[baq]

This was true for Intel for at least 10 years and I’m pretty sure for much longer than that. It was probably true for nvidia for about as long as they exist.

Hardware without software is just expensive sand. Every semiconductor company knows this. Intel was the one to perfect the whole package with x86 in the first place…

In the GPU compute space CUDA is x86. It’s ubiquitous, de facto standard and will be disrupted. Question is if it takes a year or a decade.

[JonChesterfield]

The stereotype is hardware engineers all think software is easy. So while semiconductor firms know software is important, they're often optimistic about the ease of creating it.

Cuda is enormous, very complicated and fits together relatively well. All the semiconductor startups have a business plan about being transparent drop in replacements for cuda systems, built by some tens of software engineers in a year or two. That only really makes sense if you've totally misjudged the problem difficulty.