Editorial read
In depth · editorial + model
A cluster scheduler is the orchestration software that decides which training or inference jobs run on which chips, and when, so that expensive accelerators never sit idle. It matters because installed hardware is not the same as usable hardware: the scheduler is what converts a warehouse of GPUs into economical, high-utilisation throughput. In a world where compute is scarce, this software layer behaves as invisible capacity - better scheduling wrings more useful work out of exactly the same silicon.
If scheduling is poor or fragile, the penalty is idleness and failure - jobs that stall, chips that wait, and scarce compute wasted at the moment it is most valuable. Because this is a software moat rather than a hardware one, the exposure sits with the cloud operators that build and run orchestration at scale, such as Oracle and Nebius, whose unit economics hinge directly on keeping utilisation high.