NVIDIA Is the Best Business in America. That Doesn't Mean You Should Buy It.

NVIDIA closed fiscal 2026 with $96.6 billion in free cash flow, $54.1 billion of net cash on the balance sheet, and 71% gross margins. The crowd is looking for the right moment to sell. I'm looking for the right moment to buy. Neither moment is today, and the reason isn't the one you'd expect.

Spend a day with NVIDIA's 10-K and the numbers start to look unlike anything that ordinarily appears on the same income statement. Seventy-one percent gross margins. Sixty percent operating margins. Ninety-six and a half billion dollars of free cash flow from a company that, a decade ago, most investors thought of as a gaming supplier.

Then you write down the share price, and the enthusiasm has to meet arithmetic.

The Part Ben Graham Would Love

Ben Graham used to teach his students to look at a balance sheet the way a banker looks at a loan application. Not with excitement. With a pen and a calculator. That is where I started.

NVIDIA holds $62.6 billion in cash and marketable securities. It owes $8.5 billion across all short-term and long-term debt. That leaves $54.1 billion of net cash sitting on the balance sheet, more than the market value of many large public companies. If you liquidated every note, every bond, every short-term obligation, you would have enough cash left over to buy a major American airline outright and still have change for a cheeseburger at the Dairy Queen.

Now turn the page and look at what this company earns on a dollar of sales.

Look at hardware businesses going back fifty years, and you will not find many that clear 55 cents of net income and 45 cents of free cash flow on every dollar of revenue. Those are software margins on a physical product. When you see economics like that, one of two things is true. Either the company has a structural advantage that competitors cannot easily replicate, or the numbers are a mirage that a few quarters of competition will wash away. The job is to figure out which.

The Moat Is the Stack

A moat is not a product. A moat is a set of conditions that keeps a rational competitor from taking your customers even when they want to. See's Candies has a moat because people in California associate it with their grandmother's Christmas box, not because the chocolate is better. Coca-Cola has a moat because the distribution system took a century to build. NVIDIA does not have one moat. It has three, and they reinforce each other in a way that very few businesses ever achieve.

Start with the CUDA stack. NVIDIA's first and most-discussed moat is not the silicon. Competent engineers at several companies can design a chip that runs a neural network. The moat is CUDA, and CUDA is not one product. It is a stack. Roughly two decades of layered libraries sit on top of the programming language itself: cuDNN for neural network math, cuBLAS for linear algebra, NCCL for communication between GPUs, TensorRT for inference throughput. Above those sit PyTorch, TensorFlow, and JAX, which is what AI engineers actually work in. For NVIDIA GPU workloads, the performance path commonly runs through CUDA libraries underneath. Think back to 1980, when Bill Gates licensed DOS to IBM. The economic value of that deal was not the operating system code alone. The value was what that code helped compel over the next twenty years as developers wrote applications for Windows, corporate IT departments standardized on Microsoft formats, and Windows became the default PC platform. By 1995, Microsoft's moat was not one product. It was a lattice of dependencies that a rational competitor with a superior product struggled to crack. CUDA has that same lattice quality.

The picture above is the part most price arguments leave out. AMD's ROCm has been in the field for about ten years against CUDA's roughly nineteen, depending on whether you date CUDA from its first introduction or its first public toolkit release. Third-party estimates vary by definition, but the range I found still puts NVIDIA far ahead: roughly 70-95% of AI chips used for training and deploying leading models, with accelerator-revenue estimates commonly clustering around 80-90%. AMD has good chips. AMD does not yet have the lattice. Closing that gap is not a quarter or a year of work. It is a multi-year platform problem.

Then there is the network. The second moat is the one most analysts ignore because it doesn't show up in benchmarks. NVIDIA does not really sell chips anymore. It sells racks. The largest AI workloads now run on systems with hundreds of GPUs talking to each other constantly, and the speed of that conversation determines whether you train a frontier model in two weeks or two months. NVIDIA bought Mellanox in 2020 for nearly $7 billion, which most observers thought was a curious price for a networking company. Six years later, Mellanox is central to the networking portfolio behind NVLink, InfiniBand, and Spectrum-X. NVIDIA says fifth-generation NVLink provides up to 1.8 terabytes per second of bidirectional throughput per GPU. AMD has credible accelerators and can win on raw compute in some workloads. NVIDIA still has a more complete system-level platform. Networking revenue grew 263% year over year in the fourth quarter of fiscal 2026, hitting $11 billion in a single quarter. That is not the line item of a business being commoditized.

And finally, the developer flywheel. The third moat is the one I find most economically interesting. NVIDIA has said the CUDA ecosystem includes more than six million developers and nearly 6,000 CUDA applications. Most of those developers do not work for NVIDIA. They work at Google and Microsoft and Amazon and thousands of startups, and every hour they spend optimizing a model for CUDA is an hour of sunk cost that makes the NVIDIA ecosystem more valuable. The model your researcher downloaded from Hugging Face this morning is more likely to have mature CUDA support than first-class support for a newer custom accelerator. New AI research is more likely to have a CUDA implementation early. NVIDIA does not commission most of this work. The community does much of it because CUDA is what often runs in production. A moat that deepens while the company sleeps is the rarest kind of moat there is.

These three moats compound. The CUDA stack pulls developers in. Those developers write code that benefits from NVIDIA's networked systems. The network is what makes the next generation of training jobs run. Each layer reinforces the next. When a Fortune 500 CTO sits down to evaluate a custom AMD or Amazon accelerator, what she is really evaluating is a multi-quarter migration that may not work as expected and may consume some of her team's best engineers. Some workloads will move. The hard question is how much of the production estate moves, how quickly, and at what opportunity cost.

The Twenty-Year Question

Three moats reinforcing each other is what gives NVIDIA seventy-one percent gross margins on a hardware product. Whether they will still be there in 2046 is the question that decides what I should pay today.

Two scenarios would dissolve this franchise. The first is a fundamental architectural shift: neuromorphic computing, optical accelerators, or some quantum approach that makes general-purpose GPUs the wrong abstraction for the workloads that matter. The second is a coordinated hyperscaler consortium that funds and commits to an open alternative, the way the industry eventually standardized on x86 servers and on Kubernetes. Either could happen. Neither feels close in the next ten years. Over twenty, the probability of one of them maturing is not zero. That uncertainty is one of the inputs that goes into the price I am willing to pay.

A Capable Owner-Operator

Management is the second test after the moat, and it is where most analysis goes thin. I want to know that the person allocating my capital has done it for a long time, has done it through bad cycles as well as good, and has personal money on the line.

Jensen Huang has run NVIDIA since 1993. The latest proxy available showed 3.77% beneficial ownership, including family trusts and foundation shares; even excluding foundation shares for which the proxy says he has no pecuniary interest, the position was worth roughly $170 billion at recent prices. He survived the crypto winter. He survived the Ethereum merge, which ended proof-of-work GPU mining for Ethereum and helped expose the prior gaming-channel correction. He survived an export-control regime that removed China data centre compute revenue from the first-quarter fiscal 2027 outlook. NVIDIA then returned $41.1 billion to shareholders in the most recent fiscal year and ended the year with another $58.5 billion remaining under its repurchase authorization.

That is a capable owner-operator. The executives who show up on the page as a one-line bio are not the ones I worry about.

The Part Where Enthusiasm Meets Price

Here is where I stop sounding enthusiastic.

At the April 20, 2026 close of about $202 a share, this company traded at roughly 41 times trailing earnings and 22 times sales. A ten-year Treasury bond paid you more than the free cash flow yield of the stock. A ten-year Treasury bond does not require you to correctly forecast the cost curve of hyperscaler custom silicon, the success of the Rubin memory stack, or the geopolitical position of Taiwan in 2032. A ten-year Treasury bond simply pays you.

Apply Graham's margin-of-safety discipline to a conservative DCF, using a 9% discount rate and a reasonable glide path on operating margins, and you get a per-share value somewhere near $200. The market is already charging that price. That is not an investment. That is paying full retail for a fine suit you hope will still fit in ten years.

Graham had a parable for this. He called the market an emotional business partner named Mr. Market, who stops by every day with two quotes. One is what he will pay you for your share of the business. The other is what he will sell you his share for. On most days you should ignore him. On some days he is terrified and will sell you wonderful things for half of what they are worth. On other days, like today, he is euphoric and is asking top dollar for every wonderful thing in the window.

You do not have to trade with him every day. You only have to trade when he is wrong in your favor. Today he is not.

The Risks I'm Not Dismissing

Every great business has a way it can fail. The discipline is to name the failure modes before the market does, not after. NVIDIA has five worth taking seriously.

The first is customer concentration. NVIDIA's 10-K shows a small number of direct and indirect customers account for a large share of revenue, and the major hyperscalers are central to the demand story. AWS, Google, Microsoft, and Meta are also funding the development of their own accelerators: Trainium, TPU, Maia, MTIA. None of these has displaced NVIDIA at scale. The alternatives are improving. A reasonable bear case is that custom silicon takes a meaningfully larger share of inference workloads by the end of this decade.

The second is China. NVIDIA's first quarter fiscal 2027 outlook excludes China data centre compute revenue entirely. Whether China comes back is a function of geopolitics, not engineering. I do not know how to handicap geopolitics, and I prefer not to underwrite revenue that depends on it.

The third is execution on Rubin, the next-generation architecture. TrendForce cut its 2026 Rubin share forecast for NVIDIA high-end GPU shipments from twenty-nine percent to twenty-two percent in April, citing HBM4 validation and related supply-chain adjustments. A meaningful slip in the Rubin ramp could give AMD's MI400 family a better opening in the 2027 customer commitments that get locked in this year.

The fourth is Jensen himself. He is sixty-three years old and there is no publicly declared successor. The leather jacket is a brand asset, and brand assets do not transfer cleanly. The day he steps down, the multiple compresses on the announcement.

The fifth is the one nobody likes to discuss. NVIDIA designs the chips. TSMC fabricates them. TSMC is in Taiwan. The probability of an event that disrupts that arrangement in any given year is low. The probability over twenty years is not low. This is a tail risk that doesn't deserve to drive the base case but does deserve to be on the page.

None of these risks invalidates the thesis at $150. All of them are reasons not to pay $202.

The Price I Would Pay

A reasonable margin of safety on a business of this quality is twenty-five percent. Twenty-five percent off a fair value near $200 gets me to $150. At $150, the price-to-earnings ratio drops to about 31 times trailing and the forward multiple lands in the high teens. The free cash flow yield climbs to 2.7%. I still would not call that cheap. I would call it a price at which the story has room to be slightly less than perfect and you still make money.

Ted Williams, in The Science of Hitting, divided the strike zone into seventy-seven cells, each the size of a baseball. He swung only at pitches in his best cells and let the rest go by. Many of those rejected pitches were strikes. Many of them, swung at, would have produced base hits. He let them go anyway, because his career average depended on swinging at the right pitches, not at every pitch. Investing is the same game with one critical advantage: there is no called third strike. You can stand at the plate and let pitches go by for years if that is what discipline requires. There are tens of thousands of publicly traded companies in the world. NVIDIA at $202 is a strike on the outside corner. You do not have to swing.

Does NVIDIA ever reach $150? I have no idea. Sir John Templeton warned that the most dangerous words in investing are "this time is different." A companion danger is assuming a great company can never come back to a fair price. In October 2022 this stock, split-adjusted, was trading below $15. Few investors wanted it. The same founder and platform assets were there. The price was ninety percent lower.

Great businesses do not stop being great. They do stop being bargains. They then go on sale again, usually when something that looks permanent turns out to be temporary. Patience is the only competitive advantage a retail investor has that a Wall Street analyst cannot replicate. Use it.

What This Means For You

If you already own NVIDIA, the case for holding it is straightforward. The business has never been more profitable, the moat has never been wider, and the management team has never had more capital to allocate. Selling because the stock is at $202 instead of last year's high is selling a great business because the price has come down. That is Mr. Market's game and you do not have to play it. Trim into strength if the position has grown past its natural weight in your portfolio. Set a stop twenty percent below your last trim, not because the fundamentals will break, but because the multiple might.

If you do not own NVIDIA and you have been waiting, keep waiting. Park a limit order at $150 and stop watching the screen. If you are uncomfortable owning zero, put a small starter tranche in around $170 and add only on weakness. The full position belongs at $150 or lower.

If you are an aggressive value investor, the price below which the discount becomes interesting in its own right is $120. That is below the bear-case fair value. At that price, something would have to be genuinely broken. If you cannot articulate what is broken, the market is telling you it knows something you do not. Do the work before you write the check.

The Last Thing

The price you pay determines the return you earn. Today's price assumes Jensen executes flawlessly for a decade. He may. He may not. Either way, $202 doesn't pay you to find out. $150 does.

Twenty years from now, NVIDIA will either be the operating system of artificial intelligence, the way Microsoft became the operating system of personal computing, or it will be a cautionary tale about what happens when a single vendor tries to own an entire industry. The first outcome is more likely than the second. That isn't enough certainty to pay full retail for the privilege of being right.

The pitches keep coming. Wait for yours.