[vonnik]

It works in all directions. Microsoft managed to commoditize upstack (Internet Explorer) and downstack (hardware) from its OS. The cloud vendors are attempting to commoditize the chips deeper in the stack (see Google's TPU and Amazon's new CPU thing) as well as AI tooling higher in the stack (Tensorflow). If you're a Vegas casino (and I mean, who hasn't been one at least once), you subsidize the flights to Vegas as well as the food on site. The flights get customers in the door and the food keeps them going. For the GCP, ML software gets them in the door. Cheaper processing keeps them going. Chipmakers like Nvidia and Intel are trying to do the same thing to ML software with projects like Rapids and BigDL.

Almost any layer can commoditize all around itself, as long as it has the critical mass of users. Google first got them with the search layer (who knew that was going to be such a thing in the late 90s?), which commoditized content. Netscape got it and then lost it.

Once you achieve dominance around a standard and necessary protocol at one layer of the stack, which is the whole game in tech, you can lay waste to competition in all directions, to ensure that no other layer of the stack commoditizes you.

[zozbot123]

How is Google's TPU chip a commodity? If anyone is commoditizing a ML chip, it's ironically-enough Nvidia <nvdla.org>. (And the RISC-V ecosystem will most likely get there eventually, but that's kindof a given anyway. Systolic arrays are very old tech, of Atari Transputer fame.)

[vonnik]

Google's TPU decreases demand for chips from Nvidia, by introducing a competing chip within Google that runs ML workloads really quickly, so people pay less for algorithm training. Since TPUs are not commercially available outside of GCP, GPUs are only commoditized for Google and whoever uses TPUs on its cloud.

[wbl]

That's not commoditization. It's the exact opposite where Google is trying to have a differentiated TPU product they use.

[vonnik]

It doesn't matter how TPUs are different from GPUs. It matters what the two chips have in common, which is the kind of workloads you can run on them. Any good that is a cheaper and plausible alternative to another good commoditizes the latter.