrant a lower growth rate over the next few years as well. Overall I believe I’ve been fairly conservative as I’ve assumed it takes five years for Gaming revenues to return to their FY22 peak.
This segment is now clearly the growth driver of the business, having grown at a CAGR of 60% from FY17-23 to reach almost $15bn revenue. Forecasting future growth is a bit tricky however due to lack of transparent market information. What’s important to realise however is that the spend on Nvidia products (namely GPUs) is a subset of the overall data center server spend, which includes not just accelerators like GPUs, but also CPUs, memory, networking equipment, other chips such as FPGAs, ASICs etc. As such, the forecast CAGR of 18% I assumed from ‘23-28 can be benchmarked against a few other data point
On the first data point above, according to the below data from Jefferies, overall data center spend on processors has been growing at a CAGR of 15% over the last ten years, with Nvidia’s growth being much faster as it went from almost nothing in 2016 to capturing a growing share of compute cycles. Jefferies expects overall server spend to keep growing at 10-15% over the next 5-10 years, with the accelerator segment (led by Nvidia) potentially growing faster than as it continues to take share from CPUs due to the growth in compute-intensive AI and HPC workloads.
Another relevant benchmark could be the forecast capex growth of AWS, the largest public cloud provider and Nvidia’s largest customer. Spend on Nvidia products would largely sit as a capex item for the cloud providers. According to Morgan Stanley, AWS’s capex is expected to grow at a CAGR of c. 20% from 22-27, which feels broadly in line with my assumed growth rates for Nvidia. Unfortunately it is difficult to find capex forecasts for the other cloud providers (Azure, GCP) but we could assume it would probably not too different from AWS.
As a final data point, a recent report by Research and Markets estimated the total data center accelerator market was $12bn in 2021 and expected to grow to $55bn in 2027. A couple of other forecasts floating around the internet were close to $60bn. Assuming for a minute that these market forecasts will be right (big if), under a 18% CAGR from 23-28, Nvidia’s data center revenue reaches c.$34bn in FY28, which would imply about a 60% market share. As described in my Part 3 article, currently Nvidia’s share is believed to be somewhere in the 80-90% range. For instance its most recent share of accelerator deployments in public cloud was 84%. So a reduction to 60% in the outer years seems reasonable given the competitive pressures it is facing from its peers like AMD, hyperscalers and their in-house chips (AWS and its Inferentia/Trainium chips, Google and its TPUs), and also more generally the expectation that future technology trends may move into a more heteregenous compute direction where different specialised chips are used for different purposes. I think that this is the area where my forecasts may be most debated, with the bulls claiming these growth rates are too low.
Prof Viz, Auto
I haven’t really explored these segments in depth yet to have an informed opinion but their current combined contribution is small (10% of revenue). I’ve just assumed that they largely retain their historic growth rates of between 8-10% CAGR (together accounts for less than 10% of revenue in the outer years). I recognise that there could be significant upside here if things like Omniverse (which appears to be booked under Prof Viz) and the efforts in autonomous vehicle inflect. It’s fair to say that Nvidia is seeing some early traction in both. In auto in particular Nvidia has built out an impressive platform (e.g., DRIVE Orin, Atlan, Hyperion 8) and established many high profile partnerships (e.g., BMW, Jaguar, Hyundai). These are certainty interesting growth options that I may explore in the future.
Putting all of this together, the group level revenues go from c. $26 in this financial year FY23 to $51bn in FY28, or a CAGR of 14%, which is a significant reduction from the 25% CAGR over the previous five years, which should be expected as the business grows in size. On the margin side, in FY23 both the gross margin and EBIT margin took a large hit due to the significant $1.3bn inventory write-off announced in Q2, so FY24 should see a return to more normalised levels, however from there on I’ve assumed a steady margin decline due to competitive pressure.
As I’ll come to in a bit, I have valued Nvidia on a free cash flow multiple basis, although the street and analysts mostly use P/Es. Given the relatively high cashflow conversion between NPAT and FCF for Nvidia, it would be a reasonable approach to apply a FCF multiple with reference to the company’s traded P/E range. At the current price of $112 it is trading at 32.5x forward P/E. The stock has come off over 65% from its highs last year, and as can be seen below, its current multiple is below its 3, 5 and 7 year averages.
For the purpose of calculating returns I’ve assumed an FY27 exit off the FY28 cash flows, at which point I feel like a contraction to 30x multiple is reasonable. As discussed by Michael Mauboussin in his 2014 paper on understanding P/E multiples, premium multiples tend to migrate lower over time as growth slows and ROIC is competed away, which is what I’ve assumed here. I think a 30x is fair on account of the slower growth in the outer years but still giving credit for the moat and runway of the business and its relatively high ROICs of 30% in those years. Even though competition is likely to chip away at the business, the long term growth themes of AI, HPC, gaming, autonomous vehicles and the other end markets that Nvidia is exposed to should allow reinvestment at reasonably high ROICs (albeit slightly lower than today) for a long period of time, and a valuation should reflect that. It would still be a premium multiple relative to its key peers such as AMD (currently trading at c. 16x P/E)
The below table summarises the cashflows and return calculations. This takes the above earnings projections with some assumptions for tax, capex and working capital. It’s important to note that the cashflows in FY23 have been significantly impacted by the $1.3bn inventory charge and large prepayments of supply agreements (mostly with TSMC), and thus should revert back to a more normal level next year. In regards to the treatment of share-based comp (SBC) which in Nvidia’s case is about 10% of revenue, I have modelled in the associated dilution from SBC through the forecast period, although this is more than offset by the buybacks I've assumed which is why the shares outstanding decreases over time. I then fully included the SBC expense in the final year cashflow that’s used to calculate the exit price, as otherwise you’d be ignoring the fact that SBC would persist into perpetuity.
Assuming a 30x FCF multiple, I derive a share price in 4 years of $175, which is an 11% IRR from today’s share price of $112. Not bad, but in this market we may want to wait for a bit more of a margin of safety. At an entry price of $100 for instance we are getting closer to 15% IRR, which may provide more buffer against things going wrong in the future.
The terminal multiple can obviously be a big point of debate. The bulls could argue that Nvidia should revert back to its prior 40x range in a more normal market environment, and as can be seen in the sensitivity below you would need to believe that would indeed be the case in order to achieve a more attractive double digit return from here. I would think for that to happen however the business would need to really inflect on a new growth driver such as autonomous vehicles, just like it did with data center a few years ago. From where we sit today, I think that is something that is difficult to underwrite.
As explored over the last few articles, I think the moat that Nvidia has established in its core businesses of gaming and data center currently appears robust, however the risks of competition cannot be ignored. As such I feel it is prudent to reflect this in lower growth rates and lower multiple going forward. The stock is down over 65% from its highs, and I believe annualised returns could be north of 10% from here under the base case I’ve spelled out above, however everyone should take these projections with a grain of salt. There is a wide range of outcomes with a business like Nvidia, and I completely understand that readers may take either a more bullish or bearish view. For me personally, I would wait it to drop closer to $100 for a greater margin of safety.
As a fun way to end this, here is a GeForce 4090 card next to an Xbox Series S. It’s a monster.
Dec 22 update: My financial model is included below. If you find any errors let me know.
Some good sources that I used for researching this piece:
Nvidia in The Valley (Stratechery)
Ada Lovelace GPUs shows how desperate Nvidia is - AMD RDNA 3 Cost Comparison – Nvidia’s first cost disadvantage versus AMD in a decade (SemiAnalysis)
Good technical breakdown of Nvidia’s Ada Lovelace cards, misleading marketing and pushback (YouTube)
Why AMD’s chiplets work (Asianometry Youtube channel)
To understand why AMD’s chiplet designs may have a cost advantage I recommend watching Asianometry’s video “Why AMD’s chiplets work” in the links above.
In that 2014 paper on breaking down P/E multiples, Michael Mauboussin outlines some guidelines for calculating firm fair value and P/Es based on the ROIIC, investment size, competitive advantage period and cost of capital. As a rough cross-check, inputting a range of these assumptions for Nvidia into Mauboussin’s fair value formula broadly supports a 30x multiple.
I wanted to discuss this treatment of SBC a bit more as it’s an issue I’ve spent some time thinking about. From what I’ve seen I don’t think many people use this approach of fully including the SBC expense in the earnings that get capitalised for the exit value. A common approach is to dilute the share count in the interim forecasts (as I’ve done) but then just use non-GAAP earnings in the capitalised value at the end, like SBC no longer applies from that point on. Whether you include or exclude SBC in the exit price has a material impact. For example excluding SBC in my analysis, the exit price goes up from $175 to $221 and the IRR increases from 11% to 17% - quite a large difference which may all of a sudden make the returns seem attractive. I personally think that including it is the intellectually honest approach as otherwise you’d be ignoring the fact that SBC is a real expense that will continue to dilute the share count into perpetuity. A simple analogy is that if the compensation structure was changed from SBC to cash bonuses, then it would without question be included in your earnings, so why would SBC be treated any different? I am keen to hear from people on how they think about treating SBC from a valuation standpoint.
Thanks for the article. Read all 4. Enjoyed it so much. Merry Christmas!
Thanks for the work.
I have to admit that I share your scepticism about being able to model NVIDIA. There are simply too many factors that can skew the models.
Demand shift from Cloud Providers, competition, TSMC blow up risk (or even temporary blocks), recession leading to a general slowdown for a couple of years on many levels.
All in all, I simply won't touch this without a huge margin of safety.