This seems like the option that would make the most sense. If developers can "write once, run everywhere", they might as well do that instead of Cuda. But if they have to "write once, run on Intel, or AMD, or Nvidia", why would they bother with anything other than Nvidia considering their market share? If you're an underdog you go for open standards that makes it easy to switch to your products, but it seems like AMD have seen Nvidia's Cuda and jealously decided they wanted their own version, but 15 years too late.