[btchip]

It's based on public ARM specifications (MPU + service calls) so the attack surface is way smaller compared to a virtual machine and the performance is of course way better. Applications can still corrupt their own memory space temporarily (not rewrite their code) but can't touch other applications or the Operating System.

[pjmlp]

So I don't see any benefit to go from possible logical errors to possible logical errors + memory corruption bugs.

[btchip]

the benefit is running native ARM code that you can verify vs running slow interpreted code in a black box. Have a look at the SHA-512 implementation in Java Card quoted in the article and it should be fairly obvious why you want to avoid that kind of pain.