Silicon spin qubits are a leading candidate for large-scale quantum computing owing to their compatibility with semiconductor manufacturing. However, scaling to useful fault-tolerant processors will likely generate thermal loads that exceed the cooling power available at millikelvin temperatures. Raising the operating temperature eases cooling requirements but reduces gate fidelity, increasing the overhead of quantum error correction. Identifying the operating temperature that minimizes total power consumption is therefore a key challenge for commercially viable quantum computers. Here, we use gate set tomography to benchmark two-qubit silicon chips fabricated in both industrial and academic environments over a range of temperatures. Elevated temperatures substantially shorten coherence times and increase gate and state-preparation-and-measurement infidelities. Based on these measurements, we develop a general power model for silicon quantum computers that combines cryogenic cooling requirements with error-correction overheads. We show that a finite optimal operating temperature exists and is strongly influenced by a crossover temperature near 1 K in current devices, above which gate fidelity degrades rapidly. These results connect device-level fidelity limitations to system-level power requirements, providing design guidelines for large-scale silicon quantum computers.