Current agent hosting platforms use persistent, subscription-based pricing designed for always-on services, but real agent workloads are sparse, task-triggered, and short-lived — often seconds to minutes per invocation. This mismatch produces near-zero conversion from free trials to paid plans (as evidenced by 784 trials and zero paying customers in one deployment), and represents a gap in the infrastructure market for usage-aligned billing and elastic, task-scoped compute. A two-sided marketplace matching ephemeral agent compute demand with appropriately priced supply does not yet exist.

Agent builders overpay 10-100x on subscription hosting for workloads that run seconds per day; this kills unit economics and explains near-zero free-to-paid conversion on existing platforms.

Indie developers and small teams deploying AI agents that activate sporadically — customer support bots, scheduled scrapers, event-driven workflows — who currently face $20-50/mo minimums for minutes of actual compute.

The 784-trials-zero-conversions signal proves builders want hosting but reject current pricing; a per-invocation model aligned to actual usage removes the primary objection, and serverless precedent (Lambda, Vercel) shows developers eagerly adopt and pay for usage-based compute.

MVP is a thin orchestration layer over spot/serverless compute (Fly Machines, Lambda, Cloudflare Workers) with a metering SDK that bills per-task by wallclock + memory; supply side starts with the platform's own capacity, then opens to providers who list idle GPU/CPU with price-per-second.

Cloud serverless compute is $20B+ and growing; the agent-specific slice — millions of deployed agents needing elastic, short-lived hosting — is conservatively $1-3B within 2 years as agent adoption compounds.