“self host to reduce power consumption” isn’t entirely accurate, even within a leased environment at Equinix or using a typical ‘powered shell’ provider like Sabey, the larger providers will still be customizing heavily many aspects downstream of the switchgear (and that’s where some but not all of these efficiencies can be located).
Tidbits from Peter Desantis’ keynotes and before that James Hamilton at re:Invent indicate that like Meta, Amazon doesn’t use 19” form-factor racks, has in-rack UPS solutions, DC power delivery and a sled-based server architecture. All these videos are on YouTube and the last 3-4 years (2018 onward) are really interesting and worth watching.
I’m short I’d guess Amazon and others partnerships with Equinix (and equivalents) are about two things: time to market, and some markets not being worth the investment [1] to drop 3x 30MW facilities into.
[1] https://www.datacenterdynamics.com/en/marketwatch/cutting-da...
A smaller provider usually sees a financial advantage to lease or build traditional datacenter space. “Smaller” can be quite large… I worked with an enterprise team that built out a large private data center circa 2015. Something on the order of low six figure server count. At that time the dollars worked out to build vs. cloud, mostly due to security control requirements. YMMV.