Scientific (matlab but faster): who cares, you just want the answer, not the software, right?
Not always true. Sure, there is plenty of well-motivated ad-hoc coding in scientific research. However, we sometimes have supercomputers working for months to generate the answer. Even with well-written software this could mean many core-years of number crunching. Without good high-performing software we would not get the answer at all. Developing good scientific software takes time and effort too, but if the software can be used over and over to efficiently solve >1000 problems (for instance, the GROMACS papers have been cited by users ~15,000 times), then the time invested can be very good use of taxpayer money. C++ is not a bad choice for such software in that it enables very good performance and decent maintainability.