It's way too premature to say we are turning into fragmentation cycle regarding bespoke hardware. GPUs are still universal computers, they just use a different

It's way too premature to say we are turning into fragmentation cycle regarding bespoke hardware. GPUs are still universal computers, they just use a different model with different performance characteristics (and applications, which stem from these performance characteristics and also inertia/maturity of tooling) than CPUs. ASICs and specialized processors have been a thing for a long time and their application towards some "new" things like crypto, hft, and a handful of cloud optimizations is hardly a new trend when seen in the larger context of their being used all the time for now-boring, once also-new technologies like networking hardware and embedded applications. I'd argue that the last three decades have actually had massive shifts towards more generalized and ubiquitous computing that may continue even further. First, with the introduction of PCs the computing landscape shifted away from bespoke hardware and software to more standardized plug-and-play hardware and software. With the internet, everybody got a middleware called a "web browser" that standardized UI development on a single computing model. With mobile computing, we got a more restricted model of userspace that standardized what applications could do and even how (app review, software signing) you could do it. Cloud did the same things to various extents to server software. Except for at the height of Windows' PC dominance, it's never been easier to write software usable by such a large portion of the market, and the absolute number of devices and users that applies to is probably an order of magnitude more than existed during peak-Windows. Everybody in tech knows single-core cpu performance gains are slowing down and that the remaining incremental improvements will eventually end too, so what comes next is on peoples' minds. IMO this article is jumping the gun though - it's still way too big an undertaking to use specialized computers (and I don't count GPUs as these) for all but the largest scale or most performance sensitive use cases, as it's always been.