Right now? For broad applicability it is the computational overhead. There are some applications that are doable now, such as reading data anonymously from a sm

Right now? For broad applicability it is the computational overhead. There are some applications that are doable now, such as reading data anonymously from a small-ish dataset (or a low-volume service where you can afford to wait a minute for an answer, or using really expensive hosting). An example of that was a Wikipedia server that someone made which would serve you pages without the server knowing which page your client was actually after (https://news.ycombinator.com/item?id=31668814 https://news.ycombinator.com/item?id=31668814 4 months ago, 119 comments). It's still not really efficient; you can't simply swap out the real Wikipedia for this system and expect it to simply work. > the server [needs] to scan through the entire encrypted dataset [for every request] (this is unavoidable, otherwise its I/O patterns would leak information) Imagine Wikipedia servers needs to read every byte written on Wikipedia and operate on it before being able to formulate an answer to a random pageload. Additionally, if I remember correctly, things like autocomplete worked by just downloading the entire list of articles and doing that locally. It's all not impossible, but not a drop-in solution. And then when you have a situation where you can practically apply it, there aren't popular/trusted/already-audited software packages out there for you to just use with confidence.