Besides horrible management---echoing Zigurd, a friend and contemporary, I was at one point reliably told that manufacturing, R&D and marketing were paying no attention to each other, such that manufacturing had built a factory that couldn't make the latest hardware R&D had developed, which was done completely independently of what marketing thought was needed---they were killed by non-recurring engineering (NRE) costs.
Basically they couldn't amortize the NRE for their custom hardware and later most especially chips (from memory, first, a chipset that spread the CPU across several chips, rather like the one Western Digital did that among other things enabled the LSI-11, then of course an all in one chip) across the huge number of units that Intel and Motorola sold. They also canceled their RISC rethink of their basic low level architecture on the day it was supposed to tape out; don't know if it had a low enough gate count to be like the first SPARC processor, which was implemented on 2 20,000 gate arrays, one for the CPU and one for the floating point unit (gate arrays are all alike until a few layers of metal are put on top).
So soon enough you could run a full Common Lisp, almost certainly without as much run-time error checking, faster on cheap commodity hardware than on a Lisp Machine.
Something like that seems to have happened to Azul Systems, which apparently isn't developing any new hardware, but is selling their version of the HotSpot JVM to run on x86_64 hardware. A prior generation of their pauseless GC (vs. 1 second per GiB, a big deal if your heap is 100s of GiB) required a software write or read barrier that cost ~20% of performance (all this from memory). It's likely that soon enough, even if it perhaps ran slower than their custom hardware, it was a lot cheaper to run it on commodity Intel/AMD hardware.