Not really. RAM is only expensive because of the transistor size used. Fab plants are expensive. Packaging is expensive. Shipping is expensive. Silicon is expensive. If you add all that up, you end up with expensive products.
Because fab plants are running very large transistor sizes, you get low yields and high overheads.
Let's see what happens when you cut the transistor size by three orders of magnitude...
For the same size of packaging, you get three orders of magnitude more RAM. So, per megabyte, packaging drops in cost also by three orders of magnitude.
Now, that means your average block of RAM is now around 8 Tb, which is not a perfect fit but it's good enough. The same amount of silicon is used, so there's no extra cost there. The shipping cost doesn't change. As mentioned, the packaging doesn't change. So all your major costs don't change at all.
Yield? The yield for microprocessors is just fine and they're on about the scale discussed here. In fact, you get better. A processor has to work completely. A memory chip also has to work completely, but it's much smaller. If the three round it fail testing, it doesn't affect that one. So you end up with around a quarter of the rejection rate per unit area of silicon to a full microprocessor.
So you've got great yield, same overheads, but... yes... you can use the fab plant to produce ASICs and microprocessors when demand for memory is low, so you've not got idle plant. Ever.
The cost of this memory is therefore exactly the same as the cost of a stick of conventional RAM of 1/1000th the capacity.
Size - Exactly the same as the stick of RAM.
Power budget - of no consequence. When the machine is running, you're drawing from mains power. When the machine is not running, you are refreshing the dirty bits of memory only, nothing else. And 99.9% of the time, there won't be any because sensible OS' like Linux sync before a shutdown. The 0.1% of the time, the time when your server has been hit by a power cut, the hard drive is spun down to save UPS and the main box is in the lowest possible energy mode, that's when this sort of system matters. Even on low energy mode, buffers will need flushing, housekeeping will need to be done, transactions will need to be completed. This system would give you all that.
And the time when the machine is fully powered, fully up? Your hard drive spends most of its time still spun down. Not for power, although it'll chew through a fair bit - mechanical devices always do and the high-speed drives being proposed will chew through far, far more. They'll be spun down because a running hard drive suffers rapid deterioration. Can you believe hard drives only last 5 years??! Keep the damn thing switched off until last minute, then do continuous write. Minimizes read head movement (there's practically none), minimizes bearing wear-and-tear, eliminates read head misalignment (a lot of times, you can write the entire disk in one go, so what the hell do you care if the tracks are not perfectly in line with the ones they're replacing?) and (by minimizing read head time over the drive) minimizes the risk of a head crash.
I reckon this strategy should double the expected lifetime of drives, so take the cost of one 10 Tb drive and calculate how much power you'd need to consume extra for the memory in order for the memory's power budget to exceed the value of what you're doing.
Oh, and another thing. Because I'm talking memory sticks, you only need to buy one, subsequent drives of the same or lower capacity would not need to have memory there. You could simply migrate it. RAM seems to hold up ok on old computers, so you can probably say that the stick is good for the original drive and the replacement. That halves the cost of the memory per drive.
So, no, I don't see anything unduly optimistic. I think your view of what the companies could be doing is unduly pessimistic and more in line with what the chip companies tell you that you should think than what the chip companies can actually do.