I think they were saying those were the reality-based markets driving the "banal and boring" valuation before the hype bubbles.
I'd add also that on the technical level the new NPU stuff really is very banal and boring, comparing to the old "GPU" stuff.
The pre-AI compute was very complex, having to strike a good balance between competing needs of having flow control and doing actual computations. On top of it, the actual GPUs had a lot of extra specialization for drawing triangles on top of triangles.
The AI specific compute, it's mostly just accelerated multiply-adds on large arrays of data. No branching, no actual low level code as such. Far more multiply-adds than evaluations of the transfer function, too.
Even if you don't want to use a machine learning library like Tensorflow or Torch because you're doing some unusual math, you can use a generic math library like e.g. Jax.
You don't have to write any code that runs on GPU, at all, unless you're doing something very unusual like e.g. implementing a differentiable raytracer because you're reconstructing 3D scenes and synthesizing views / relighting them / etc (something with far more relevance to Hollywood than any of the overhyped generative garbage). And even if you're doing the latter with modern hardware you may be better off finding a way to express it in terms of large array operations, without any branchy on-GPU code.
One thing that is rather infuriating about generative AI hype is that it is even sucking the air out of the room for fundamentally interesting AI work.
edit: One big idiotic promise is to replace all actual AI work with just generating some garbage using an off the shelf LLM architecture, going as far as an example in this very thread (generating chemical formulas as sequences of characters with an LLM, as if it was ever likely to work even OK considering how the underlying problem is in 3D and not in fact a linear sequence, and how a million training samples is actually a lot less than a trillion).
