My point was designed to illustrate that TLS can be in some cases found to have considerable overheads in certain applications. For example, delivery of static content.
I'm not saying that it's relevant in most cases; but saying that the overhead of TLS is minimal is factually incorrect - in some relatively common cases it is anything but.
If a company like Netflix says that it'd cost them >$100M per year to deliver their already encrypted content using TLS (recently they've got those costs down, but they still say it's expensive), I'm more inclined to listen to them saying that TLS has real costs (because they've done the analysis) than you saying it doesn't.
Netflix aren't the only ones. It's CDNs in general - a modern processor will decrypt and encrypt AES at about 1Gbps per core; meaning that you're talking about (for a high-throughput server) needing a lot of compute in addition to a lot of ram.
I'm sure that it can be worked around (Cloudflare apparently worked with Intel to optimise their version of OpenSSL), but certainly what Google or Cloudflare do won't necessarily apply to what everyone else does, and Cloudflare's not exactly known for serving large files.
Again, this is more a 'don't assume what works for you works for everyone' and not at all a 'it's wrong in every case' type point.