I am surprised that they say that full custom poses a higher 'risk' - that's only true for very complex chips like cpu's, not for the very simple (and I mean very simple) functions found in SHA256.
The added 'risk' is because for the first time they are getting outside the 'standard cell' design flow.
I have doubt that the new design is truly full custom. Their previous designs were simple unrolled hashers. True 'full custom' optimized design would be rolled. And switching from unrolled to rolled would involve redesign of the I/O protocol.
My bet is on them purchasing a custom standard macro library: lower-power by lower-area and lower-noise-margins. Sort-of like bitfury did for his first chip: 55nm-drawn transistors in the 65nm-nominal process.
Such a 'extra-low-power' library may be violating some default DRC's (design rule checks) of their foundry. Thus the foundry makes them explicitly waive DRC conformance warranty with their mask order.