CUDA: How NVIDIA Forged an Unbreakable Empire in the Age of AI

Lesson 1 in tech is Jensen never loses. Lesson 2 is that if you study Lesson 1 you can figure out why and how to apply those lessons in your corner of tech. That's what this piece is for!

We hear a lot about moats, and I thought it was time to write about the greatest moat of all time: CUDA. Not many people properly understand CUDA and why it matters, and NVIDIA doesn’t talk about it strategically in public (if you have a moat you don’t talk about it like a moat). So I took some time to break it down, describe how CUDA gives NVIDIA an edge, and then draw out lessons learned for builders in an age when moats feel increasingly out of reach.

Let’s rewind 25 years…

In the early 2000s, a silent revolution simmered in the shadow of a very public CPU war. Intel and AMD were taking turns bragging about clock speeds and multicore architectures, each vying to dominate the heart of personal computing. GPUs, meanwhile, were seen mainly as gaming hardware—great for rendering pixel-perfect explosions but not much else. Behind closed doors, however, researchers at universities like Stanford and MIT were experimenting with GPUs for scientific computing, running molecular dynamics simulations and image processing tasks. The results were promising but messy—programming GPUs for non-graphics tasks required low-level hacks, arcane knowledge, and lots of patience.

NVIDIA took note of these hacks, saw the enormous potential, and made a bold decision. In 2006, they introduced a software platform that would transform their graphics cards into parallel computing juggernauts. That platform was CUDA, and it would not only reshape the GPU industry but also pave the way for the AI boom. While AMD and Intel battled in broad daylight, NVIDIA quietly erected a fortress—one whose walls were made not of transistors alone, but of a comprehensive software ecosystem that would soon dominate deep learning, high-performance computing (HPC), and data-center workloads.