Yes. CPU usage of RAM lowers GPU available bandwidth more than what the CPU consumes, likely partly due to bank contention. There's a graph near the OP.
That being said, this rumor only makes sense if they're 8GB stacks. 10GB stacks (for a total of 20GB), don't exist AFAIK because you're looking for an even number of die given how the channels are constructed in the standard.
That being said, a 2048-bit HBM2 interface will actually take less die space than a 256-bit GDDR6 one. I'll also mention that if this is fake, the constructor of this rumor knows a lot about the industry to make it sound believable. Calling out TSMC's INFO_MS is clever shit.
OK thx. Interesting