One agent request crosses eight separately metered services, and each of them prices a different unit of that request. Doubling traffic doubles every one of those meters at once, along with the egress charged for moving data between them.
Akka runs agents, orchestration, memory, streaming, endpoints, guardrails, evaluations, training, and inference on shared compute. One request is one unit of consumption, and the invoice is billed on platform tier and service cores.
An agent request does not stop at the model. Orchestration, memory reads and writes, streaming, the gateway, telemetry and the safety check all happen on the way. On a stack assembled from separate cloud services, each of those steps belongs to a service with its own price list.
Each service in that stack is sized on its own, and each carries headroom so it does not become the constraint. Headroom is provisioned eight times over, and the amount held grows with the deployed base.
Shared compute removes the separate meters. Memory, streaming, observability and governance are properties of the Akka runtime, so none of them carries a line of its own. No data moves between services that would charge for the move. The infrastructure cost reduction reaches 90%. Fox shrank its AI personalization engine from 150,000 cores to 22,000 after porting to Akka.
An Enterprise Agreement and a volume tier both lower the rate charged by each service. The number of services a request crosses stays the same, so the bill still grows on eight meters at once.
Growth follows the shape of the agent. A multi-step agent adds orchestration steps and memory queries faster than it adds tokens. A long-conversation agent adds tokens faster than it adds steps, because the whole history is re-sent on every call.
Open-weight models change the token line. Orchestration, memory, streaming, the gateway, observability, safety and egress are billed by the services that provide them, and none of those charges depends on which model was called.
The reduction comes from shared compute, sub-10ms memory native to the runtime, scale-to-zero, the removal of inter-service egress, and evaluation, guardrails, observability and governance built into the platform.
