Jensen Huang · Free lesson guide

Why CUDA won: a bet on programmability and scale

In 2006 NVIDIA started adding a general-purpose programming layer to graphics cards sold to gamers — spending die area, engineering budget and gross margin on a capability almost no paying customer had asked for. Investors objected for years. That decision, not any single chip, is why the company sits at the centre of the AI economy today, and the reasoning behind it is transferable well beyond semiconductors.

Platform strategyAI infrastructureTechnology strategy

What you'll learn

What CUDA actually is and what it changed about using a GPU
Why NVIDIA subsidised developers before any market existed
How accumulated third-party code becomes a switching cost competitors cannot buy past
Why backwards compatibility across generations compounds the advantage
How to tell a real platform bet from an expensive feature

Key concepts

Programmability: Before CUDA, using a GPU for non-graphics work meant expressing your problem as a graphics operation — theoretically possible, practically hostile. CUDA let developers write parallel code in a familiar language, which converted specialist hardware into a general computer.
The installed-base bet: By shipping CUDA on ordinary consumer cards rather than an expensive professional line, NVIDIA put a supercomputing-capable device on the desk of every student and researcher. The market did not exist, so it manufactured the population that would create one.
Switching cost as accumulated code: Fifteen years of libraries, frameworks, tutorials, university courses and trained engineers now sit on top of CUDA. A competitor must match not the chip but that entire body of work, which no amount of capital produces quickly.
Generational compatibility: Each new architecture ran existing CUDA code. That turns every past adopter into a likely future buyer, and it means the ecosystem's value grows monotonically instead of resetting.

The bet, stated honestly

NVIDIA committed sustained investment to a capability whose customers were hypothetical. The internal logic was conditional: if any large class of workloads ever became massively parallel, the winner would be whichever architecture developers already knew — not whichever was fastest at the moment the shift occurred. That is a bet on the durability of human skill and existing code rather than on silicon performance, and it is why the eventual AI boom found NVIDIA already holding the position.

Why the moat is software, not transistors

Competing accelerators have repeatedly matched or exceeded NVIDIA on raw specifications. They still struggle, because a research team choosing hardware is choosing whether their existing code runs, whether the framework they use is supported on day one, and whether they can hire people who already know the toolchain. The performance comparison happens after those questions, and frequently never happens at all. Platform advantage means the customer's decision is made before your product is evaluated.

What generalises

Three things transfer to businesses with no silicon. Compete on the layer where other people accumulate work, because that accumulation belongs to you economically without appearing on your balance sheet. Subsidise adoption early, when it is cheap, rather than buying it later at market price. And protect compatibility obsessively, because every break resets the accumulation and hands rivals an opening they could not otherwise create.

How this connects to Jensen Huang

This is the clearest instance of Huang's recurring move: identify a workload that would matter if the cost curve bent, then spend years bending it while defending the margin dilution to public shareholders. NVIDIA's later expansions — libraries, DGX systems, networking — extend the same layer logic upward.

Explore all Jensen Huang courses The CUDA Catalyst: Building an AI Operating System

Inside the membership

Why CUDA Won: A Bet on Programmability and Scale

Part of “The CUDA Catalyst: Building an AI Operating System” · 8 min

The complete lesson covering the internal decision, the sceptical years and the AlexNet inflection
The platform-strategy framework applied to non-hardware businesses
The framework saved to your Playbook
Localised audio summary and Master chat

Unlock the full lesson Members: open the lesson

Jensen Huang

Developer lock-in: ecosystem as a moat

Why an ecosystem of developers, libraries and trained engineers is harder to displace than any product advantage — and the conditions under which an ecosystem moat forms or fails.

Warren Buffett

Brand power: Coca-Cola's intangible asset

Why Buffett treats Coca-Cola's brand as a durable economic moat, how brand power converts into pricing power and distribution, and the test he uses to check whether a moat is real.

Steve Jobs

Killing products: the power of 'no' and focus

How Jobs cut roughly forty Apple products to four using a two-by-two grid in 1997, why saying no was a resource decision rather than a taste preference, and how to apply the method.

Independent educational analysis of the ideas, principles and documented work of Jensen Huang. Tiptico is not affiliated with, endorsed by or authored by Jensen Huang.