Why can adding more threads make a CPU-bound service slower?

When runnable threads exceed useful CPU capacity, scheduling, context switching, cache disruption, and lock contention can outweigh any parallelism benefit. Shared data may also cause cache-coherence traffic or serialization in critical sections. Measure CPU utilization, run queues, contention, and throughput under a realistic workload. Size worker pools according to CPU, memory, and downstream limits rather than maximizing thread count. For I/O-bound work, more concurrency can help until another bottleneck is reached, but unbounded concurrency can still exhaust memory or overload dependencies.