NVIDIA: How a Graphics-Card Company Became a $4.7 Trillion Nation of Compute

Panoramic research report · longitudinal history + cross-sectional rivalry + synthesis Subject: NVIDIA Corporation (NASDAQ: NVDA) Report date: 2026-06-26 · Data cutoff: FY2026 annual report (ended 2026-01-31) + same-day quote Sources: Yahoo Finance (quote / fundamentals / financials / analysts), company annual filings For information and research purposes only. Not investment advice.

Before we start

To understand the $4.74 trillion company NVIDIA is today, you first have to accept something counterintuitive: as recently as late 2022, most people — including plenty of professional investors — still thought of it as "the company that makes gaming graphics cards." It was. But it had also spent fifteen quiet years turning a chip originally built to make Crysis run smoother into the default substrate on which humanity trains artificial intelligence.

The pattern of "badly underrated for a long time, then repriced by the entire world in a single moment" is not rare in business history. What is rare is how NVIDIA stretched that "moment" into a three-year detonation that carried it to the top of the global market-cap rankings. And the part worth studying is not the detonation itself, but the fifteen years before it — the decisions that looked, at the time, like distractions; that Wall Street questioned again and again; and that ultimately decided the outcome.

This report runs along two axes. Longitudinally, we follow the timeline from a sketch three engineers drew in a Denny's booth in 1993 to a company that today ships, every quarter, more data-center silicon than AMD's entire annual revenue. We spend real effort reconstructing the decision logic at each pivotal moment — why it nearly died in 1995, why in 2006 it began burning money on something with no visible payoff, why in 2020 it paid $6.9 billion for a company that, on the surface, made networking cables. Cross-sectionally, we place it back on the 2026 competitive board and ask what AMD, Google, Amazon, Broadcom and Huawei are actually doing to attack it — and why, after three years of assault, its share of data-center AI training silicon has barely moved.

Finally, we bring the two axes together to answer the question a Motley Fool headline posed just two days before this was written — "Why is NVIDIA stock so cheap?" (a $4.74 trillion market cap, yet a forward P/E of only about 15, while it still grows 60%+ a year). That apparent contradiction is itself a key — one that opens the market's deepest ambivalence about this company.

One convention note: NVIDIA's fiscal year ends at the end of January and is named one year ahead. "FY2026" means the fiscal year ended January 31, 2026, which actually spans most of calendar 2025. All figures below use that convention, and per-share numbers are adjusted for the 10-for-1 split of June 2024 so they stay comparable.

Part One · Longitudinal: thirty-three years, three bets

If you had to compress NVIDIA's history into one sentence, it would be this: it has made three bet-the-company wagers, all on the same conviction — that parallel computing would eventually rule the world — and won all three. The first, in the 1990s, was that graphics deserved a dedicated chip. The second, in 2006, was that what a GPU could do reached far beyond graphics (CUDA). The third, around 2020, was that AI would need not chips but entire "factories" of compute. Let's open them one at a time.

1. Origins: a "graphics accelerator" company that nearly didn't survive three years (1993–1999)

1.1 A judgment call in a Denny's booth

In April 1993, Jensen Huang, Chris Malachowsky and Curtis Priem decided to start a company in a booth at a Denny's in east San Jose. The setting later became part of Silicon Valley lore — Huang had waited tables and cleaned restrooms at Denny's as a young man, and thirty years later would choose a booth in the same chain to found a company that reshaped the history of computing.

But the myth tends to obscure the reality of the moment: what these three were betting on was a thesis that was both unproven and brutally contested — that personal computers would eventually need a chip dedicated to handling graphics, rather than dumping all that work on the CPU. In 1993 that was hardly a unique view. There were dozens of companies making graphics chips (then still called "accelerator cards"): 3dfx, ATI, S3, Matrox, Trident, Cirrus Logic — all of them with money, customers and channels. What chance did a three-person startup, launched on a few million dollars from Sequoia and others, have of surviving?

The founders' division of labor would prove crucial: Huang ran business and strategy, Priem was chief technical architect (he'd designed graphics systems at IBM and Sun), and Malachowsky ran engineering. The name NVIDIA traces to the Latin invidia ("envy"), and echoes the "NV" — for "Next Version" — they used to label all their early files.

1.2 The NV1 failure: a bet that was "too clever"

NVIDIA's first product, NV1 (1995), is a textbook failure — and that failure defined the company's character.

The whole industry was converging on a single technical standard: building 3D graphics out of polygons (triangles). Microsoft's forthcoming DirectX was built on the same idea. NVIDIA, however, chose a path that was technically "more elegant" — quadratic surface modeling. In theory, building objects from curved surfaces yields smoother results using fewer resources than stitching together piles of triangles.

The problem was that the entire software world refused to follow. Game developers were already writing for polygons and DirectX; nobody wanted to rewrite their engine for one small company's non-standard approach. NV1 also bundled a sound card, a gamepad port and a grab-bag of other features in an attempt to be an "all-in-one" card — and ended up excellent at none of them, while costing more. It found almost no buyers in the market. Worse, once Microsoft's DirectX shipped, NV1's technical path was effectively sentenced to death as "incompatible."

By 1996, NVIDIA's cash had dwindled to barely a month's runway, and headcount had been cut from over a hundred to around thirty. The company was one step from bankruptcy.

1.3 The cold pivot: RIVA 128 and the art of staying alive

Here we meet the first trait in NVIDIA's character — one that recurs throughout its history: it will bet everything on a long-term direction, but the moment it discovers a specific path is wrong, it will pivot coldly and without sentiment.

A widely cited detail: at the darkest moment, with NV1 dead and the money nearly gone, Huang made two decisions. First, admit the quadratic-surface path was wrong and turn the whole company toward industry-standard polygons and Direct3D. Second — and harder — without the cash to run the normal "test fully, then tape out" process, he bet the company's remaining money on skipping part of the verification and pushing the chip straight to market. It was a decision that, had it failed, would have left no second chance.

That chip was the RIVA 128 (1997). It embraced the Direct3D standard, put both 2D and 3D performance in the top tier of its day, and priced competitively. It sold a million units within four months of launch. The company survived.

Worth noting, too, is the role of Japan's Sega. In the NV1 era, Sega had developed console games on NVIDIA's approach; it later moved to polygons and could simply have abandoned a near-bankrupt supplier. Instead, Sega still paid out a contract — buying NVIDIA precious breathing room. Huang has thanked that "saved by a rival's goodwill" episode many times in public; it became a motif in NVIDIA's culture of facing failure honestly.

1.4 GeForce 256: invent a word, define a category

1999 was NVIDIA's pivot year. In January, it listed on the Nasdaq. Later that year it launched the GeForce 256 and coined a word for it that would later be written into countless textbooks — and into the entire industry's vocabulary: the GPU (Graphics Processing Unit).

The naming was itself a carefully designed positioning move. The prior term of art was "graphics accelerator card" — and that one word, "accelerator," gave away its subordinate status: it was merely a part that offloaded work from the CPU. By naming it a "Processing Unit," parallel to "CPU" (Central Processing Unit), NVIDIA was announcing to the world that graphics processing was not an accessory function of the CPU, but an independent class of computation worthy of its own dedicated processor.

GeForce 256's substantive technical contribution was moving geometry transform and lighting (T&L) — previously the CPU's job — onto the GPU, giving the card the ability, for the first time, to complete an entire graphics pipeline on its own. That was the first step in the GPU's journey from "accelerator accessory" to "independent compute unit" — and the foreshadowing of the day CUDA would turn it into a "general-purpose compute unit."

1.5 What this era was really about

Looking back on 1993–1999, the things NVIDIA actually learned and wrote into its DNA come down to three, and they recur over the next three decades:

1. You can hold the direction, but you must constantly correct the path. It never wavered on the big thesis — graphics deserve a dedicated chip — but it pivoted off NV1's specific technical path the instant it had to. That combination of "stubborn on strategy, flexible on tactics" is the underlying reason none of its later bets crashed.
2. Cadence is a weapon. After RIVA 128, NVIDIA pioneered a "new product every six months" rhythm and used speed to wear down a field of rivals. That "out-iterate the competition" playbook would reappear twenty-five years later, in the form of "one architecture a year," as its core strategy against AMD and custom silicon.
3. Naming is strategy. From the moment it called its product a "GPU," NVIDIA understood that defining a category's name is defining its standard — and whoever defines the standard ultimately takes the most profit.

By the end of 1999, NVIDIA was a successful company — but essentially still a hardware-component supplier. To transform itself, it would need a second bet — one that would burn a decade of patience.

2. The first real bet: turning a gaming card into a "general-purpose computer" (2006–2012)

2.1 An "off-mission" idea

Had the story stopped at selling cards, NVIDIA would have been an excellent company with a limited ceiling. What truly changed its fate was the decision, in 2006, that looked thoroughly "off-mission" at the time — CUDA (Compute Unified Device Architecture).

To see why CUDA matters, you first have to grasp the fundamental physical difference between a GPU and a CPU. A CPU is like a brilliant generalist: a handful of very powerful cores, good at handling complex, ever-changing tasks in sequence — it can do anything, but only a few things at a time. A GPU is the opposite: hundreds or thousands of relatively simple small cores, no single one of them clever, but winning on sheer number — ideally suited to the dumb, heavy work of "doing the same computation a few million times over."

Graphics rendering is exactly that kind of work: millions of pixels on screen, each needing nearly identical color and lighting math. So a GPU is, by nature, a "parallel-computing monster." And Huang, then-chief-scientist David Kirk, and a young researcher named Ian Buck realized something: the world is full of tasks — scientific computing, financial modeling, physics simulation, image processing — that are also, at bottom, "the same computation a few million times over." In principle, they could all run tens of times faster on a GPU than on a CPU.

Ian Buck had built a project called Brook in his Stanford PhD, an early attempt to make GPUs do general-purpose computing. NVIDIA hired him and put him in charge of turning that academic idea into an industrial-grade product.

2.2 What CUDA actually solved

The issue was the barrier to entry. Before CUDA, to do non-graphics computing on a GPU, you had to "disguise" your problem as a graphics problem — pretend your data was pixels and textures, trick the GPU into computing via a graphics API like OpenGL, then translate the "image" results back. The process was profoundly unnatural, and the number of people in the world who could do it fluently was no more than a few hundred. It was like giving the GPU a hardened set of muscles and then handing it a single chopstick to work with.

What CUDA did was give the GPU a toolkit that let you write ordinary C-like programs: you no longer needed to understand graphics; if you could write code and knew how to break a task into "parallelizable chunks," you could call the GPU's thousands of cores directly. It dropped the barrier to general-purpose GPU computing from a few hundred graphics PhDs to the millions of engineers and students worldwide who could write C.

To make this work, NVIDIA made a far-reaching companion decision: from 2006 on, nearly every GPU it made — including the consumer cards sold to gamers — supported CUDA. That meant any university lab, any broke student, could do world-class parallel-computing research in a dorm room on a few-hundred-dollar GeForce gaming card. NVIDIA had, in effect, laid down a vast, free "CUDA developer training network" across the entire world.

2.3 Ten years of burning money — and Wall Street's ten years of doubt

CUDA was a bet that, at the time, had no visible payoff and demanded nearly a decade of sustained R&D and ecosystem investment.

From 2008 to 2015, Wall Street's attitude ran from puzzled to impatient: why would a graphics-card company pour such a large share of its R&D budget into university labs and research institutes that bought hardly any chips? Those customers neither purchased in volume nor offered a clear business model. GPU computing (then called GPGPU) sounded cool, but it contributed almost nothing to NVIDIA's bottom line. In those years, NVIDIA's revenue and stock price were still driven mainly by the gaming-card market and its cyclicality.

Huang's public line through that decade barely changed. The gist was always: we don't know exactly which day, or in what form, parallel computing will erupt, but we are 100% certain that parallel computing is the future of computing, so we will keep investing until it happens. It was a textbook case of betting on a first principle in the absence of any evidence — the kind of wager that either makes you a legend or a cautionary tale.

Worth emphasizing: that persistence was not free at the time. The investment weighed on NVIDIA's margins in those years, and left its cost structure heavier than rivals who made pure gaming chips and carried no CUDA baggage. In other words, before CUDA became a moat, it spent many years as a cost sink.

2.4 The true shape of the moat: not chips, but muscle memory

CUDA ultimately became the foundation of NVIDIA's entire empire — but its value lies not in hardware, nor even in the toolkit itself. Its real value is an almost irreversible form of ecosystem lock-in:

When generation after generation of researchers, engineers and students learn parallel programming with CUDA at the very start of their careers, CUDA becomes their "native language" and muscle memory. University courses teaching GPU computing default to CUDA. The code in academic papers defaults to CUDA. Libraries and tools in the open-source community support CUDA first. By the time those people graduate, join companies and become technical decision-makers, they instinctively — almost without thinking — choose NVIDIA. Not because they compared specs, but because "what I know is this, what my team knows is this, and the whole community's resources are built on this."

That is a moat deeper and more durable than any hardware-performance edge. Hardware can be matched within a generation or two (and AMD, as we'll see, did exactly that), but prying loose a whole generation's muscle memory and a whole software ecosystem's inertia takes not a better product, but a decade-plus and a dose of luck. This is why AMD, later, with hardware half the price, would still struggle to take frontier AI customers from NVIDIA — we'll devote "Rival Four" in the cross-sectional part to this battle over CUDA.

By 2012, CUDA had built a real presence in scientific and high-performance computing circles, but its financial contribution to NVIDIA was still limited. The whole bet looked stuck in the state of "theoretically right, commercially unproven." Until, late that year, something happened in a lab at the University of Toronto that pressed the detonator — and that is the next section's story.

3. The spark: when AlexNet chose two GeForce cards (2012–2020)

3.1 The creation moment: deep learning happened to be born on top of CUDA

In September 2012, two students from Geoffrey Hinton's lab at the University of Toronto — Alex Krizhevsky and Ilya Sutskever (the latter would become OpenAI's chief scientist) — entered a deep neural network called AlexNet in that year's ImageNet large-scale image-recognition challenge. The result was a rout: AlexNet's error rate beat second place by nearly ten percentage points — in a contest usually decided by fractions of a point, that amounted to announcing a changing of the eras.

This is the widely acknowledged "creation moment" of deep learning, the starting gun of the modern AI wave. And for NVIDIA, the most important detail is this: AlexNet was trained on two consumer-grade GeForce GTX 580 gaming cards, via CUDA.

It's worth pausing on the causal chain. Krizhevsky used GPUs not because NVIDIA had designed some "AI chip" — the concept didn't even exist yet. He used GPUs simply because training a neural network requires vast amounts of matrix multiplication (again, the classic "same computation a few million times over"), and CUDA was already there, letting him do on a machine a student could afford what had previously needed an entire server room.

In other words, the six years NVIDIA had spent burning money to lay down the CUDA ecosystem meant deep learning's ignition point happened to land on its chips. There was luck in it — no one could have foreseen that deep learning would erupt this way. But luck only favored the one who had already built the field, turned on the lights, and waited there for six years. This was the first, and decisive, payoff of the CUDA bet.

3.2 Turning luck into inevitability: from "happens to do AI" to "built for AI"

The next eight years were a process of NVIDIA both enjoying that luck and, through a string of deliberate decisions, turning it into "inevitability." It did not sit on AlexNet's good fortune waiting for orders; it systematically — almost aggressively — remade itself into an "AI company." A few pivotal nodes:

2016: hand-delivering the first DGX-1 to OpenAI. That year NVIDIA launched the DGX-1 — a machine integrating multiple top GPUs, an "AI supercomputer" purpose-built for deep learning. A much-retold image: Huang personally delivered the first DGX-1, his signature on the chassis, to what was then a small non-profit lab called OpenAI. The significance runs well past PR — it shows that at a moment when no one had yet seen the direction clearly, NVIDIA was already hand-cultivating, and deeply binding itself to, its single most important future customer base. The people and machines that would later train ChatGPT were on exactly this thread.

2017: the Volta architecture and Tensor Cores. This was the watershed where NVIDIA shifted formally from "a graphics card that happens to do AI" to "a chip designed for AI." Volta was the first to embed hardware units — Tensor Cores — purpose-built for deep learning's core operation (matrix multiply-accumulate). From then on, NVIDIA's GPUs were no longer just "a general-purpose calculator that happened to be good at parallelism"; they began, at the silicon level, to specialize for AI as a specific workload. It deliberately opened a gulf between itself and the "generic GPU" that rivals would struggle to cross.

2018–2019: a "software empire" on top of CUDA. Over these years NVIDIA rolled out a dense series of specialized software libraries and frameworks built on CUDA (such as cuDNN for deep learning and TensorRT for inference). This step was critical: it deepened the CUDA moat from "general parallel computing" to "full-stack optimization of AI computing." To do AI, a team using NVIDIA wasn't just buying hardware — it was buying an entire, deeply tuned, out-of-the-box toolchain from the silicon up to the framework. To compete, a rival no longer just had to build a faster chip; it had to replicate this entire software empire.

3.3 The underrated $6.9 billion: acquiring Mellanox

The acquisition of Mellanox (about $6.9 billion), announced in 2019 and completed in 2020, is the most underrated deal in NVIDIA's history — and the key to understanding why it can today "sell an entire machine room, not just a chip."

Mellanox made high-speed interconnect technology, anchored by InfiniBand. Why does that matter? Because once an AI model grows too big for one chip, or ten, and needs thousands of chips training the same model in concert, how the chips connect — how data flows between them at high speed — becomes as important as the performance of a single chip, if not more so. A cluster of ten thousand top GPUs, if the network between chips is the bottleneck, will leave even the fastest silicon idling and waiting on one another in great numbers.

Acquiring Mellanox let NVIDIA complete the "network between chips" piece of the puzzle in one stroke. Its product form could now level up from "a GPU" to "an entire, efficiently-coordinated GPU cluster." That was the technical precondition for it to later ship the GB200 NVL72 — a full rack that "connects 72 chips into one giant GPU." What Huang saw was that the unit of AI competition was shifting from "chip" to "system" — and that half the system's outcome lay in the network. At the time, many viewed the deal as a strange move, "a graphics company buying a cabling company." Three years on, it reads as the decisive step in NVIDIA's elevation from component supplier to system supplier.

3.4 An interlude: the false boom and bust of crypto

On the road to the AI detonation, NVIDIA also lived through an important — and painful — interlude: cryptocurrency mining.

In the crypto manias of 2017 and 2020–2021, miners frantically bought up GPUs to mine coins like Ethereum that needed heavy parallel computation. This handed NVIDIA wave after wave of seemingly strong demand; gaming cards went out of stock and doubled in price. But that demand had a fatal trait: it was wildly unstable, and it bites back. When coin prices fell, or Ethereum moved to a mechanism that didn't need GPUs, miners would dump their used cards onto the market all at once, instantly puncturing NVIDIA's new-card demand and price structure.

In 2022 that scene played out for real. A crypto winter, layered on a post-pandemic pullback in consumer-electronics demand, blew up NVIDIA's gaming inventory and sank its revenue. FY2023 (ended January 2023) was an ugly year: full-year revenue of $26.97 billion, essentially flat; net income of just $4.37 billion, roughly halved; gross margin down to a 56.9% trough. The stock fell about two-thirds from its high.

This interlude matters for two reasons. First, it left the market with a deeply embedded label — "NVIDIA is a cyclical semiconductor company" whose results swing with some external boom. It was precisely that label that led many to badly underestimate, in early 2023, the nature of the AI demand about to arrive. Second, it is a genuine reminder: NVIDIA's high growth has always been bound to some "external source of demand" — once gaming, once crypto, now AI. The durability of that demand source is forever the question this company most deserves to be asked. We'll raise it again, more seriously, in the synthesis.

3.5 What this era was about: completing the "system supplier" transformation

Across 2012–2020, NVIDIA completed a thorough transformation in product and strategy — but financially the eruption hadn't yet come. As FY2023 closed, it still wore the valuation and label of a "cyclical semiconductor company."

But beneath that label, it had quietly assembled every ingredient the detonation would require:

• The CUDA ecosystem — a whole generation's muscle memory;
• Tensor Cores — silicon-level hardware specialized for AI;
• a software empire of cuDNN/TensorRT and more — full-stack tuning for AI;
• Mellanox networking — the ability to weave chips into a system;
• early, deep ties to OpenAI, Microsoft and Google — the relationships with its largest future customers.

Before November 2022, each of those five cards, viewed alone, was "useful but not decisive." Then a product called ChatGPT lit them all at once, fusing them into a straight flush no one could beat. That is the next section — and the moment NVIDIA's third, and to date largest, bet paid off.

4. The third bet pays off: the "compute detonation" after ChatGPT (2023–2026)

4.1 Three months, and the whole world placed its order

On November 30, 2022, OpenAI released ChatGPT. It crossed a million users in five days and a hundred million in two months, the fastest-growing consumer app in history. But for NVIDIA, the real ignition point wasn't the user count — it was a collective epiphany inside corporate boardrooms and engineering teams everywhere: to build large language models, you had to buy NVIDIA's chips, and the more, the faster, the better.

The epiphany hit nearly every player at once. Microsoft needed compute for OpenAI and its own Copilot; Google had to catch up; Meta was building the open-source Llama; Amazon had to give cloud customers AI capability; and countless startups and even nation-state-scale projects kicked off simultaneously. They all surged toward the same supplier, at once, regardless of cost. Every bit of NVIDIA's fifteen years of groundwork — the CUDA ecosystem, Tensor Cores, the cuDNN software empire, Mellanox networking, the early relationships with OpenAI/Microsoft/Google — converted, in that instant, into a backlog of orders you couldn't get even by paying up. H100s were unobtainable; lead times stretched to half a year or more.

4.2 A financial curve with almost no precedent in corporate history

What happened next has almost no comparison in the history of large companies. Look straight at the numbers (convention: fiscal year ended end-January; USD):

¹ EPS adjusted for the 10-for-1 split of June 2024 to stay comparable.

Read that table closely and several startling facts emerge:

• Eightfold in three years. Revenue went from $27 billion in FY2023 to $216 billion in FY2026 — a full eight times. Net income went from $4.4 billion to $120 billion — twenty-seven times.
• One year's profit exceeds four years of prior revenue. FY2026's net income ($120 billion) is more than four times FY2023's full-year revenue ($27 billion).
• Software-company margins on a hardware company. FY2026 gross margin of 71%, operating margin around 60%, net margin of 55.6%. For a company that "sells chips" to post margins like these is without precedent in semiconductor history. For contrast, traditional semiconductor gross margins typically run 30%–50%.
• From cyclical trough to peak in three years. FY2023 was a 56.9% gross-margin trough, halved profit, the "cyclical" label; three years later it is one of the most profitable companies on Earth.

That 71% gross margin is the single number in this whole report that most deserves to be interrogated. Where does it come from? Not from silicon cost, but from pricing power — and the root of that pricing power runs back to CUDA. When a customer buys NVIDIA, it isn't just buying a chip's compute; it's buying the certainty that "everyone uses this, the ecosystem is the most complete, the odds of something going wrong are lowest, and the engineers you hire already know how to use it." When compute is desperately scarce and model iteration speed is a matter of corporate life and death, that certainty is worth a premium. What NVIDIA collects is, in essence, an "industry-standard tax."

4.3 The cadence war, escalated: "one architecture a year"

At the very moment demand was exploding and it could sell out lying down, NVIDIA made a counterintuitive, deeply aggressive decision: compress the iteration cycle of its data-center chips from the industry norm of two years to one.

From 2024, its roadmap became a clear annual beat: the Hopper architecture (H100/H200, 2022–2023) → Blackwell (B200/GB200, 2024–2025) → Rubin (from 2026). Each generation a large leap in performance and efficiency.

The ferocity of "one architecture a year" lies in this: it uses NVIDIA's own crushing profit and R&D advantage to force every rival to race in a rhythm they can never match. Suppose a rival needs two to three years to design a chip benchmarking the H100; by the time it finally ships and reaches volume, NVIDIA has already launched something a generation or two ahead. The rival is forever chasing a moving target, forever competing with "last year's NVIDIA" while never able to sell at "today's NVIDIA's" price. It is, in essence, the contemporary version of the "new product every six months to wear down rivals" playbook from after RIVA 128 in 1997 — only this time the stakes are the compute standard of the entire AI era.

4.4 From selling chips to selling "AI factories"

NVIDIA's flagship product today is no longer "a chip." Take the GB200 NVL72: it is a full rack that connects 72 Blackwell GPUs, via NVIDIA's own high-speed network (NVLink, descended from the Mellanox thread), into "a single logical giant GPU," complete with liquid cooling, power, switching and the software stack — turnkey.

Huang coined a new term for this — the "AI factory." The naming is, again, a classic positioning move (echoing the coinage of "GPU" in 1999): he is telling customers and the market that NVIDIA sells not parts, but a production facility that can "turn electricity and data into intelligence." An AI factory can be priced at several million, even tens of millions of dollars per rack.

The strategic significance is profound: it lifts the battlefield, wholesale, from "whose single chip is faster" to the system level of "chip + network + full machine + software + cooling + service." Even if a rival builds a chip that matches or beats NVIDIA's on raw performance, it still faces a brutal reality: the customer doesn't want a chip, it wants an entire factory that can go into production immediately — and that factory's networking, software and co-optimization were stacked up by NVIDIA through the Mellanox acquisition and a decade-plus of CUDA. They can't be routed around piecemeal. This is precisely the wall that AMD and custom silicon find hardest to scale, in the cross-sectional part.

4.5 The summit: the most valuable company on Earth

In June 2024, NVIDIA completed a 10-for-1 stock split, pulling the per-share price from over a thousand dollars back to just over a hundred to make it easier for more retail investors to participate — itself a way of riding the "AI for everyone" narrative.

By the time of this writing in June 2026, NVIDIA's market cap is about $4.74 trillion, the most valuable company on the planet. Behind its share price (around $196) is a conclusion stacked up over thirty-three years and three bets:

NVIDIA is no longer "a company that sells chips" but a "nation of compute" that has defined the standard, the cadence and the ecosystem of compute as infrastructure. It defined the language AI computing is written in (CUDA), the hardware it runs on (GPU + Tensor Core), the cadence it iterates on (one architecture a year), and the form it ships in (the AI factory). Inside the standard it has drawn, everyone — including its most powerful customers and rivals — must play by its rules.

But standing at the summit means every direction from here is downhill, and every rival is climbing toward you. The longitudinal history pauses here; next, we swing the camera from the timeline to the cross-section, and look at exactly who, by what means, is trying to breach this seemingly impregnable city of compute in 2026.

Part Two · Cross-sectional: everyone is storming the walls — so why are they still standing?

The longitudinal part told how NVIDIA reached the summit; the cross-sectional part answers a sharper question: standing at the 2026 mark, with so many deep-pocketed, world-class rivals attacking, why is NVIDIA's share of data-center AI silicon still pinned around 90%?

By the panoramic method's read on the competitive scene, NVIDIA is in Scenario C (ample rivals, three or more) — but with a peculiarity: its rivals are less "taking its share" than doing three different things. AMD is fighting it head-on in hardware, trying for a slice. The cloud giants are cutting their own costs with in-house silicon — self-sufficiency. Broadcom and friends are selling shovels to the miners. And Chinese players, forced by circumstance, are building a separate stack from scratch. Each chisels at the city from one direction, but none has broken through the wall built of CUDA + system integration + iteration cadence. Below, we take the five most representative rivals one at a time.

Rival One: AMD — the only one fighting head-on in hardware

5.1 The only rival with a fully matching product line

Of all the rivals, AMD is the only company whose product line maps almost one-to-one onto NVIDIA's. It has data-center GPUs (the Instinct series, from MI300X to MI325X and MI350), its own software stack to counter CUDA (ROCm), an interconnect answer to NVLink — even a leader cut from the same cloth: a bottom-up technologist of fierce execution, the Taiwanese-American engineer Lisa Su (who, in a piece of trivia the media loves for "family duel of the AI era" headlines, is a distant relative of Jensen Huang).

If anyone on Earth can arm-wrestle NVIDIA on the hardware dimension, it is AMD. Over the past decade-plus, by dragging Intel off its throne in CPUs, it proved it has both the script and the execution for "underdog topples giant." So when the AI wave arrived, everyone naturally cast AMD as the "most promising challenger."

5.2 Hardware: the gap is already small, and locally even reversed

First the good news (for AMD). On raw hardware specs, the gap between AMD and NVIDIA has narrowed from "a generation" to "neck and neck."

The Instinct series, led by the MI300X, at one point delivered more high-bandwidth memory (HBM) — a metric that matters enormously for large models — than NVIDIA's same-generation parts; more memory means a single card can hold a larger model and cut the losses from splitting models across chips. On certain inference workloads, the MI300X is genuinely competitive on price-performance. And AMD typically prices a notch below NVIDIA's comparable parts — a real temptation when a compute purchase runs into the billions.

More importantly, AMD has learned the "sell the system" lesson too, rolling out rack-scale answers to NVIDIA's full racks and filling in networking and software capability through a string of acquisitions. It has not made the mistake of "build only the chip, not the system."

5.3 Software: between ROCm and CUDA, still a chasm

But close on hardware is not the same as winning. AMD's real — and hardest to fill — shortfall is in software: the moat we kept stressing in the longitudinal part.

ROCm is AMD's answer to CUDA, and to be fair it has improved fast in recent years. But against CUDA's nearly twenty years of accumulation, it remains a chasm in three places: documentation and maturity (CUDA's pitfalls were long ago paved over by predecessors, its docs exhaustive; ROCm is still catching up), community and ecosystem (the vast majority of the world's open-source AI projects, tutorials and pre-trained models support CUDA first, sometimes only), and stability and coverage (the many new operators and optimizations frontier models rely on are often available on CUDA day one; on ROCm you wait).

What does that mean for a real engineering team? It means migrating a codebase that already runs fine on CUDA over to ROCm often costs extra weeks filling compatibility pitfalls, tuning performance and chasing down odd problems "that wouldn't happen on NVIDIA." When compute is desperately scarce and model-iteration speed directly determines corporate life and death, those weeks of engineer time and uncertainty aren't worth paying for "slightly cheaper hardware," for many companies. That is the most concrete expression of the CUDA moat: not a spec, but an invisible migration cost buried in engineer hours.

5.4 The user's view: what's the real word of mouth

If you scan engineering communities for the real verdict on AMD Instinct, the two highest-frequency lines are roughly: "the price-performance is genuinely great" and "but you have to know what you're doing and be willing to tinker."

Those two lines capture AMD's situation precisely. Its cards suit customers who are strong in engineering, motivated to escape a single supplier, and running relatively standardized workloads — typically giants like Meta and Microsoft, who staff top-tier AI-infrastructure teams. They have the ability to absorb ROCm's migration cost and a strong incentive to cultivate a "second supplier" to check NVIDIA's pricing power. In fact, these players are already deploying AMD cards in real life, mainly for inference (the day-to-day serving after a model is trained), where the ecosystem dependency is comparatively lower.

But for the broad second tier and the startups that lack that engineering muscle, or that chase the fastest possible iteration, AMD is still not the "default." Their instinct remains NVIDIA — not because they compared, but because "everyone uses this, it's easy to find answers when something breaks, and the people we hire already know it."

5.5 Niche and trend: it is a "price anchor," not a "replacement"

Put it together and AMD's niche in the whole arena is clear: it is NVIDIA's most important "price anchor" and "backup," not its "replacement."

Its existence has two real values. First, it gives every large customer a bargaining chip at the table — the mere ability to say "I could buy AMD" gives the customer some leverage in pricing talks with NVIDIA. Second, it genuinely catches the most cost-sensitive slice of the inference market.

But be clear about the nature of its threat: for the foreseeable future, as long as CUDA's ecosystem inertia holds, AMD will struggle to turn from "challenger" into "replacement." Its most realistic threat to NVIDIA is not how much share it takes, but pushing NVIDIA's terrifying 71% gross margin down a few points — by giving customers an "affordable alternative" that limits NVIDIA's ability to name any price it likes. That is real, but far from fatal, pressure on NVIDIA. For AMD itself, the "AI second supplier" position, even at just 10%–20% of the market, is already a huge business — big enough to re-rate AMD's own valuation. That is why the market is willing to give AMD an "AI beneficiary" multiple even though it may never be number one.

Rival Two: the custom-ASIC camp (Google TPU / Amazon Trainium / Microsoft Maia / Meta MTIA) — the real structural threat

6.1 Why this line deserves more of NVIDIA's vigilance than AMD

If AMD storms the city from the front, the custom-ASIC (application-specific integrated circuit) line goes for NVIDIA's jugular — its largest customer base itself.

The logic is almost cruelly simple. NVIDIA's single biggest revenue bloc comes from a handful of hyperscale cloud providers: Microsoft, Google, Amazon, Meta. Together they funnel hundreds of billions of dollars a year to NVIDIA — and pay roughly 70% gross margin for the privilege; that is, of every chip they buy, seven-tenths is NVIDIA's profit. For giants holding some of the world's best chip-design talent, and large enough that any cost saving is worth doing, one thought becomes impossible to suppress: why don't I build the chip myself and keep that 70%?

So they all did. That is the origin of the custom-ASIC camp. It is fundamentally different from AMD: AMD builds chips to "sell to everyone," while the cloud giants build chips to "use themselves and save themselves money." Which means they don't need to win the whole market — they only need to migrate the largest, most standardized slice of their own internal AI workloads off NVIDIA to save astronomical sums. This is erosion from the inside, from the biggest customers, and it is harder to defend against than outside competition.

6.2 The progress: Google TPU has already proven "routing around NVIDIA works"

Within this camp, progress varies widely:

• Google TPU (Tensor Processing Unit) is by far the most mature and persuasive case. Google began self-designing TPUs around 2015 and is now on its sixth or seventh generation. The most important fact: Google's own flagship large model, Gemini, is mainly trained and run on TPUs, with little reliance on NVIDIA. The symbolic weight is enormous — it proved to the world that a sufficiently capable company can build an end-to-end, usable compute path that routes entirely around NVIDIA. Google has even begun renting TPU compute externally via its cloud, probing from "self-use" toward "external supply."
• Amazon Trainium / Inferentia: AWS's in-house training chip (Trainium) and inference chip (Inferentia) are iterating fast, aiming to migrate ever more AI workloads on AWS onto its own silicon and to attract cloud customers with lower prices. Amazon has also invested heavily in Anthropic and pushed it toward Trainium — in effect cultivating a "major customer for its own chips."
• Microsoft Maia: as OpenAI's largest compute provider, Microsoft has every incentive to self-design Maia — its annual compute bill for OpenAI is large enough that any in-house cost saving pays off. But Maia started relatively late and is, for now, more supplement than replacement.
• Meta MTIA: Meta's in-house silicon mainly serves its vast recommendation systems and internal AI workloads — the classic "start with your own best-understood, most standardized workload."

6.3 The three obstacles on this road

Custom ASIC sounds like NVIDIA's nightmare, but it faces three structural obstacles that make it "chronic erosion" rather than "a fatal blow":

Obstacle one: it serves only itself, with no ecosystem. A cloud provider's in-house chip is tailored to its own workloads; it lacks NVIDIA's general ecosystem of "sold to the whole world, optimized by the whole world, supported by mountains of third-party software." Its R&D cost can only be amortized over one company, and it struggles to attract outside developers to optimize for it. Its flexibility is poorer, too — a chip tuned for today's model architecture may not suit the new architecture that appears tomorrow.

Obstacle two: designing one successful AI chip is brutally hard. It needs a top chip-design team, multi-year cycles, and the huge cost of tape-out after tape-out. Google's TPU took nearly a decade to get where it is. Most players' in-house chips can currently cover only the most standardized, least-changing slice of their own workloads (typically inference and recommendation), while the most frontier, most variable, most flexibility-hungry training tasks still go to NVIDIA.

Obstacle three: forever chasing a moving target. This is exactly the power of NVIDIA's "one architecture a year." A cloud provider finally designs a generation benchmarking the H100; by the time it reaches volume, NVIDIA's Blackwell and Rubin are a generation or two ahead. In-house silicon is forever competing with "the NVIDIA of the past," which makes "full replacement" all but impossible on cadence alone.

6.4 The real winner may be the "shovel sellers": Broadcom and Marvell

Here lies one of the most elegant — and most overlooked — judgments in the whole AI-hardware landscape: the truly safe earners on the custom-ASIC track may not be the cloud giants themselves, but Broadcom and Marvell, who design the chips for them.

The cloud giants want in-house silicon, but most of them lack a complete chip-design capability and need specialists like Broadcom and Marvell to supply the core design IP, networking interconnect and engineering services. This forms a classic "sell shovels in the gold rush" niche: whether Google TPU or Amazon Trainium ultimately wins, as long as the "self-designed ASIC replacing NVIDIA" trend runs, Broadcom and Marvell take a cut from every project. So in the capital markets, Broadcom's story in recent years is essentially being re-rated as "the general contractor of the custom-AI-chip wave" — one of the few roles in this siege of NVIDIA that profits no matter who breaches the wall.

For NVIDIA, this means it faces not just a few customers' in-house projects, but a structural rival camp that is continually resupplied by professional arms dealers (Broadcom/Marvell) and will be around for the long haul.

6.5 Niche and trend

Place the custom-ASIC camp back on the whole board: its niche is carving meat from "the edge and the inner layer of the cake" — taking the largest-volume, most standardized, most predictable workloads inside the cloud giants (especially inference). That slice happens to be the "fattest but also most easily standardized-away" part of NVIDIA's profit.

But the trend judgment is clear: for the foreseeable future, this is chronic erosion of share, not disruption. Frontier training, the new-model research that needs fast trial-and-error flexibility, and all the second-tier and startup companies that can't afford or staff an in-house chip team — they will, and can only, return to NVIDIA. In other words, custom ASIC will slowly press down NVIDIA's dominance and growth slope in "inference," but will struggle to shake its standard-setter status in "training" and general frontier compute.

For investors, the signal to watch closely is this: the day self-designed ASIC starts taking real share in "frontier training" — not just inference. That will be the inflection where erosion moves from "the edge" to "the core." We'll put it on the monitoring list in the final synthesis.

Rival Three: Intel — a reminder that "leadership is not forever"

7.1 The laggard

If AMD is an evenly-matched rival and custom ASIC a structural threat, Intel's role in this AI-compute war can basically be summed up in one word: behind.

Intel's data-center AI accelerator (the Gaudi line, from its 2019 acquisition of Israel's Habana Labs) has had a faint market presence for years. It is stuck in the most awkward position: its hardware is competitive with neither NVIDIA nor AMD; its software ecosystem is nonexistent — it has neither CUDA nor anything like ROCm, even as a chaser; and the chip-manufacturing prowess it once relied on, once unmatched, has these years been overtaken by TSMC. With neither a performance edge nor an ecosystem edge, nor the cloud giants' "must-build-our-own" internal motive, Gaudi sits far down customers' purchase lists.

7.2 Intel's real value: a mirror

But Intel earns a section in this report not because it is a threat to NVIDIA — it barely is — but because it is an extraordinarily important mirror.

Don't forget that Intel was once the absolute hegemon of the semiconductor industry — the one that "defined the standard, the cadence, the ecosystem." In the PC and server eras, "Intel Inside" was a synonym for computing, the x86 architecture was an unshakable standard, and its "tick-tock" cadence (a process year, then an architecture year) left every rival in the dust. That is almost a carbon copy of NVIDIA's position in AI today. The ecosystem lock-in, the standard-setting power, the cadence advantage Intel held back then are uncannily similar to NVIDIA's now.

And yet Intel went from peak to laggard in less than a decade. It misjudged the mobile era (missing smartphone chips), stumbled repeatedly on manufacturing process, and — locked in by its own success — reacted slowly to new paradigms. A standard-setter that once looked impregnable lost its throne across one or two paradigm shifts.

7.3 What the mirror shows

Intel's story provides a necessary sobering agent for this report: in semiconductors, there is no unshakable throne; a moat, however deep, can be drained by a shift in the era's paradigm.

How stable is NVIDIA's dominance today? Part of the answer hides in exactly how Intel lost its almost equally solid dominance. Intel's decline came not because a rival built a better x86 chip, but because: (1) the paradigm of computing shifted from PC to mobile and cloud — the shape of demand changed; and (2) it dropped the ball on manufacturing, its former core advantage. Transpose those two onto NVIDIA and you get the two long-term risks it should most fear — the next shift in the computing paradigm (say, some entirely new mode of AI computation that doesn't depend on the current GPU paradigm), and its heavy reliance on TSMC's leading-edge process (an external link it cannot fully control).

This is not to say NVIDIA will repeat Intel's fate — its CUDA lock-in runs deeper than Intel's old x86 lock-in, and its vigilance toward paradigm shifts is higher (Huang himself made his name by betting on a paradigm shift). But the Intel mirror reminds us: in assessing NVIDIA, the real risk has never been "who can build a faster chip," but "whether a new paradigm will emerge that makes the current rules of the game obsolete." With that lens, we should turn to the next section — the moat itself, and how it is being besieged.

Rival Four: CUDA vs everything trying to route around it — the real battlefield

8.1 Not a company, but the moat itself under siege

Strictly, the "rival" in this section is not a single company but NVIDIA's deepest moat — CUDA's dominance — being besieged from several directions at once. This is the single most important, and most worth continuously tracking, variable for NVIDIA's long-term value. Full stop.

Recall the longitudinal conclusion: what makes NVIDIA truly irreplaceable is not the chip, but the fact that "doing AI means defaulting to CUDA" is etched into a generation of engineers' muscle memory. That 71% gross margin is rooted right here. So anything that can weaken the default of "must use CUDA" is going for NVIDIA's jugular.

8.2 The besiegers: abstraction layers, compilers and frameworks

The siege comes from several forces, different in direction but united in aim — to stop AI code being chained to CUDA:

• Higher-level frameworks "masking" hardware differences. Most AI engineers today don't write CUDA directly; they use high-level frameworks like PyTorch. And PyTorch and its peers have worked hard in recent years to support multiple hardware back-ends — in theory, if the framework layer fully smooths over the underlying hardware differences, then a developer writes one set of code that runs painlessly on NVIDIA, AMD, even in-house silicon. That would fundamentally weaken CUDA's lock-in.
• "Intermediate compiler layers" like OpenAI's Triton. Projects like Triton try to let developers write high-performance operators in a higher-level, hardware-agnostic way, with a compiler translating to different hardware. If it matures enough to approach hand-written CUDA's performance, NVIDIA's implicit edge — "only a CUDA engineer can wring out the hardware" — dissolves.
• Industry-alliance "de-NVIDIA" efforts. Some industry alliances, formed by NVIDIA's rivals and largest customers, are jointly pushing open-standard software stacks, with a blunt aim — to cut the whole industry's dependence on a single supplier. They have money, motive, and top engineering teams.

8.3 Why, as of 2026, the wall still stands

These sieges are real and rightly aimed. But to judge the progress objectively — as of mid-2026, the wall is nowhere near breached — for three reasons:

One: inertia is measured in decades. CUDA has accumulated nearly twenty years of optimization, documentation, community and talent. To get a substitute to fully match on "maturity + performance + ecosystem coverage" is not the work of one or two releases; it is an undertaking measured in decades.

Two: NVIDIA does not stand still. It iterates not just hardware each year but its software stack, furiously. When a rival finally approaches CUDA at some abstraction layer, NVIDIA has often already co-optimized CUDA with its latest hardware more deeply, opening the gap again. You are forever fighting "last year's CUDA."

Three: the reality of performance loss. A hardware-agnostic abstraction layer usually trades away some performance. And when compute is money, even a 10%–15% performance hit can be unacceptable to a frontier lab burning billions on compute — it would rather keep eating CUDA's lock-in than give up that last bit of performance.

8.4 Trend judgment: the string to watch for the long term

So this section's trend judgment is: the CUDA moat is being continuously eroded, but slowly, and NVIDIA is actively hedging — no near-term threat. But —

the day migrating a model from NVIDIA to other hardware becomes as painless as switching cloud providers, NVIDIA's 71% gross margin instantly loses its foundation. That is the most fundamental sword hanging over NVIDIA's valuation. It hangs high and has not fallen today, but everyone who cares about NVIDIA's long-term value should treat "the maturity of CUDA alternatives" as the most important string, and watch it, continuously, for the long term. Once that string truly snaps, every moat story told above has to be rewritten.

We'll place this at the very center of the judgment in the synthesis.

Rival Five: China and export controls — a rival the geopolitics is "hand-raising"

9.1 A peculiar "rival"

Listing China as a "rival" dimension is warranted because the competitive logic here is entirely different from the previous four — it is not the outcome of natural market evolution, but manufactured and accelerated by geopolitics.

Over the past few years, the US has steadily tightened export controls on advanced AI chips to China: from initially restricting the top-end H100/A100, to repeatedly lowering the performance red line for exportable chips, to layering on curbs even for "China-special" parts. NVIDIA's response has been to repeatedly design deliberately cut-down "special-edition" chips for the China market (such as the H20 and its successors), tuned to sit just under the control line. But the red line itself keeps moving, putting NVIDIA in an intensely passive position in China: each compliant special-edition chip it designs may be knocked back to square one by the next round of controls.

For NVIDIA's financials, China was once a heavyweight market (contributing a meaningful share of its data-center revenue at the peak). Export controls erode that revenue directly and continuously — a risk variable decided not by the company's operations at all, but purely by policy.

9.2 The real cost: it has hand-raised Huawei

But export controls' deepest harm to NVIDIA is not the China revenue it loses — it is that the vacuum it leaves in the China market is being filled, at unprecedented speed, by a domestic rival: Huawei.

The logic is ironic. Under normal market competition, Huawei's Ascend AI chip, starting from scratch to chase NVIDIA's mature hardware + CUDA ecosystem, would be a near-impossible task — it would face the same "hardware catchable, ecosystem unmovable" bind as AMD, from a lower starting point. But export controls change the game: when Chinese customers can't buy NVIDIA's best chips, NVIDIA's strongest weapon — ecosystem inertia — suddenly stops working. A Chinese large-model company that can't get top CUDA hardware may, rather than use a cut-down special edition, grit its teeth and migrate to Huawei Ascend — since it has to endure inconvenience either way, better to endure one that ships reliably and isn't held hostage by US policy.

In other words, export controls amount to forcibly clearing away Huawei's largest domestic competitor, handing Ascend a window — inside a protected market — to catch up on ecosystem, accumulate customers and iterate products. And Huawei is indeed seizing it, rapidly advancing the Ascend chips and its own software stack (CANN, its answer to CUDA) and integrating its own networking and full-machine capabilities — walking, in effect, a Chinese version of NVIDIA's full-stack path.

9.3 Niche and trend

Place this dimension back on the whole board: for NVIDIA, the China market is a certain near-term revenue loss, and the long-term rise of a parallel rival — cultivated by policy — independent of the Western ecosystem.

The trend judgment has two layers. First, inside China, NVIDIA's share will be continually replaced by Ascend; barring a policy reversal, this market's "outflow" is basically one-directional. Second — and more worth long-term vigilance — if Huawei Ascend matures its ecosystem inside China's vast, protected market, it could in time spill over into other markets sensitive to dependence on US technology (parts of the emerging world, Belt-and-Road countries), forming a "non-NVIDIA, non-Western" compute camp. That would be a long-term, geopolitical-layer challenge to NVIDIA's status as global standard-setter.

But keep a sense of proportion: as of 2026, Ascend still has a clear generational gap behind NVIDIA in hardware sophistication, ecosystem maturity, and — most critically — leading-edge manufacturing capacity, and its rise is still largely confined within China. Its threat to NVIDIA's global base remains "distant and geopolitical," not "present and commercial." But it is a variable that will keep evolving with policy and time, and it belongs on the long-term monitoring list.

Cross-sectional summary table

Before the final synthesis, here are the five rivals side by side:

The right way to read this table is not "NVIDIA leads on everything, so it can rest easy," but to see clearly that it faces a well-divided siege coalition: AMD presses its price, the cloud giants nibble its inference share, Broadcom and friends resupply the besiegers, CUDA alternatives grind at its root, and Huawei builds anew on the geopolitical other side. None can defeat it alone, but together they form a long-term gravity — "NVIDIA's excess profit will, in time, be ground down." The only question is: how fast.

Part Three · Synthesis: standing on a $4.7 trillion summit, where to look?

Lay the longitudinal history over the cross-sectional landscape and three judgments emerge, each clearer — and more important — than either axis alone. This part is not an abridgment of what came before; it twists the two threads into one and reaches new conclusions.

Judgment one: the moat's true body is "the standard," and the standard's natural enemy is "paradigm shift"

The longitudinal history proves, again and again: the fight NVIDIA truly won was never any single chip, but CUDA — begun in 2006, burning money for nearly a decade before it paid off — which turned a whole generation of engineers' muscle memory into the most valuable, invisible line on its balance sheet. Cross-sectionally, the five rivals either close in on hardware (AMD) or are good-enough for self-use (ASIC, Huawei), but none can break through head-on on "ecosystem inertia."

So the first synthesis judgment is: in the near term, what can shake NVIDIA will never be a faster chip, but only "a software abstraction layer maturing to the point that the underlying hardware no longer matters" — the sword hanging in Rival Four (the CUDA siege). And the Intel mirror (Rival Three) reminds us further: a standard-setter's true death never comes from being beaten by a rival under the same rules, but from being overturned by a "paradigm shift that makes the old rules obsolete." For NVIDIA, that shift could come from software (a hardware-agnostic abstraction layer maturing) or from some entirely new mode of AI computation that doesn't rely on the current GPU paradigm. This is its most fundamental — but also its slowest — risk.

Judgment two: the next question isn't "can it sell?" but "who pays for the compute?"

This is the most concrete, and most pressing, judgment from fusing the longitudinal financial curve with the cross-sectional customer structure.

FY2026 revenue has reached $216 billion. To sustain high growth means its customers — chiefly those few cloud giants and large-model firms — must pour in another several hundred billion dollars a year buying compute. And the cross-sectional part tells us those same customers are simultaneously racing to self-design chips (Rival Two) to escape dependence on NVIDIA. Which raises the single biggest suspense in the entire AI investment thesis:

This astronomical capex on chips must ultimately be paid back by the money AI applications actually earn. If, over the next year or two, large-model monetization can't keep pace with compute investment, then once cloud-giant capex growth slows — even from "frenzied growth" to "moderate growth" — NVIDIA will be the first, and the most violently, to feel the chill in the whole chain. Look back at the crypto interlude in the longitudinal part (Section 3.4): NVIDIA's high growth has historically always been bound to the cyclicality of some external demand source — once gaming, once mining, now AI capex. The durability of that demand source is forever the question this company most deserves to be asked. Its prosperity today is, in essence, bound to the whole industry's "confidence in AI's future" — and confidence fluctuates.

Judgment three: the tension in the valuation is exactly the market "pricing in" the two judgments above

This is the last judgment, and the one that gathers everything in.

The Motley Fool headline quoted at the open — "Why is NVIDIA so cheap?" — points at a surface contradiction: a market cap of $4.74 trillion, the most valuable company in the world, yet a forward P/E of only about 15 and a PEG of just 0.61. By traditional standards, a company still growing 60%+ at that multiple is unbelievably "cheap."

But the market isn't dim. That "cheapness" is not an oversight but a pricing — it uses the valuation to spell out, plainly, the worries in Judgments one and two: the market expects NVIDIA's historically rare hyper-growth to be unsustainable, with growth slowing markedly over the next few years (the "who pays for the compute" fear from Judgment two), and the excess gross margin to be ground down, in time, by competition and ecosystem substitution (the moat fear from Judgment one). In other words, NVIDIA's ~15× forward P/E already prices in that "the party will cool."

So the core game on this company is no longer "is it good?" — it is plainly a great company: thirty-three years, three bets, the most profitable on Earth, the deepest moat. The core game becomes three more precise questions: how fast will the cooldown come? how hard? and is today's "looks cheap" price pricing that cooldown in fully enough? That is what to actually watch from atop a $4.74 trillion summit. The optimist says "15× forward for a company like this is a giveaway"; the cautious say "it's 15× precisely because the market knows the E can't keep growing — the denominator could stall at any time." Both are right — they have simply placed different bets on the speed of the cooldown.

Key data snapshot (as of 2026-06-26)

Source: Yahoo Finance. Price and market cap are intraday; financials are FY2026 annual; TTM metrics are trailing twelve months.

A few points worth a second look in this table: (1) debt-to-equity of just 6.6%, ample net cash, $46.3B of free cash flow — an exceptionally clean, exceptionally resilient balance sheet, giving it the means to keep pouring into R&D, buybacks and riding out cycles through any industry cooldown. (2) Beta of 2.2 — its stock swings more than twice the market, meaning it is both the highest-beta name in an AI bull market and the one that falls hardest when sentiment reverses, exactly matching the high-stakes nature of Judgment three. (3) 59 analysts, a mean target of $298.9, roughly 50% above the current price, and not one rating below buy — sell-side consensus is extremely bullish, but remember that in the "party" phase sell-side consensus tends to be pro-cyclical and lagging; it is no margin of safety in itself.

Core risk matrix

What to track (what strengthens / falsifies the judgments above)

Signals that strengthen the thesis (the bull script continues):

• Major cloud providers keep revising capex guidance upward, explicitly pointing at AI;
• Blackwell / Rubin reach volume on schedule, with gross margin holding around 70%;
• No scalable, low-loss CUDA alternative emerges; PyTorch multi-backend still defaults to CUDA in practice;
• Clear, scalable monetization appears on the AI-application side (giving "who pays for the compute" an answer).

Signals that falsify the thesis (watch closely):

• Big-customer capex growth slows sequentially for the first time — the most sensitive string in Judgment two;
• Self-designed ASIC takes real share in frontier training (not just inference) — erosion moving from edge to core;
• A "hardware-agnostic" software layer is suddenly adopted widely by frontier labs, collapsing migration cost;
• Export controls tighten further, or Huawei Ascend's ecosystem begins spilling beyond China;
• Gross margin shows a sustained, non-one-off decline — the most direct thermometer of pricing-power softening.

Synthesis · the one-line close

Longitudinally, NVIDIA spent thirty-three years and three bets turning "compute" from a procured good into an infrastructure whose standard, cadence and ecosystem it defines — it never won on chips, it won on the standard. Cross-sectionally, five kinds of rivals besiege this city of compute from five directions; none can breach it alone, but together they form the long-term gravity that "excess profit will, in time, be ground down." And projecting both onto that ~15× forward P/E, the market has, in fact, already rendered its verdict: this is a great company standing on a great — but possibly near-peak — cycle; its price today is written full of the expectation that "the party will cool." The real disagreement was never about whether NVIDIA is good, but about how fast and how hard the cooldown comes, and whether today's price prices that in enough — a question no one can answer for you, and one this report only undertakes to lay out clearly, variable by variable.

Disclaimer: This report is compiled from public information and third-party market data; all key figures are labeled with their basis and time point, and parts of the historical narrative are reconstructions from public sources that may contain inaccuracies or be out of date. It is for information and research/learning purposes only, and does not constitute investment advice, an offer, a buy/sell instruction, or any judgment on the value of any security. Markets carry risk; make decisions with care, consult a licensed professional, and bear your own risk.

— End —

FAQ

What is the bull case for NVIDIA (NVDA)?

It won the standard, not the chip. CUDA locked in a generation of engineers, giving NVIDIA pricing power — a ~71% gross margin — that rivals' faster or cheaper chips struggle to break. As long as AI compute demand keeps cloud capex rising, NVIDIA captures most of it.

Why does NVIDIA trade at only ~15× forward earnings despite a $4.7T market cap?

Because earnings grew so fast — revenue rose roughly 8× in three years — that the forward multiple compresses. The low multiple is the market pricing in a growth slowdown, i.e. that the party eventually cools.

What are the biggest risks to NVIDIA?

AI capex slowing (the 'who pays for the compute' question), CUDA being replaced by hardware-agnostic software layers, cloud customers' custom ASICs taking inference share, and the stock's high sensitivity to any growth stall.