I believe this practice actually started out with the enterprise stuff, i.e. the "large corporate buyers", and if anything, they are more likely to not care or embrace it.

Hyperscalers have already been bought in for years. There's custom Amazon, Google, etc. silicon in Xeons that either gets fused off in the general case or needs a secret MSR knock to turn on, or what have you.

Is there any public documentation as to what these features would be, and has anyone discovered ways to turn them on?

I remember hearing a few years back about Intel rolling a special weird Xeon SKU with VNNI/bfloat16 tacked onto a different core architecture that got cancelled, specifically because Facebook backed out of the project. However, I think that was intended to be a publicly-available SKU that just didn't make monetary sense to produce in volume.

There's also the occasional off-roadmap SKUs that eventually wind up on eBay or wherever... but my impression was that these were just ways for Intel to obfuscate whatever special volume deals they were giving certain hyperscalars.

Intel Cooper Lake - it's a 14 nm server part that fit a particular niche (4S support, VNNI/bfloat16). It was partly due to FB, but also partly to provide a platform that was larger than ICX server could / can.

enterprise customers have been the leading group demanding software-defined "capacity on demand" features for literally decades