Allwinner A20 (4 core ARMv8 Cortex A53, 512KB L2 cache, 128KB L1 cache, 2 core Mali 400, video interface, etc.) costs 6$ in volume. So I guess just in plain die area, the 64 core Mars ARMv8 (32MB L2 cache) could take just 8 times the area of Allwinner A20, without including L3 cache. That would mean 48$ just in silicon, with naive extrapolation. Also, there are two versions of the chip: the high-end ("Mars") and a low-end ("Earth"), may be without L3 cache ([1]).
With progressive die reduction, it could be even cheaper, if demand reaches enough critical mass (e.g. billions of chips for routing devices, servers, etc.).
In addition to Cavium, there is another potential competitor: Mediatek (after Atheros aquisition), selling crazy cheap SoCs for SOHO routers. What if they start in the high-end and server market? (they would need just IP available from ARM for the inter-core memory routing, for making cheap SoCs).
[1] https://tweakimg.net/files/upload/HC27.24.320-Mars-64core-Ga...