BM1387 ASIC inside the Antminer S9: 189 pcs per unit, each unit 14 TH/s
So, each BM1387 is capable of ca 0.07 TH/s, consumption roughly 5 watts as well
To achieve the current 5.000.000 TH/s, you would need 71 millions of those.
Assume 50% mobile users with one chip, and 50% desktop users with two chips, that means a completely fully decentralized network of roughly 50 millions users.
The issues would seem to be more on the network bandwidth / latency etc side than on the ASIC side I believe.