Direct compilation to x86-64 machine code does not get you performance on par with C (by which I assume the author means GCC or Clang). The optimization pipelines of GCC and Clang have had decades of work put into them by some of the best compiler engineers in the world.
Since the author states that the compilation time is linear, this would seem to imply that a full suite of optimizations are not being done, since many optimizations done by GCC and Clang have nonlinear complexity. It is easy to get fast compilation if you don't perform optimizations.
> - Thread safety and guaranteed absence of data races. You no longer have to constantly ask yourself: "Is this thread safe?" Everything is! No perfomance costs either. For example, if you are using a hash map in a concurrent function, a thread safe hash map is used automatically. Otherwise a faster single thread hash map is used.
This description doesn't guarantee freedom from data races. (Java's memory model basically fits this description, for instance, except for the specific case of hash tables, which aren't built into the language.) Even if it did, the tricky part is determining what a "concurrent function" is. The obvious ways one might imagine doing this tend to fall down in the face of higher-order functions.