Service providers degrade artificial intelligence models during server congestion under the belief that smaller models save money. Standard accounting assumes each query represents a finished transaction, but real customers pay for working answers rather than raw computational attempts. When a downgraded model outputs an inaccurate answer, the user immediately sends a retry query or leaves the platform entirely.
The physical breakdown begins when overloaded servers truncate context windows or switch to smaller neural networks to clear queue backlogs. This response resembles a store cashier rushing customers out with broken merchandise, forcing angry shoppers to line up again at the counter. Each returned request re-enters the arrival line, creating a wave of traffic that multiplies the incoming workload. If throttling passes an ignition threshold, the influx of retries manufactures more server load than the degraded model sheds, locking the system into a permanent degraded state.
Researchers modeled this operational behavior by combining inventory loss, retry multipliers, and queuing theory. Their analysis evaluated how cheaper models affect energy consumption compared to the capacity needed for each satisfied answer. The calculations revealed that running smaller models can save energy per completed answer while consuming strictly more server capacity.
The researchers demonstrated that providers can calculate the exact shadow price of intelligence for each customer class and hour in milliseconds. Server operators can use these closed-form trajectories to ration model capacity by critical ratio instead of firing reactive throttles during demand surges.