[CamperBob2]

I'm tempted to call that pure luck. As far as they knew, crypto would be the killer app.

However, if you start with the assumption that at some point, people are going to need a lot of fast parallel compute for something, you could rationally justify their long-term strategy. They skated where the proverbial puck was going. They couldn't see the puck, but they were pretty sure there was one. In hindsight that really does look like a safe bet.

[dist-epoch]

People were (ab)using OpenGL to run compute on GPUs in 2004-2006, doing stuff like rendering 2 triangles covering the whole screen and then doing the actual compute in the pixel shaders, getting 10x speedups over CPUs for some problems.

NVIDIA just had their eyes open to an obvious market demand and made it easier by creating CUDA.

[KK7NIL]

Nvidia subsidized machine learning research for years (both with CUDA, hardware donations and developing what was a very niche product line just for them) before deep learning became big, much less the advent of LLMs.

Certainly Jensen seemed to have an extremely long view on this burgeoning machine learning market in the early 2010's.

[Danox]

It didn’t hurt that they had a two companies named Intel and Microsoft that completely missed the boat where GPUs or mobile computing were concerned both are currently the top two companies in tech today by market cap?

[fhn]

CUDA came out of the need for running parallel cores in their GPUs. This is not luck, it's product evolution. They did it first, they did it best, and they are reaping the benefits. The alternative here is to Not have CUDA and continue writing sub-optimal code for GPUs.