NVIDIA’s $3 Trillion+ Valuation Isn’t a Bubble

NVIDIA has joined the elite $3 trillion club, leaving Alphabet, Amazon and Meta in the dust. Many are quick to label it as an AI bubble. This post will show it’s not, based on the company’s exceptional fundamentals. However, there’s never been a company of this scale with so much upside and downside potential.

Fundamentals

Intrinsic value

NVIDIA’s intrinsic valuation is based on the present value of its future cash-flows. But calculating those (i.e. doing a DCF analysis) would require making a ton of assumptions about the future that cannot be tested. Even famed DCF-investor, Warren Buffet, was exposed by his old friend Charlie Munger to not actually do any DCFs. Buffett rather uses DCFs as a mental model, it seems.

Rule of 40 / Rule of X

Luckily there’s a simpler way to see if $3 trillion is justifiable: Bessemer’s Rule of X. That’s basically a regression analysis to see how growth and profitability impact valuation multiples.

Growth & Profitability

NVIDIA grows really, really fast. Its most recent quarter, revenue was up 262% YoY. The current fiscal year it will easily grow 80-100% YoY. Forward-looking, the analyst consensus is that NVIDIA will grow about 25% annually from 2024 to 2026 (calendar years). On top of that, NVIDIA has EBITDA margins of 65%. As a combination of growth and profitability, few companies come close. The typical tradeoff between growth and profitability is clearly visible in the chart below: most companies are scattered in the bottom left of the chart. NVIDIA smashes right through this tradeoff. Microsoft and Meta also perform well.

For the Rule of 40, performance = growth + margins. Bessemer's Rule of X takes the same concept but adds a weighting to growth, called X. The below chart shows what this looks like where X = 3, because the regression line has the best predictive power at around that value (R-square = 70% including the Magnificent Seven and 75% excluding them).

The regression line is like a "Fair value" line. Companies above the line are trading at an implied premium. Companies below the line are trading at an implied discount.

As the chart shows, NVIDIA trades currently at a ±25% premium. But that premium can be explained by the fact that NVIDIA has been consistently beating its guidance, and grows so fast that analysts can barely keep up with adjusting their growth estimates. Other explanations for the 25% premium include hype/momentum (not quite the same as a bubble) and upside potential (more on that later).

And as for tech in general, a BofA chart shared by Azeem Azhar suggests tech stocks on the whole haven't decoupled from fundamentals, unlike during the Internet Boom.

NVIDIA's future revenue growth

Analysts predict NVIDIA's fiscal year 2025 (ending January 2025) revenues will reach approximately $112 billion. Given its huge order backlog, 2025 seems mostly in the bag. Beyond that, NVIDIA is expected to grow 20-25% per year thereafter, reaching $163B by 2027. The 20-25% seems modest compared to its current growth rate, although with hardware sales, volatile demand and supply chain risk, sudden sub-0% growth is always a possibility.

Headroom: Apple vs. NVIDIA

Let's take Apple, another $3 trillion+ company, as a point of reference. Apple does roughly $400 billion in revenue, with EBITDA margins around 35% (vs. NVIDIA's 65% margins). Apple is growing only around 1-5% annually at the moment, basically plateauing. About $200B of Apple's $400 billion revenues are from selling iPhone devices, a saturated market.

There are 5 billion smartphone users, of which Apple dominates the 30% premium segment = 1.5 billion iPhone users. It currently manages to convince ±250 million people each year to buy a new iPhone. It might be able to accelerate Services Revenues (another $100B revenues) through AI, but that's still completely unproven.

NVIDIA on the other hand, is just getting started with AI computing, and the entire AI ecosystem is just getting started. It's currently selling its iPhone 3, as it were. Apple's did $80 billion when it sold the iPhone 3.

The replacement cycle for GPUs is about maybe 3 years, funnily enough this is similar to an iPhone, but heavily influenced by the rate of innovation and seems much faster in the current cycle. Still a lot of unknowns here on the side of NVIDIA.

NVIDIA's top clients: hyperscalers

NVIDIA's top clients as percentage of its revenues are listed below. NVIDIA's run-rate revenue is about $100B today (rounded).

Source: NVIDIA statements for percentages, client names are rumored.

Take Microsoft: growing from $13 billion spend today, to say $20 billion by 2027 sounds entirely plausible. All hyperscalers are planning to aggressively roll out their data centre capacity, according to a year-old Hyperscale Self-Build Report. For reference, 1 GW equates to roughly $10 billion in Capex for the data center itself, and GPUs come on top of that.

That report is more than a year old. Their latest report probably shows a much more dramatic increase (if anyone has it, or similar data, please let me know).

Microsoft has already announced plans for a $100 billion supercluster which aren't included in the above chart. Same applies for all others. Now there are even talks of $1 trillion superclusters. 

With these hyperscalers also comes concentration risk. To avoid too much market concentration, Nvidia is backing alternative cloud providers with venture capital (in the case of Corewave) and allocations of GPUs (hidden). Companies like Coreweave and Vantage are raising billions in debt and equity to accelerate their data center build outs.

However, it's worth noting that each of NVIDIA's major clients is developing proprietary AI chips:

For a detailed exploration of hyperscalers' in-house chip development efforts, see Chris Zeoli's excellent article on the subject.

Other NVIDIA client groups

Governments

Then there are governments. As Nathan Benaich noted "NVIDIA is likely to be the prime beneficiary of the government GPU trade that’s taking off."

Foundational AI model builders

According to an NVIDIA exec: "Big training clusters these days .. at OpenAI, Microsoft and Google have been on the order of tens of thousands of GPUs. 30,000 is a typical number. And people are talking about building hundreds of thousands of GPUs to continue scaling the sizes of the models" (via Tae Kim). For reference, 30,000 new GPUs cost $1 billion, roughly.

There are 230 GenAI model builders on Dealroom, and together they've raised over $35 billion since 2023, per the Dealroom database. Some like X.ai and Anthropic are horizontal/universal model builders. Others like Wayve, which just raised over $1B, are building vertical AI applications with own model. The future might include a network of different big universal and smaller specialist models.

Corporate adoption of AI across verticals

So far, the vast majority of AI investment has been going into the foundational layer. The next phase is in applied AI. Tesla, a major NVIDIA customer, is a great example of this. Soon large corporates like JPMorgan, Eli Lilly, Visa, Novo Nordisk might follow. Across automotive, healthcare, defense, and other verticals the demand for AI is just getting started.

NVIDIA risks

Risk #1: Competition

As mentioned above, NVIDIA faces competition from hyperscalers building their own custom AI chips. Then there are semiconductor companies like AMD and Broadcom, which builds custom chips for Tikitok owner ByteDance.

The word on the street is that Google's TPUs are so far no match for Nvidia's services. Not just in terms of pure performance, but also the overall level of service and ease of use.

When it comes to AI computing, AMD, Intel, and Arm don't yet seem to be anywhere close to NVIDIA. 

OK, now that I’m out, I can finally say this publicly: LOL, no, sorry, you are not catching up to NVIDIA any time this decade.

— Bojan Tunguz (@tunguz) April 16, 2024

Nvidia is also releasing new generation chips at an accelerated pace, as seen with the recent Rubin release. Rubin is not even yet on the below chart.

Source: Linas Beliūnas’ Post

Risk #2: New Entrants

Groq is a scaleup that builds LPUs (Language Processing Units) that provide high-performance at low cost for inference. Its founder Jonathan Ross was responsible for developing TPUs at Google. Cerebras and 100+ other startups & scaleups are also active in the AI chips space. Habana and Graphcore have the scar tissue to show it is not easy to compete with NVIDIA.

One area of high anticipation is energy efficiency combined with high performance. Companies in this area include Vaire Computing (reversible computing), Extropic (thermodynamic computing), SiPearl, Innolight Technology, Immedia PA Semi, Syntiant (Microsoft & Amazon backed). The Dealroom platform includes 50+ high performance / low power chip companies.

These companies are not an immediate threat to NVIDIA, but an exciting addition to the space.

Quantum computing is a dark horse in the accelerated computing space. Companies such as Quantinuum, PsiQuantum, IQM, Pasqal are promising but very far away from being a meaningful threat to NVIDIA. To learn more about this space check out 300+ companies in this Quantum computing landscape.

Risk #3: Energy as a limiting factor

The International Energy Agency projects that global energy demand from data centers, AI, crypto could more than double from '22 to '26 to 114 GW (energy usage of Japan). Will energy become a major limiting factor?

A more aggressive estimate from Altimeter / SemiAnlaysis shows that data centers could soon consume up to 19% of the US national power usage. This comes on top of growing electricity demands from electric vehicles, heat pumps, etc.

All this could be a limiting factor for the industry and for NVIDIA, but could also be a strength if NVIDIA manages to build GPUs that are ahead of others in terms of efficiency.

Risk #4: China/Taiwan tension

Semiconductors are a complex ecosystem with a hugely complex supply chain. For instance NVIDIA is highly dependent on TSMC, and large chunks of the supply chain are embedded in Taiwan, and probably China too. I have no idea what kind of probability to assign to this risk, and over what timeframe.

This probably explains why TSMC is relatively cheap. In contrast, Arm trades at a significant premium (only 10% float, rest still owned by Softbank). It might be less exposed to geopolitical risk, with its royalty-based revenue model.

Pro tip: to compare the metrics and multiples of NVIDIA against its peers Dealroom Multiples, go to the "companies similar to" dropdown menu, and select NVIDIA.

Risk #5: AI running out data / steam

One of the biggest and most real risks is this. The current AI hype is largely founded on excitement about LLMs. Some like Yann LeCun have pointed out that "The beginning of a sigmoid looks like an exponential." Startups like Cartesia and Symbolica are working on new models beyond LLMs. Which other ones do you know? Let us know in the chat!

Dive deeper

All of this is not investment advice obviously.These insights were largely drawn from the Dealroom Magnificent Seven report. This report contains lots of deep links to underlying live data.