NVIDIA Corporation

Let me be honest at the outset: this is the hardest business I have ever tried to appraise, and I will end up admiring it far more than I can bring myself to own it. But let us start simply. NVIDIA designs the computer chips that artificial intelligence runs on — the "picks and shovels" of the AI gold rush. When a company wants to train or run a large AI model, it buys (or rents) NVIDIA chips. Almost everyone does. That is the whole story, and for now it is a magnificent one.

Founded in 1993 in Santa Clara, NVIDIA spent its first fifteen years making graphics cards for video games. Then it made a strange, patient bet — CUDA (see Part II) — that let its gaming chips do general mathematics. A decade later, that bet turned out to be the key to artificial intelligence, and NVIDIA found itself holding the only shovel in a gold rush. In the year ended January 2026 it booked $215.9 billion in revenue and $120 billion in profit — a 63% net margin — and on the trailing twelve months it is already running near $253 billion. It is worth about $4.85 trillion: the most valuable company on Earth.

"The GPU is the engine of the AI revolution." — the bull case in one line. The whole question is how long the engine runs, and who else learns to build one.

Crucially, NVIDIA is fabless: it designs the chips but builds none of them. The actual manufacturing is done by TSMC in Taiwan, with memory from SK Hynix, Samsung and Micron. This keeps NVIDIA astonishingly capital-light — it spends barely 2.6% of revenue on physical plant — while TSMC bears the tens of billions of factory cost. NVIDIA captures the design, the software, and the lion's share of the profit. It is a toll-collector that owns no roads.

You cannot judge where NVIDIA is going without knowing the improbable road it travelled. The story matters because it contains both halves of the investment case: a visionary, patient moat-builder, and a business whose fortunes have lurched violently with cycles more than once.

Two lessons an owner should carry from this timeline. First, the CUDA bet is the finest example of patient moat-building I have seen in technology — Huang spent a decade and untold R&D on something with no obvious payoff, and it became the foundation of an era. That is management of the highest order. Second — and I cannot stress this enough — NVIDIA's demand has repeatedly proven cyclical: revenue fell from FY2019 to FY2020, and profit collapsed 55% in the crypto/gaming bust of FY2023. The company you see today, doubling every year, has been humbled by its own cycles before. History does not have to rhyme. But it often does.

Here is where I must be most honest with you. The what is understandable; the who-wins-in-ten-years is not. Let me walk the machine end to end:

This is the crux of the whole report. Buffett's rule isn't merely "understand what they do" — it is "be able to predict, with confidence, what the business looks like in a decade." I can do that for Coca-Cola. I cannot do it for a semiconductor company on a one-year product cadence whose biggest customers are busy building their own chips. That does not make NVIDIA a bad business — it makes it, for me, an unpredictable one. And unpredictable is not the same as bad; it is the same as hard.

A decade ago NVIDIA was a gaming company. Today it is, overwhelmingly, an AI data-centre company — one segment now dwarfs everything else.

Who pays? A tiny number of enormous buyers — the cloud giants (Microsoft, Amazon, Google, Meta, Oracle) and a wave of "AI factory" builders. This is both the strength (deep-pocketed, urgent demand) and the danger (see Part X: two customers were ~39% of a recent quarter's revenue).

Why are the margins so extreme? NVIDIA earns a 74% gross margin and a 63% net margin — numbers almost unheard of in hardware. Four reasons: (1) a near-monopoly in AI training lets it price like one; (2) customers buy not a chip but a full stack — silicon + networking + software — which is hard to unbundle and shop on price; (3) the fabless model pushes factory costs onto TSMC; and (4) genuine scarcity of leading-edge supply. The server-makers who assemble NVIDIA's boxes earn low-single-digit margins on the same hardware. NVIDIA keeps almost all of the "GPU dollar."

The bull case and the bear case both live here. Buffett's key distinction is between a durable structural moat (a castle competitors cannot storm even with unlimited money) and a temporary technology lead (a head start that can be leapfrogged). NVIDIA's argument is that it has converted the second into the first.

The case that it's a fortress — CUDA. Built continuously since 2006, CUDA is the software layer AI runs on — described as "the operating system of AI." It carries real switching costs and network effects: ~4 million developers, every major framework built on it, universities teaching it, and codebases and talent that are all CUDA-native. Migrating away can cost months of engineering per workflow. Add the full stack (chips + networking + rack-scale systems + software) and a one-year product cadence (Hopper → Blackwell → Rubin) that forces rivals to chase a moving target. This is a genuine, wide moat — the real thing.

The case that it's 'only' a lead. A moat built on being technically ahead is not the same as a captive customer who cannot leave — and NVIDIA's biggest customers are the ones funding the alternatives (Part VI). Software moats in technology have eroded before once a "good-enough" open standard appears (PyTorch already abstracts the hardware beneath it). The moat is thinnest exactly where the market is growing fastest — inference. So: a wide moat today, of genuinely debated durability. I score it an 8, not a 10 — the discount is for the uncertainty, and semiconductors have never been a place where today's king reigns for decades.

The single most important competitive fact: NVIDIA dominates AI training (~80–90% share), where its full stack is decisive. But inference — running models, projected to be the far larger long-run workload — is more cost-sensitive, less tied to CUDA, and better suited to cheaper custom chips. That is where the war will be fought.

The one to watch is not AMD — it is NVIDIA's own customers. Google (TPU), Amazon (Trainium), Microsoft (Maia) and Meta (MTIA), designing with Broadcom and Marvell, are pouring billions into their own chips to escape NVIDIA's pricing. These are captive for now — aimed at internal inference, not sold to outsiders — and analysts think custom silicon might take ~15–25% of the market by 2030, concentrated in that inference tier. But make no mistake: the people who most want NVIDIA's margins to fall are the people writing NVIDIA its biggest cheques. That is an uncomfortable place for any franchise to sit.

If there is one unambiguous strength here, it is the person at the top.

The ownership picture is a genuine positive: the founder still has tens of billions riding on the outcome, which is exactly the alignment I look for. The honest asterisk is the steady drumbeat of insider selling — routine for any company whose stock has risen tenfold, and mostly pre-scheduled, but a reminder that those closest to the business are diversifying away from it, not adding, at these prices.

Purely as financial statements, I have never seen better: 63 cents of every revenue dollar falls to profit, returns on capital sit above 60%, and there is no debt to speak of. One honest wrinkle: free cash flow ($4.90/share) trails reported earnings ($6.57) — but, unlike Microsoft, not because of capex (which is tiny). It's because NVIDIA is tying up cash in inventory and receivables to feed its own explosive growth (inventory sits on the books ~140 days). That is the arithmetic of a company growing faster than it can collect — a good problem, but a reminder that the reported profits are, for now, partly a paper story racing ahead of the cash.

Here is the whole valuation in a sentence: on today's earnings NVIDIA is dear (30×); on next year's estimates it is cheap (22×); on the year after, a bargain (16×). Everything depends on whether those forward earnings arrive. If AI capex compounds as the Street assumes — revenue marching from $216B toward $1 trillion by decade's end — then $200 is a gift. If the boom pauses to "digest," as chip cycles reliably have, those estimates evaporate and 19× sales becomes a trap. You are not really buying a P/E here; you are buying a forecast of permanence.