Jensen Huang and NVIDIA: The Bet That Built an AI Empire

I was not at the diner in San Jose in 1993, and I want that understood. The men who were there have told the story so often, and told it so well, that the napkin has acquired the weight of a founding document. Three engineers, a booth, a pot of coffee, and a decision to build a chip for a market that did not exist. I have told a few tall tales myself, and I know how much a good napkin can be made to carry.

What is not in dispute is the arithmetic. The company that came out of that booth now supplies the machinery that trains nearly every frontier model on earth, and the man in the leather jacket has become the industry’s most reliable showman.

The first decade was nearly the last decade

Here is the part the folklore skips. NVIDIA spent its first ten years going broke in installments. An early chipset shipped with a flaw so expensive that the company had to recall it, take the writedown, and start over on an architecture it had already bet the business on. The founders who had promised a chip a year found themselves explaining why the year would have to be repeated.

Out of that failure came one idea, and it was not complicated: if a graphics processor could be made to run ordinary code instead of only drawing triangles, the chip would be worth more than the pictures it drew. NVIDIA then did the unglamorous thing. It spent years writing the tools and libraries for a market that had not arrived, and kept spending after the first two attempts convinced nobody.

A moat made out of other people’s code

• Fifteen years of libraries and dependencies, most of them written by somebody else.
• Relationships with every major foundry, system builder and cloud provider.
• An architecture cadence that competitors have never matched.
• A developer community trained on these tools since its undergraduate days.

The buyers who would rather not be customers

A rival can build a faster chip. What a rival cannot do before Tuesday is persuade several million programmers to rewrite everything they know. Even so, the largest purchasers of these chips — the hyperscale clouds and the biggest laboratories — have both the motive and the money to design their own. Every one of them is trying, and I would not bet against the effort on principle, having watched a great many monopolies get bored of themselves.

Huang’s answer has been to climb. Networking, systems, software, and now whole racks sold as a single sealed unit rather than as parts a competitor might undercut. It is the same manoeuvre he has run from the beginning: change what you sell before somebody else turns it into a commodity.

The retreat that became the house style

Inside the company the near-death of the 1990s is treated as the origin of its manners. The lesson recorded from it was that a technology firm must be willing to throw away its own architecture the moment the evidence turns against it, and to do so in public, which is the part most boards cannot stomach.

It has been tested repeatedly since. Each generation of products has required customers to rewrite their code, and each time the company accepted the ill will because the alternative was to be overtaken by somebody less sentimental.

One man, one jacket, one terminal value

The risk most discussed in the analyst reports is not competitive but personal. Strategy, culture and public identity are fused to one individual more tightly than is comfortable for a company of this size, and any investor running a terminal-value calculation is quietly pricing the day the jacket comes off the rack for good.

There is a deep bench, and the firm has survived handovers before. But the question will be asked at every earnings call until events answer it. I have outlived several men who were said to be indispensable. It is a distinction only ever awarded posthumously.