I didn't make any assertion about AI, only about "AI" (note the quotes in my GP comment) — i.e. the same old machine-learning-based features like super-resolution upscaling, patch-match, etc, that people have been adding to image-editing software for more than a decade now, but which now get branded as "AI" because people recognize them by this highly-saturated marketing term.
Few artists want generative-AI diffusion models in their paint program; but most artists appreciate "classical" ML-based tools and effects — many of which they might not even think of as being ML-based. Because, until recently, "classical ML" tools and effects have been things run client-side on the system, and so necessarily small and lightweight, only being shipped if they'll work on the lowest-common-denominator GPU (esp. "amount of VRAM") that artists might be using.
The interesting thing is that, due to the genAI craze, GPU training and inference clusters have been highly commoditized / brought into reach for the developers of these "classical ML" models. You don't need to invest in your own hyperscale on-prem GPU cluster to train models bigger than fit on a gaming PC any more. And this has led to increased interest in, and development of, larger "classical ML" models, because now they're not so tightly-bounded by running client-side on lowest-common-denominator hardware. They can instead throw (time on) a cloud GPU cluster to train their model; and then expect the downstream consumer of that model (= a company like Canva) to solve the problem of running the resulting model not by pushing back for something size-optimized to be run locally on user machines, but rather by standing up an model-inference-API backend running it on the same kind of GPU IaaS infra that was used to train it.
