> If you are developing fundamentally new ML techniques (like the folks at Google)

Nope, still no superstars needed. The most technically challenging thing they've built is probably is what they are doing with the TPUs, but even there you "just" need people expierienced in the right fields. Most groundbraking research stuff these days is mostly on the mathematical side, not the implementation.