If I am reading this correctly, @Alex says his line is performing well, but the speedtest 'estimate' is not consistently giving him good results? Isn't this more weight to my point you can't rely on the speedtest.net results? If its using 1 + 3 random servers, you will need avg ~250 from each, Alex was not getting that, but getting good 'real world' results:
I will confirm however that multi-threaded speeds are absolutely fine especially real-world.
Another theory, I will assume the clever people that have designed the speedtest.net system and also those that have installed the speedtest servers, have built some filtering and restrictions in place to prevent users constantly running speedtests impacting users streaming video, music, making VoIP calls, ie speedtest is considered low priority network traffic, and network designed accordingly.
Maybe this is why speed.cloudflare.com or fast.com give higher 'estimates', as they have more control of the servers on their network, ISP will have a faster connection to those netflix/AWS/cloudflare networks than a random speedtest server at another ISP?
@bogoff has already confirmed that there is more than 1G connection from AWS to Zen speedtest server. I want to see the results from AWS to @bogoff, I happy to pay the 25p