Why do competing ZK provers waste most of their compute racing each other?

In every decentralized ZK prover network today, multiple provers start generating the same proof simultaneously. Only one submission can land on-chain; the rest revert after spending 20 to 60 minutes of GPU time per collision. The effective utilization rate in a competitive pool is roughly one over the number of active provers, so adding more provers makes the waste worse. Commit-reveal schemes reduce collisions but add latency that erodes the benefit of decentralization; reputation-based pre-assignment reintroduces a trusted scheduler. No mechanism has solved task allocation without recreating the centralization it was meant to replace.