Say Copilot makes an engineer 2x as productive, their all-in salary is $500k to make the math easy, 1,000 MS sw engineers are using it, and Copilot took $5mm to train (GPT-3 took $4.6mm). Those 1k MS employees now being twice as productive are doing the work of an extra 1k people at $500k, or $500 million's worth in a year. That means $5 million in Copilot training costs are paid for in... 4 days. I have no inside information, so those number are all made up, but I'm pretty sure the initial training costs have already been paid off internally.
We also don't know how many multiples of $5mm it took to produce the initial version of Copilot, nor how many subsequent training runs there have been.
Point is, any significant productivity gains made, across an organization the size of Microsoft engineering, easily pays for big expensive training runs.