Wall Street Refuses to Price This Bull Like the Mammoth It Truly Is

Introduction

Nvidia’s market cap before open on Tuesday, May 26 is $5.21T. The crazy thing is this stock has been a juggernaut in the compute space for a very long time. An embarrassing thing to admit is I only first heard about it back at the start of COVID when exploring one of about 5.21T hobbies that I picked up during the lockdowns and exploring making my own PC — I never did — but that isn’t the point; the point is back in May of 2020 this company was already worth north of $200B, getting a market cap as high as $823B on 11/15/21 during the pandemic. When I started looking up “how to build a PC” I quickly started looking up: “Nvidia”, “What is Nvidia?”, “How does Nvidia make so much money?”, and finally “Who is the CEO of Nvidia?”

The CEO is Jensen Huang. I was and still am fascinated at his story because 1) he went to a state school in the Pacific Northwest — Oregon State — which is atypical to say the least of tech founders, especially of tech founders that have created companies like Nvidia; and 2) he started the company at a Denny’s, a company he worked at for half a decade as a bus boy and waiter back in the day. I hope you continue reading this piece because I swear I am going to get into the technicals in a moment, but would understand if you went and started looking up his backstory if you haven’t heard of it to this point; and if you have some free time you should listen to some interviews that Huang gives. He is insightful in ways that few CEOs are and he is leading the most valuable company in the world into this new era, as the leader of the entire frontier.

I say all this to say I thought I missed the ship during COVID when I looked at the stock chart. Although I kept tabs on it — I almost got in when it was cut in half in fall of 2022 — but by then thought it might have been a “COVID stock” like $PTON (still ~95% off its 2021 peak) or $Z (similar).

I can keep going on and on and on about all the time I looked at this company but it really doesn’t matter; it’s been up and to the right for much of the time since that 2022 reminiscing I just walked you through.

That corresponds with the launch of a new product from a very small regional business, OpenAI; has anyone reading this heard of ChatGPT or artificial intelligence (that’s what the AI stands for in OpenAI I think)!

Since the release of ChatGPT and the corresponding hyperscaler capex spending from the other biggest companies in the world — Microsoft, Alphabet, Amazon, and Meta — Nvidia has gone from being a massive company to being the biggest company.

It has changed the world as we know it, and continues to change the world. The thing is while it 4x’d during COVID from $200B to $800B, that was in large part due to people preparing for a society in the future being more based online; well now it has scaled up more than 25x from that market cap number 6 years ago. And here’s what’s obvious: this is a company that is so clearly being priced lower than it should be. Wall Street refuses to rerate it because it will become so gigantic that I truly don’t know what would happen, but here is what I would say. If you have real confidence in the future of this economy it’s pretty easy to fall in love with this company and make it your entire portfolio. I wouldn’t suggest that but the future of the stock market’s success is really dependent on Nvidia continuing to produce like it has been over the past 3 years.

The Earnings Picture

Two things have happened in three years. The first is that Nvidia’s revenue went vertical. The company finished FY23 at $27B and FY26 at $216B — an 8x increase in three years on a starting base that was already an order of magnitude larger than most companies will ever reach.

The second thing is that the operating leverage on that revenue is unreal. Net income grew from $4.4B in FY23 to $120.1B in FY26 — a 27x increase against an 8x revenue increase. Gross margin runs at 71% on $216B of hardware revenue. Net margin runs at 56%. These are software margins on a hardware business at a scale that has, historically, only ever produced commodity margins for the hardware company in the value chain.

The result is that Nvidia, which six years ago was a $200B graphics-card company that fewer Americans could name than could name Joe Exotic but the rest of the cast of Tiger King, is now the second-most-profitable company on Earth, behind only Alphabet.

The only company ahead of it, Alphabet, is one of Nvidia’s largest customers; so is Microsoft, sitting a few rungs below. Apple, the other household megacap near the top, is comparatively flat on earnings growth. Nvidia is the one accelerating.

Growth at $5 Trillion

Here is the part of the story that has no historical comp.

Apple grew 3%, Microsoft in the mid-teens, Meta and Google at a jarring 28% and 30% respectively. Compare all these massive companies to Nvidia and you’ll see something stand out.

Nvidia grew earnings 145% in FY25. It grew them 65% again in FY26. Both years it was already a megacap. Both years it was already the largest or second-largest semiconductor company in the world by revenue and the most valuable company in the world by market capitalization. Both years it was supposed to slow down because the law of large numbers said it had to slow down. It did not slow down.

We have never before seen this in the history of capitalism — a company this big, growing this fast, this quickly. If I showed you something like this a decade ago, or even three years ago, you’d have told me something crazy happened. Something crazy did happen. The new era is here.

This isn’t hyperbole. The nearest historical analog is Microsoft in the late 1990s — and Microsoft at its absolute peak fiscal year of that era added approximately $4B in incremental net income year-over-year — roughly $7.7 billion in today’s dollars. Nvidia added $47B in incremental net income last fiscal year alone. There is no precedent.

What Makes Nvidia Special

I actually asked Claude — my personal favorite of the AIs — to write the section that follows, because it understands the industry it is built on better than I can. The next 3 paragraphs were generated by Claude Opus 4.7, where I asked it to ‘Describe Nvidia, what it does, and why it does it better than everyone else.’

Modern AI compute runs on a specific kind of processor: the GPU, originally designed for rendering graphics, now repurposed for the matrix-multiplication workloads that underlie modern neural networks. An Nvidia H100 or B200 chip is a parallel-computing engine — tens of billions of transistors arranged as thousands of specialized cores that can multiply large arrays of numbers simultaneously. When a language model generates a sentence, every token requires running an input vector through hundreds of billions of weights, all of which must be multiplied, summed, and re-routed in milliseconds. CPUs do this serially and slowly. GPUs do it in parallel and at speeds CPUs cannot approach. That parallelism is the architectural feature that makes large-scale neural-network inference economically viable.

The unit economics are heavy. A single Nvidia H100 GPU lists at roughly $25,000 to $40,000, with secondary-market premiums driving prices substantially higher during supply-constrained periods. The newer Blackwell-architecture parts — the B200 and the GB200 NVL72 rack — list higher still; a fully configured GB200 NVL72 rack, containing 72 Blackwell GPUs interconnected with high-bandwidth networking, lists at roughly three million dollars. Hyperscalers buy these by the tens of thousands per facility. A single state-of-the-art training cluster, built to train a frontier-scale language model, can require 25,000 to 100,000 of these chips connected with Nvidia networking — a single-facility hardware bill measured in the billions before any data-center construction is counted.

Nvidia’s dominant share — by most estimates, roughly 80 to 85 percent of the AI training-and-inference accelerator market — is not principally a chip-design advantage. It is a software advantage. CUDA, the parallel-computing platform Nvidia has been investing in since 2007, is what the world’s machine-learning engineers actually write code against. Every major model framework, every researcher’s notebook, every production inference stack is built on CUDA primitives. Competing chips — Google’s TPU, Amazon’s Trainium, AMD’s MI300 — must either emulate CUDA at a performance penalty or persuade developers to rewrite their stacks. Both are expensive. The chips matter, but the developer mindshare matters more, and switching costs compound with every new model trained. The moat is nearly two decades of accumulated developer behavior, not silicon alone.

Semiconductors Are Now the Index

The next point about Nvidia that doesn’t compute is the margin structure on hardware revenue at this scale.

Apple sells the most valuable consumer hardware on Earth at a 25% net margin. Samsung’s semiconductor business — the only direct competitive analog at scale — runs in the low teens in good cycles. Intel, which used to define the category Nvidia now occupies, ran at 25-30% net margins in its monopoly era and has since collapsed.

Nvidia runs at 56%. On $216B of revenue. Selling physical objects.

The simplest way to think about it is this: Nvidia’s gross margin reflects a hardware product, but the cost structure underneath it is software-like. The R&D and headcount required to design a chip generation is largely fixed; the marginal cost of each additional chip is small relative to average selling price; the entire revenue line scales without scaling the cost line proportionally. Software businesses have always had this profile — it’s why they trade at premium multiples — but Nvidia has built that profile while shipping silicon, which has historically been a category that thoroughly punishes scale leaders the moment competition shows up.

The 71% gross margin compressed from 75% in FY25, by the way — the Blackwell ramp and tariff exposure pulled gross margin down 4%. Even if the next two fiscal years see another 2% compression, gross margin lands at 69% and net margin somewhere around 50%, which still meaningfully exceeds the entire history of the semiconductor industry’s best margin years.

What Hyperscaler Capex Says About the Future

The demand-side argument is the strongest single piece of evidence in this thesis, and I think it’s underrated by the people who write about Nvidia for a living.

Microsoft, Alphabet, Amazon, and Meta combined for approximately $162B of capital expenditure in 2022. In 2023 the number was essentially flat at $160B. Then ChatGPT happened. In 2024 the four companies spent $251B. In 2025 they spent $416B. For 2026 they have guided to approximately $685B combined.

To put $685B in perspective: in a single year, that is roughly the size of the entire 2008–09 financial-crisis bailout. Congress authorized $700B for TARP — the number seared into a generation’s memory as “the bailout” — and the four hyperscalers are now guiding to spend nearly that much on AI infrastructure in 2026 alone. The difference is who pays. TARP was taxpayer money, marshaled by the federal government to stop a collapse. This is private capital — four companies’ own balance-sheet cash and free cash flow — spent voluntarily, at the scale of a national emergency response, on a bet that AI demand is real. No appropriations bill, no taxpayer backstop, no public vote. A handful of corporations are financing the largest infrastructure buildout of the era essentially out of pocket. Whether that proves the most rational capital deployment in history or the largest voluntary misallocation ever attempted is the whole question — but the sheer scale of it, privately funded, has no precedent.

That is a 4.2x increase in the cash spending of the four most disciplined capital allocators in technology over four calendar years. These are not venture capitalists. These are companies whose CEOs answer to shareholders quarterly and whose CFOs personally feel the pain of every dollar of capex that does not return capital. They are, in aggregate, spending nearly $700B a year on data-center buildout. Where do they think that capex is going?

The answer is mostly to Nvidia. Nvidia’s data-center segment generated $193.7B in FY26 — roughly 30% of total combined hyperscaler capex on a calendar-year basis. The remainder goes to land, power, networking, cooling, real estate, and increasingly the hyperscalers’ own custom silicon — which is a real competitive threat that I engage with directly in the bear case.

But the relevant point here is not what fraction of the capex dollar goes to Nvidia. The relevant point is that the four companies that account for the majority of Nvidia’s revenue are guiding to spend more than four times what they spent four years ago. If you believe that capex commitment, you have a credible forward demand picture for Nvidia for the next several years. If you don’t believe it — if you think the hyperscalers are about to walk back $400B+ of guided 2026 capex — then a lot of investment theses about the AI era need to be revised, and Nvidia is the least of them.

I think the capex commitments are largely real.

What This Stock Is Worth

The valuation question is whether $5.21T is a defensible price for what Nvidia is today, and whether $10T is a defensible target for what Nvidia is likely to be in two to three years.

I’ll argue that $5.21T is reasonable on present numbers — pricing in significant future growth, but not implausibly — and that $10T is more likely than not within 24-30 months based on Street consensus that I think is, if anything, too low.

The base case is the Street’s own number. Consensus has Nvidia at $375.7B of revenue in FY27 and $498.6B in FY28, with EPS of $8.47 and $11.57 respectively. At the current diluted share count of approximately 24.5B, that’s roughly $207B of net income in FY27 and $283B in FY28. At a compressed multiple of 35x — meaningful multiple compression from today’s 43x, reflecting business maturation — the FY28 number alone supports a market cap of $9.9T.

The Street’s own consensus, with a compressed multiple, essentially gets you to $10T. The bull case requires either modest upside to consensus or modest multiple holdup, not both.

The bear case — engaged in its own section below — is real but bounded; the credible downside is sideways trading for two to three years, not a collapse. The sideways trade happened already as investors got spooked last year about accelerated investment from these big tech stocks. Those nerves have seemed to be put to rest…

I will address that as well as a few of the other downside risks that could potentially shake out that I don’t think the company would particularly be at fault over. However best to provide all sides to the possibilities.

A Word For The Bears

Start with what financing costs. A chunk of this buildout is no longer being funded out of free cash flow; it’s being financed (Meta, Oracle, and others are now issuing serious debt to build). Financing only makes sense if the data center on the other end clears the bar it needs to clear to be worth building — and that bar moves with where rates sit. With rates sitting at 4 to 5 percent, every additional data center has to earn its keep against a much higher bar than it would have in, say, 2021. If rates stay higher for longer (and the fiscal picture below and what we are hearing from the Fed suggests they might), the math on the next round of capex likely gets worse, not better.

So where is the revenue, or when will it start pumping? The four hyperscalers are guiding to something like $685B of capex in 2026. The actual revenue customers are paying for AI products today is a small fraction of that. That gap is the whole ballgame. It can close (the bull case is that it closes fast and these turn out to be the best investments anyone ever made), or it can persist long enough that a CFO somewhere blinks, cuts capex, and the rest follow. This was the idea that sent tech down earlier this year, but here is the thing: it seems like the sentiment has changed and Jensen knows it. If one of the hyperscalers stops paying, someone else will step up.

Then there’s the United States balance sheet sitting underneath all of it. U.S. federal debt is north of a truly jaw-dropping $39 trillion, net interest expense now runs ahead of the defense budget, and the bond market’s willingness to keep absorbing supply at current yields is not guaranteed. I’m not predicting a fiscal crisis, but this is probably the closest I have been to sounding the alarm on our fiscal wellbeing as a nation heading for a disaster at some point.

The labor and consumer side is the slower-burning version of the same worry. The productivity case for AI is, in large part, a labor-displacement case (that’s the polite way and the impolite way to say the same thing). If displacement runs ahead of the new demand AI is supposed to create, you get a hole on the consumer side of the economy, and the consumer is what ultimately funds the ad budgets, the software seats, and the cloud bills that pay for the capex that pays Nvidia. Upward mobility stalling and consumer confidence cracking are both very serious social and economic problems. An answer floated for this is an AI dividend, sort of like the UBI that the Yang Gang talked about during the 2020 presidential election. Any UBI program is far, far away in my outlook.

None of this is a knock on Nvidia, and I think these are all unlikely individually, but there is a negative compounding effect that can take place if sentiment changes. That might be the reason this company is still valued where it is and not where it should be. I am biased in my opinion but unbiased in my procedure of research. If you polled 100 random people and asked them what the biggest company in the world is, I wonder how many would say Nvidia; but here’s the thing, if you polled 100 random investors on what the best company in the market is, I don’t think I’d have to wonder what a strong percentage of them would say. In Nvidia, Jensen Huang has created a mammoth, and I can’t believe I am saying a company worth over $5T is cheap.

China + theft of IP. The greater-China region accounted for approximately $19.7B of Nvidia’s FY26 revenue — 9% of the total, down from 21% in FY23. The decline reflects the cumulative effect of U.S. export controls on advanced AI accelerators. Further restrictions would compress this further, and the political environment is not improving.

The most aggressive bear case here is that China revenue goes to zero. That’s a 9% hit to revenue that gets absorbed by less than two quarters of growth at current rates. The risk is real and the trajectory is the wrong direction, but the absolute number is small relative to the trajectory.

There is a wrinkle that cuts the other way. Jensen Huang has, by most accounts, kept a deliberately cordial relationship with Beijing — lobbying against the harshest export controls and continuing to design China-compliant parts rather than walking away from the market — and that pragmatism has limited the damage so far. The real tail risk is not commercial but geopolitical, and it runs through Taiwan. Nvidia designs its chips; TSMC, in Taiwan, makes them. Any Chinese move on Taiwan — blockade, conflict, or an outright attempt to take the island — would threaten the single point of failure in Nvidia’s entire supply chain, with no quick substitute for TSMC’s leading-edge capacity. How concentrated has that supply chain become? Taiwan’s stock market just surpassed India’s to become the fifth-largest in the world, at roughly $4.95 trillion — a nation of about 23 million people, smaller than many single cities, now carrying more market value than a nation of 1.4 billion, almost entirely on the back of the semiconductor industry. We will see whether any future conflict over Taiwan changes the calculus.

The U.S. fails to build out data centers. There is a genuine, bipartisan populist turn against data centers: the right dislikes their ties to Big Tech, the left dislikes their energy and water footprint, and both increasingly show up at the county zoning meeting to say no. Much of it rests on bad information about what these facilities actually consume. So the risk is real enough to name — permitting fights, grid-interconnection queues measured in years, local opposition, power prices that turn ratepayers into opponents. But I don’t think it ends up mattering much for Nvidia, and here’s why: demand for compute is global. If America makes itself too hostile to build, the build doesn’t stop — it relocates. The chips still get bought; they just get installed somewhere outside U.S. control. That is a serious problem for American strategic and economic leadership, and I care about it as a citizen. But for the Nvidia thesis specifically, it’s closer to a routing question than a demand question — the GPUs sell either way. Worth watching, not worth fearing. The thing that would actually break this thesis is demand evaporating, not demand moving.

Custom silicon and the 43x multiple — briefer notes. Hyperscalers are building their own chips (Google TPU, Amazon Trainium, Microsoft Maia) and that is a real but bounded threat to Nvidia’s pricing power — not to its absolute volume, as long as CUDA remains the developer standard. And at 43x last year’s earnings, the stock is expensive; if the growth rate cracks the multiple cracks faster than the earnings do.

Conclusion

This is the most asymmetric large-cap setup I have ever looked at. The price tag is rich relative to what the company earned last year; the growth rate at scale has no historical comp; the demand-side commitments from the customers who pay the bills are guiding to four-times-five-years-ago spending; the margin structure turns every dollar of revenue into more than half a dollar of net income; the balance sheet carries $50B+ in cash and marketable debt securities plus another $22B in strategic equity stakes against $9.5B of debt. Free cash flow in FY26 was $97B. The company is more profitable than every public company except Alphabet, growing faster than any of them, and the customers who pay its bills are guiding to multi-year demand that justifies the trajectory.

Most investment theses are strong in one dimension. Nvidia’s thesis is strong in three or four — earnings scale, growth rate at scale, margin profile, and customer-side capex commitments — with one or two areas of genuine risk that I have engaged with directly above. The downside is contained; the bull case needs no help from the Street’s existing consensus; and the base case lands at a $7.5T to $10T market cap over the next 24-30 months without any multiple expansion.

If you have confidence in the economy and the AI build-out as the consensus narrative, this is not a hard call.

The future is limitless for this mammoth.

view all sources here

Suggested Citation: Rettke, Sterling. “Wall Street Refuses to Price This Bull Like the Mammoth It Truly Is.” sterlingrettke.com, May 27, 2026.