Nvidia: The 90% Drawdown That Built the AI Age

It is the autumn of 1996, in an office in California where the lights are kept low to save money on electricity, and Jensen Huang is doing the math on how many weeks of payroll his company has left.

The answer is not many. Nvidia’s first product, a graphics chip called the NV1, has failed. A deal with Sega to power its consoles has collapsed. The company has perhaps nine months of runway, and Huang is about to make the hardest decision of his early career: he will cut the staff from roughly a hundred people to about forty, gambling everything that remains on a single new chip.

He will spend the next several years opening company meetings with a sentence that becomes Nvidia’s unofficial motto: Our company is thirty days from going out of business.

He was not being dramatic. He was describing, with engineering precision, the actual condition of the company. And it matters, because Nvidia would three decades later become one of the most valuable enterprises in the history of capitalism, and the path between those two points runs directly through a stock-market collapse so severe that almost no rational investor could have held through it. Nvidia’s maximum drawdown was a fall of roughly ninety percent. That is the price history charged for the single best-performing stock of its generation.

The World Before Nvidia

In the early 1990s, the graphics card was understood as a peripheral. A useful, secondary piece of hardware that took instructions from the computer’s central processor and turned them into pixels. The brain was the CPU, and giants like Intel owned the brain.

The market for graphics chips was crowded and brutal. Dozens of companies fought over thin margins and short product cycles, and a single mistimed chip could kill a company. It was not, on its face, a market in which anyone expected the most valuable company on the planet to be born.

What no one understood was that the humble graphics chip had a hidden property. Rendering graphics requires performing a very large number of simple mathematical operations all at the same time, in parallel. A CPU does a few complex things very fast, one after another. A graphics chip does thousands of simple things at once. For displaying a game, that parallelism was merely useful. For a kind of mathematics almost no one was doing yet, it would turn out to be the key to the future.

The Founders

In 1993, three engineers founded Nvidia: Jensen Huang, who had been a director at LSI Logic and an engineer at AMD; Chris Malachowsky; and Curtis Priem. The founding meetings took place in a booth at a Denny’s on Berryessa Road in East San Jose, a chain Huang knew intimately because he had worked there as a dishwasher and busboy as a teenager, his first job after his family emigrated from Taiwan.

They raised about $2 million from Sequoia Capital and Sutter Hill at a $6 million post-money valuation, and named the company from the Latin invidia, meaning envy. The first engineers worked out of a townhouse, some without salaries, because the funding had not yet closed.

The founder who matters most for the rest of this story is Huang, and the trait that matters most would not become fully visible for another decade: a temperamental refusal to let the company be only what it currently was. Huang’s defining belief, repeated for thirty years, is that a company that is not reinventing itself is slowly dying at the rate of Moore’s Law. He did not found a graphics-card company and defend it. He founded one and then, repeatedly, bet its survival on becoming something else. The company is the subject of this story, but its serial reinventions are inseparable from one man’s conviction that standing still is the only fatal move.

The Near-Death Moment

Nvidia’s first chip, the NV1, shipped in 1995. It was ambitious and wrong. It rendered graphics using quadrilateral primitives, a technically interesting choice that put it out of step with the triangle-based approach the games industry was rapidly adopting. In Huang’s own later and characteristically blunt words, it was too technically poor to render graphics well.

Worse, Nvidia had bet its early fortunes on a partnership with Sega. When the quadrilateral approach proved incompatible with where the market was going, the Sega relationship fell apart. The company was holding a failed product and had lost the partner meant to carry it.

By the end of 1996, Nvidia was within months of bankruptcy. Two things saved it. First, in an act of unusual grace, Sega’s leadership made a $5 million investment that extended the runway. Second, Huang laid off well over half the company, cutting headcount from roughly 100 to about 40, and concentrated the surviving team on a single new chip, the RIVA 128, built this time around the triangle approach the market actually wanted.

When the RIVA 128 shipped in August 1997, Nvidia had roughly one month of payroll left in the bank. This is the literal origin of the “thirty days from going out of business” mantra. It was not a slogan invented in hindsight. It was the bank balance.

The gamble paid off. The RIVA 128 sold around a million units in four months, pulling Nvidia back from the edge. In 1998 Huang secured the company’s manufacturing future by signing a foundry partnership with TSMC, after a now-famous letter to its founder Morris Chang. In 1999 Nvidia released the GeForce 256, which it marketed as the world’s first “GPU.” Nvidia had survived. And then the stock market did something that should be tattooed on the inside of every investor’s eyelids.

The Drawdown That Defines the Story

Nvidia went public in 1999, rode the dot-com boom up, and then rode the bust down. According to a detailed study by Morgan Stanley’s Counterpoint Global, using split-adjusted daily prices, Nvidia’s stock fell roughly ninety percent from its peak, bottoming on October 9, 2002. Over the same period the index of the thirty largest U.S. semiconductor stocks fell about sixty-five percent. Nvidia fell far more than its own sector, and more than the roughly eighty-five percent median maximum drawdown for all U.S. stocks in the study.

Then came the part almost no one talks about. It took Nvidia 4.1 years, from the October 2002 bottom to November 2006, simply to climb back to its prior peak. Not to make anyone rich. Just to get back to even.

Metric	Figure
Maximum drawdown	~90% (89.7%)
Date of the bottom	October 9, 2002
Time to recover to prior peak	4.1 years (to November 2006)
Semiconductor sector over same period	~65% decline
What came after	Best-performing S&P 500 stock over the 20 years to 2024

Sit with that, because it is the entire lesson rendered as a single verifiable fact. If you had bought Nvidia at its dot-com peak, you would have watched ninety cents of every dollar evaporate, then waited more than four years just to be made whole. And the same Nvidia that inflicted that drawdown went on to be the single best-performing stock in the S&P 500 over the twenty years ending 2024.

The Counterpoint Global study makes the point by pairing Nvidia with Foot Locker. Both fell about ninety percent. Nvidia went on to generational greatness; Foot Locker eventually agreed to be acquired at a fraction of its peak. A ninety percent drawdown tells you almost nothing on its own. It is a doorway that both legends and corpses walk through. The only question that matters is what is happening to the business inside the company while the stock is being destroyed.

The Inflection

In 2006, while Wall Street wanted nothing more than steady profits from gaming chips, Nvidia did something that looked like a baffling waste of money. It released CUDA.

CUDA was a software platform that let programmers use Nvidia’s graphics chips for general-purpose mathematics, not just for drawing games. Huang’s insight was that the massive parallel-processing power Nvidia had built for rendering pixels could be turned to any problem requiring enormous numbers of calculations at once. He did not know exactly which problems. He bet the company’s research budget, billions of dollars over the better part of a decade, on the conviction that such problems would come.

For years they did not, and Wall Street treated CUDA as an expensive distraction that depressed margins for a market that might never materialise. Once again Huang was betting on a market he had to imagine before he could serve it. Insiders call CUDA Nvidia’s “second founding.”

The bet paid off in 2012. A neural network called AlexNet, trained on Nvidia GPUs, shattered the records in a major image-recognition competition and ignited the modern deep-learning revolution. Training neural networks turned out to be precisely the kind of massively parallel mathematics that Nvidia’s chips, and only Nvidia’s chips with their mature CUDA software, were built to do. The market Huang imagined in 2006 arrived, and Nvidia owned the only road into it.

What Everyone Got Wrong

Mistake #1: Treating the 2002 drawdown as a verdict on the business.
Reality: the stock fell ninety percent, but the company kept shipping better chips and building toward the GPU category it had named. The market was pricing terror, not fundamentals.

Mistake #2: Reading CUDA as a waste of money.
Reality: Wall Street saw a decade of depressed margins funding a market that did not exist. It could not see that Nvidia was buying, cheaply and years early, sole ownership of the road into AI.

Mistake #3: Believing AI compute would commoditise like ordinary hardware.
Reality: chips commoditise; ecosystems do not. By the time rivals could match the silicon, the world’s AI software already ran on CUDA, and rewriting it was a cost few would pay.

The Moat

Nvidia’s moat is the deepest kind in technology, because it is two layers that reinforce each other.

The first layer is hardware leadership: Nvidia designs the most powerful AI chips in the world, manufactured by TSMC under the fabless model, and connected into vast “AI factories” using networking technology it acquired by buying Mellanox in 2020. A competitor could, in principle, design a competitive chip.

The second layer is why they cannot easily win anyway: CUDA. Over nearly two decades, essentially the entire global community of AI researchers learned to build on CUDA. The software, libraries, and accumulated knowledge of millions of developers all assume Nvidia hardware underneath. A rival chip that is merely as fast is not enough, because the world’s AI software does not run on it without enormous switching cost. Nvidia built a hardware lead and then poured a software moat around it, and the moat is what makes the lead durable. By the mid-2020s, Nvidia’s data-centre business had grown to roughly ninety percent of revenue, with record annual revenue around $130 billion in fiscal 2025.

The Wealth Created

Imagine you had put $10,000 into Nvidia at its IPO in 1999.

You would have first endured the dot-com round trip: a climb, then a fall of roughly ninety percent that left you with around a thousand dollars of paper value at the October 2002 bottom. You would have waited until late 2006 merely to get your $10,000 back. Many times across the following two decades you would have suffered further gut-wrenching declines, because even great compounding machines fall hard and often. But if you had simply held, you would have owned the best-performing stock in the S&P 500 over the twenty years to 2024. By 2025 Nvidia had at moments been the most valuable company on Earth, briefly carrying a market value in the four-to-five-trillion-dollar range.

The return was historic. And it was only ever available to the investor who could survive a ninety percent drawdown and a four-year wait just to break even, without selling. The number of people who actually did that is, in practical terms, almost zero. That is not a flaw in the story. That is the story.

The Alternative Timeline

A counterfactual, clearly hypothetical.

Picture the world where Huang, chastened by the near-death of 1996 and the ninety percent drawdown of 2002, plays it safe in 2006. He listens to Wall Street, protects his margins, and declines to pour billions into CUDA for a market no one can see.

Nvidia remains an excellent gaming-graphics company. When the deep-learning revolution arrives in 2012, the researchers have no mature platform built for general-purpose GPU computing, and the AI boom is slower, more fragmented, and bottlenecked for years. The generative-AI explosion of the 2020s, when it comes, comes later and messier. The company worth trillions does not exist.

It did not happen that way, because the same founder who survived being thirty days from bankruptcy refused, at the peak of his comfort, to stop reinventing. The lesson is not that boldness always wins. It is that the moats which produce historic returns are almost always built during the periods when caution would have been the popular, respectable choice.

Why This Matters to Investors

Nvidia is the cleanest proof of that thesis available in the public markets, because every part of it is a number you can verify. The improbable idea: a graphics chip that turned out to be a general-purpose mathematics engine. The stretch where failure looked inevitable: thirty days from bankruptcy in 1996, a ninety percent drawdown in 2002. The point where success looks obvious: a multi-trillion-dollar valuation and ninety percent of AI workloads. And the opportunity existed only for those present in the long, dark middle.

The reason to study legendary companies is pattern recognition. Almost every great fortune in market history came from a small number of exceptional businesses that, at the moment of maximum opportunity, looked broken or overvalued or absurd. Nvidia looked broken in 2002 and absurd in 2006. The asymmetry that makes legendary companies legendary, a small chance of an enormous payoff, is only available to those who can survive the stretch where it looks like certain death. The greatest opportunities almost always looked terrible before they looked inevitable.

Lessons in Order of Depth

On the surface — the move

Reinvent before you are forced to. Nvidia survived because Huang treated the company as a thing that must continually become something new: quadrilaterals to triangles, gaming to general computing, chips to AI factories. The trader’s analogue is the refusal to marry a single setup or market regime that has stopped working.

Below the surface — the Money

A ninety percent drawdown is survivable only if you sized and structured your position to survive it. Most people who “would have held” Nvidia are lying to themselves, because they would have been shaken out long before the payoff. The discipline is to participate in asymmetric upside in a way that does not force you to sell at the bottom.

Below that — the Mind

The hardest thing in the Nvidia story is not the engineering of CUDA. It is enduring four years just to get back to even while the crowd calls you a fool. That is the same test a trader faces in a long drawdown: the market offers no encouragement, the thesis is unproven, and quitting feels like wisdom. The ones who compound can tell conviction from stubbornness and hold the first while shedding the second.

At the deepest level — the question

Nvidia and Foot Locker both fell ninety percent. One became the most valuable company on Earth; the other was acquired for a fraction of its peak. The drawdown looked identical. The difference was entirely in what the business was building while the stock collapsed. The deepest lesson: a falling price tells you nothing by itself. The only thing worth knowing is whether the thing being sold off is quietly becoming more valuable or quietly dying, and answering that correctly, against the screaming of the price, is the whole job.

The Legendary Scorecard

Category	Score	Notes
Founder Vision	10 / 10	Saw a general-purpose computer inside a graphics chip
Innovation	10 / 10	Invented the GPU category; CUDA created GPU computing
Execution	10 / 10	From one month’s payroll to relentless architecture cadence
Moat	10 / 10	Hardware lead + CUDA software lock-in; near-total AI share
Capital Allocation	9 / 10	Bet billions on CUDA years before the market existed
Wealth Creation	10 / 10	Best-performing S&P 500 stock over 20 years to 2024
Durability	8 / 10	Dominant, but competition and cyclicality are real risks
Historical Importance	10 / 10	The hardware foundation of the entire AI era
Overall Legendary	9.7 / 10	A defining example of the technology gamble

Scores are an editorial verdict on the standard eight-category scale used across the Greatest Companies series. The overall is a judgment, not a weighted average.

Company Timeline

• 1993 — Founded at a Denny’s; ~$2M from Sequoia and Sutter Hill
• 1995 — NV1 chip fails
• 1996 — Sega deal collapses; staff cut from ~100 to ~40
• 1997 — RIVA 128 sells ~1M units in four months and saves the company
• 1998 — TSMC foundry partnership (letter to Morris Chang)
• 1999 — GeForce 256 marketed as the world’s first GPU; IPO
• 2002 (Oct 9) — Stock bottoms after a ~90% drawdown
• 2006 — CUDA launches, the “second founding”
• 2012 — AlexNet ignites the deep-learning revolution on Nvidia GPUs
• 2020 — Acquires Mellanox for AI networking
• 2025 — Briefly among the most valuable companies on Earth (~$4–5T)

Key Numbers

Founded	1993 (Denny’s, East San Jose)
Founders	Jensen Huang, Chris Malachowsky, Curtis Priem
Early capital	~$2M (Sequoia, Sutter Hill)
Maximum drawdown	~90%, bottoming Oct 9, 2002
Recovery to prior peak	4.1 years
The second founding (CUDA)	2006
Peak market value (2025)	briefly ~$4–5 trillion

Related Reading

More Greatest Companies

• TSMC: The Foundry That Makes the AI Age Possible (Nvidia’s fabless partner since 1998)
• ASML: The Impossible Machine Behind Every Chip (the same Technology Gamble lesson)
• SpaceX: The Company That Bet Everything on Its Last Rocket (the founder’s-bet companion)

Lesson Hubs

• Surviving Drawdowns (the ninety-percent test in full)
• Competitive Moats (how software lock-in outlasts a hardware lead)

Across the Library

• The Dot-Com Bubble (Market Mayhem — the crash that produced Nvidia’s great drawdown)
• Stanley Druckenmiller (Greatest Traders — conviction, and knowing when the thesis still holds)

This article is part of the Greatest Companies series, adapted from the book Greatest Companies, now available on Kindle. Explore the wider framework in The Complete Trader’s Edge.

This article is part of the Greatest Companies series, adapted from the book Greatest Companies, now available on Kindle.

Frequently Asked Questions

When was Nvidia founded and by whom?

Nvidia was founded in 1993 by Jensen Huang, Chris Malachowsky, and Curtis Priem, with the founding discussions famously held in a booth at a Denny’s diner in East San Jose, California.

How close did Nvidia come to bankruptcy?

Very close. After its first chip, the NV1, failed and a Sega partnership collapsed, Nvidia was within months of running out of cash in 1996. It cut staff from about 100 to 40, and by the time its RIVA 128 chip launched in 1997 it had roughly one month of payroll left.

What was Nvidia’s worst stock drawdown?

Nvidia’s maximum drawdown was roughly 90%, bottoming on October 9, 2002, at the tail of the dot-com bust. It then took about 4.1 years just to recover to its prior peak, before going on to become the best-performing stock in the S&P 500 over the twenty years to 2024.

What is CUDA and why did it matter?

CUDA, launched in 2006, is Nvidia’s software platform that lets its GPUs perform general-purpose mathematics. It let the world’s AI researchers build on Nvidia hardware, creating a software ecosystem so entrenched that it became the company’s deepest moat once the deep-learning era arrived in 2012.

Why is Nvidia so dominant in AI?

Nvidia combines the most powerful AI chips, manufactured by TSMC, with the CUDA software ecosystem that nearly all AI software is written for. Matching the chip is not enough for a rival, because the world’s AI code already runs on CUDA, making the switching cost enormous.