Nvidia

Nvidia makes the chips that artificial intelligence is trained on. In October 2025 it became the first company in history worth $5 trillion. In 1996 it had forty employees and one month’s money.

Thirty Days From Going Out of Business

For years the unofficial motto inside Nvidia was a sentence rather than a slogan: our company is thirty days from going out of business.

Companies invent that kind of line in hindsight, after they are safe, because near-death makes a better story than steady progress. Nvidia did not. It said it while it was true, and it kept saying it afterwards, which is the more interesting choice.

The reason it stuck is that the firm did not survive its first crisis by being right. It survived because a competitor felt sorry for it.

A Booth at Denny’s, 1993

Jensen Huang, Chris Malachowsky and Curtis Priem sketched the company out in a booth at a Denny’s on Berryessa Road in East San Jose, late in 1992. It was incorporated on 5 April 1993 in Sunnyvale, California, with $40,000.

The bet was that general-purpose processors would never be good enough at drawing three-dimensional graphics, and that a dedicated chip would win. That bet was correct. Almost everything about the first execution of it was not.

The Chip That Bet on the Wrong Shape

Nvidia’s first graphics accelerator, the NV1, rendered images using quadrilaterals — four-sided shapes.

Then Microsoft shipped DirectX, and DirectX rendered using triangles.

That is the entire failure, and it is worth sitting with because it is not a failure of engineering. The NV1 worked. It was simply built around a geometric primitive that the industry’s new standard did not support, and no amount of quality in the silicon could survive that. A company can be technically right and commercially dead at the same time, decided by a standard it does not control.

Sega had contracted Nvidia for the Dreamcast. Nvidia’s technology fell behind, and Sega’s president Shoichiro Irimajiri told Huang personally that the project was cancelled — and then invested $5 million in Nvidia anyway.

Huang has said that money “gave us six months to live.”

It is worth naming that plainly. The most valuable company on earth exists partly because an executive at a firm that no longer makes consoles chose to write a cheque to a supplier he had just fired.

One Month’s Payroll

In 1996 Huang cut the workforce from around a hundred people to about forty, and pointed everything that remained at a single chip built for triangles.

The RIVA 128 shipped in August 1997. By then the company had enough money left for one month of payroll.

It sold roughly one million units in four months.

Nvidia listed on 22 January 1999.

Inventing a Word

Late in 1999 Nvidia shipped the GeForce 256 and called it a graphics processing unit — the first product marketed as a GPU. It moved transformation and lighting calculations off the main processor and onto the card.

Naming a category is an underrated form of power. Before the GeForce 256 these were graphics cards, judged as accessories. Afterwards there was a GPU, sitting alongside the CPU as a second processor with its own job. Nvidia did not only build the thing; it wrote the sentence the industry used to describe it.

The Billion Dollars Nobody Asked For

Here is the decision that actually produced the $5 trillion, and it was made years before anyone could see a reason for it.

A GPU is not really a graphics device. It is a chip that does thousands of simple calculations at the same time, which is what shading pixels happens to require — and also what a great many scientific and mathematical problems require. In the mid-2000s Nvidia decided to make that capability available to ordinary programmers, and spent more than $1 billion building the software to do it. The platform is called CUDA.

At the time this looked close to indefensible. Nvidia sold gaming hardware. Its customers wanted frame rates. Instead it poured a fortune into tooling for academics running simulations, and kept paying for it year after year while the market for that work stayed small.

Then machine learning arrived, and it turned out that training a neural network is exactly the kind of massively parallel arithmetic CUDA had been built to serve. By 2025, Nvidia held more than 80 per cent of the market for GPUs used to train and deploy AI models, and about 92 per cent of the discrete GPU market overall.

The moat is not really the silicon. Competitors can build fast chips. The moat is twenty years of software, libraries and trained engineers who already know how to use it — a switching cost that no rival can buy.

When the World Turned Up

The numbers from the AI build-out are difficult to hold in the head.

Revenue was $5.0 billion in 2016 and $10.9 billion in 2020. By the financial year ending January 2024 it was $60.9 billion. A year later, $130 billion. For the year ended 25 January 2026 it was $215.9 billion, with net income of $120.1 billion.

A company with 42,000 employees earned more profit in a year than most countries collect in tax.

The market followed. Nvidia passed $1 trillion in May 2023, $2 trillion on 1 March 2024, $3 trillion on 18 June 2024 — briefly the most valuable company in the world — $4 trillion on 10 July 2025, and $5 trillion on 29 October 2025. It joined the Dow Jones Industrial Average in November 2024.

It has spent accordingly: an agreement to buy Groq for $20 billion, Hugging Face for $12.9 billion in August 2026, and in September 2026 an agreement to acquire MediaTek for $110 billion, pending regulatory approval. An earlier attempt to buy Arm for $40 billion, announced in September 2020, was abandoned in February 2022 after the UK competition regulator objected.

What Could Still Go Wrong

A profile that stopped there would be an advertisement. Three things are worth stating.

The concentration. Nvidia’s revenue depends on a small number of very large buyers building AI data centres. Those buyers are also, increasingly, designing their own chips. A handful of purchasing decisions could move the whole line.

The circularity. Nvidia has invested in companies that then buy Nvidia hardware — taking stakes in AI firms and chipmakers whose spending returns as its own revenue. That is legal and disclosed, and it also makes demand harder to read from outside.

The fragility of the story. In January 2025 the release of a cheaper AI model from DeepSeek wiped roughly $600 billion off Nvidia’s value in a single day — the largest one-day loss any American company has recorded. Nothing about the chips changed that morning. What changed was an assumption about how much compute the future would need.

That is the honest position on Nvidia. It is a company whose central bet took twenty years to pay and then paid more than anyone forecast — and whose valuation now rests on the world continuing to want a great deal more computation than it currently has a use for.

The men who agreed all this in a Denny’s booth would presumably take the risk.

Common Questions

How This Was Written

Photography is public domain or CC0 and is credited in the site’s image credits. Corrections: commissions@99founder.com.