A lab in Addis Ababa trains a mixture-of-experts layer with 256 experts on eight machines, and every word picks eight of them. Left alone, the router piles words onto favourites and the whole step waits for the busiest machine. DeepSeek-V3 nudges a bias on each expert instead of adding a balancing loss. Then switch the balancing off, make the nudges too big, or let words reach any machine. This lesson needs JavaScript.