Yeah, there's something fishy in the Lisa numbers, either Apple lied and pulled a "Crazy Eddie" overshoot kind of thing and never got caught, or they reused the AppleNet numbers. So far we have a single collision, but if we find some more, it will be interesting.
Maybe we're all wrong and there really were ~120K of them made. Really difficult to tell with the data we have now.
So yes, those 7,000 SunRem got were on consignment - so SunRem had the warehouse space and Apple could be freed from stocking them, of those, some 2300 were ordered to go into the landfill for tax rebates. So many of the rest were sold off, and the remainder picked up by VintageMicros.
Some of the early ones likely were leased or given to universities as demos, not sure, but some of those wound up in the 7,000 SunRem got because the interview with Bob Cook indicates many were already used (and some were just random incomplete pieces.)
There's indications that some went to NASA because of Lisa Project, some went to the NSA (the WD2001 crypto chip made it useful there for the 2/10s), some went to an oil company that used custom software for the AM9512 FPU to figure out where to dig (as per info from David T. Craig for this 3rd one).
I don't know if the WD2001 and AM9512 are pin compatible or not, if they are (obviously you'd have different software to talk to them) would let us know whether or not the 2/5s and 2/10s would be interchangeable in this.
It doesn't seem that the SANE library has code for them AM9512 sadly, but I could be wrong, however both the WD2001 and AM9512 aren't directly accessible by the 68000, so it's not like the FPUs for Intel or 68010+, you'd need to go through the COPS VIA to talk to either of these, so probably nowhere near as fast as an 8087, etc. as the VIA's fastest access is 625KBPS which a single 8-cycle 5MHz clock memory cycle.
But you'd need several sets of these cycles for each byte and many more for a whole operation, and the lie is actually bigger since every other 8-cycle set is given to the video display, so you'd at most get maybe 1/4th or 1/8th or 1/16th the bandwidth of that 625KBPs.
I mean, you'd need a tight loop that reads a set of bytes, then writes to one of these chips, which means the CPU is using the bus to read from memory, turn around, and write to the VIA, and then have some DBRA opcode as well and ofc read the program opcodes as well. So at the end of day, you'll never reach 625KBPS despite the lie in the Lisa HWG.
Likely for crypto operations it would be much much faster since you'd need many DES cycles to do a full DES40 operation, but as an FPU, not all that great. For this you'd need to set a key in the WD2001's memory and then push some data to it, which it would encrypt and you'd get the result back in.
So it's not like telling the FPU, multiply this floating number by this floating number and get back the result. You might get some improvement over SANE's built in routines, but it can't be that much faster. Maybe you'd get 2x if you're lucky.