[Official FAQ] Corrections and follow-up questions welcome.

Amortize hardware, power, maintenance time, and downtime across monthly tokens, then compare to API unit prices. Low volume often favors APIs; high volume plus data residency may favor local—or a hybrid.

Framework:

• Include idle cost when the GPU sits unused.
• Value privacy and offline capability explicitly—they are not free elsewhere.
• Revisit math when prices or utilization change.

Power rates vary widely; plug in your numbers. Anonymous bill comparisons welcome.