8-Nvidia

Nvidia was founded in 1993 to make graphics chips for video games. That was the whole idea, and for roughly the first fifteen years it was an accurate description of the company. Jensen Huang, Chris Malachowsky, and Curtis Priem built hardware that drew polygons quickly so that teenagers could shoot at each other in higher resolution. It was a good business in a brutally competitive market that killed most of its participants. The technical detail that turned into an industry was buried in the architecture. Drawing a 3D scene means performing the same simple arithmetic on millions of pixels simultaneously — which is a completely different problem from what a CPU is built for. A CPU is a handful of very fast, very general cores optimized to do complicated things one after another. A GPU is thousands of simple cores optimized to do one easy thing to enormous quantities of data at once. Nvidia built that architecture for shading pixels. But “do the same simple arithmetic to enormous quantities of data at once” also describes molecular dynamics, seismic imaging, financial modeling, and, critically, training neural networks. Academic researchers noticed first, and their methods were grotesque. In the early 2000s, scientists who wanted GPU speed had to disguise their math as a graphics problem: encode data as textures, express computation as shader programs, and read results back out as rendered images. It worked, barely, and it was miserable. Nvidia’s decisive move was to stop treating this as a curiosity and start treating it as a market. Drawing on Ian Buck’s academic work on general-purpose GPU programming, the company built CUDA, a way to write ordinary code for the graphics card, no textures or shaders required, and released it in 2006–2007. Then Huang did the thing the story usually skips. He spent the next several years pouring money into CUDA support, developer tools, and university outreach for a market that essentially did not exist, while investors asked why a gaming company was subsidizing scientific computing. It was a large, sustained, unpopular bet with no visible customer. The bet paid in 2012, when AlexNet won the ImageNet competition by an enormous margin, trained on two consumer Nvidia gaming cards. That result made the case that deep learning was primarily a compute problem, and Nvidia turned out to own the compute. Every subsequent wave — image recognition, then language models, then generative AI — ran on Nvidia silicon and CUDA, and the software ecosystem built up over a decade became a moat competitors still struggle to cross. Nvidia became one of the most valuable companies ever created, and the reason is a hardware architecture designed to make video games look better. The byproduct here is unusually pure. The parallel-compute capability was not a strategic insight. It was a physical consequence of the shape of the graphics problem — an accident of engineering that happened to match the shape of a problem nobody in 19