I’ve read the other replies but I remember well why it failed: only single precision floats.

You can’t run physics simulations on single precision. So the people who would have bought and tinkered and pushed it forwards were all turned off by it.

The military was allowed double precision; but there was no point making it a compute cluster otherwise.

I don’t know if it was a marketing or budget decision but basically all the early adopters who would have championed doing impressive things on there (I was so hyped at the time for the cell architecture) couldn’t do anything useful with it.