I worked on Gene - a gyrokinetics app used to study fusion reactors - and I fondly recall helping a colleague debug one of their simulations in which they set a

I worked on Gene - a gyrokinetics app used to study fusion reactors - and I fondly recall helping a colleague debug one of their simulations in which they set a boundary condition incorrectly; that simulation ran for two weeks before the mistake was found. We estimated this mistake to have costed the tax-payer multiple millions of dollars _in electricity costs_ (just from the core's that were used, their wattage, and the time duration; which is a fraction of the cost of acquiring and operating a super-computer). All our fusion models that correctly predict all the results produced by all the fusion reactors that have been built to date predict that ITER will work. I have no idea whether these models predict that these other reactor designs will also work. What I do know is that one needs _many_ simulations to study that, and the costs of doing that feels absurdly out-of-reach for any startup. One does not only need to "verify" a design, but come up with it, optimize it, etc. as well as gaining access to the supercomputing resources or buying and maintaining their own supercomputer. That puts the initial investment already in the millions. OTOH, a startup that checks a design that has been created, optimized, and verified in academia and whose goal is "only" to build it, would require a smaller investment, but no idea how big this investment ought to be.