Anyone in a CPU-like market would do well to learn from the postmortem on Azul Systems' hardware. They had some neat technology for a hardware load barrier that made it possible to make garbage collection faster and had a Java implementation tuned to use it. And they were a complete market failure.

The reason was quite simple. In most cases, Java GC is under a 30% overhead. It's often much less, but even in the target markets where it might be as high as 50%, what does that mean? It means that a company that can reduce that overhead to zero needs to reach 50% of the performance of the baseline system, and 50% of the performance per Watt, and 50% of the performance per dollar, just to break even. For workloads that are 30% GC overhead, if they reduced it to zero, they needed to reach 70% of the baseline non-GC performance of an equivalent-cost Xeon. That's possible, but it's also expensive. The development costs are one or two billion dollars and then you still don't have the economies of scale that Intel has, so getting the per-unit costs down is even harder.

I've seen quite a few hardware startups since Azul left the hardware market who have completely failed to learn this lesson.