In case anyone else wants to build an AI training farm to the right ratios, I did the math for generating AI cores at a steady-state rate. My goal was one machine training AI 9's into 10s constantly, with enough of all the other machines to keep it going constantly.

You need:
- 1 trainer training 9s
- 3 trainers training 8s
- 6 trainers training 7s
- 9 trainers training 6s
- 12 trainers training 5s
- 15 trainers training 4s
- 18 trainers training 3s
- 20 trainers training 2s
- 17 trainers training 1s
- 6 trainers training AI cores

For each ten second cycle, you need 3 AI cores, 1.01 of them goes unstable and gets recycled, and the other 1.99 turn into a mix of different levels (0.23 10s, 0.19 9s, and 0.20 8s are the important ones, though).

You can get a more efficient build by expanding - that was 107 machines to run one 9>10, but you can run ten of them with only 528 machines (10/23/38/53/67/78/85/88/68/18), because there's a lot that's idle at lower levels due to rounding errors. But tbh, the smaller farm already feels more than big enough, especially since it'll build up a surplus of each tier while you wait for the next tier to get researched.

You can always just do it as a sushi build, and tbh I think that might actually be better in principle, but I like the visibility and control you get from doing a segregated build like this.