As I understand, they don’t use it to generate code, but for scanning code for bugs. This is relevant as there are now many CVEs found by AI scanning tools.

Fun fact: Based on estimates of the value of the Linux kernel (around 2200 Million Euros in 2011), the size of the kernel (in 2011, version 2.6 was current, with a bit less than 20 million lines of code), abd average salary of kernel developers (for Germany, one could estimate 60,000 - 80,000 EUR/year), one can estimate the lines of new code that a full-time kernel developer produces - less than 1000 lines per year:

 value=2200000000
 cost_man_year = 60000
 lines_of_code= 20000000

 person_years = value / cost_man_year
lines_per_person_per_year = lines_of_code / person_years

I get to 545 lines per person-year that way.

Even if AI could reduce time to type in that code to zero, it would not speed up kernel development, since developers are 99% of the time doing something different than typing in code - determining what new code should do, organizing and structuring concepts, or fixing bugs (which is determining what existing code should do, determining what it actually does, and determining how to fix differences between the former and the latter).