Compounding

Compounding is a positive-feedback loop, not merely repeated growth. Each cycle must preserve something—capital, code, data, trust, network density, or knowledge—that improves the economics or accuracy of the next cycle. The relevant multiplier is therefore the retained output after leakage: maintenance, errors, coordination costs, dilution, and decay. Better architectures can make each unit of compute more productive, while reusable workflows prevent research and operational learning from being discarded between runs architecture-is-eating-scale frontier-workflow-problem. Exponential-looking growth persists only while those gains are reinvested and the multiplier remains above one; otherwise activity expands without capability compounding.

The strongest loops join leverage with a scarce state variable. CME’s accumulated open interest and clearing infrastructure concentrate liquidity and reduce the incentive to move contracts elsewhere; greater participation then reinforces the venue that attracts the next participant cme-group-cme. Moody’s compounds institutional acceptance because its risk conventions are embedded in investment, regulatory, and audit workflows moodys-and-the-business-of-making-trust-machine-readable. Physical networks obey the same logic only when execution closes the loop: Old Dominion turns route density, service reliability, and selective pricing into better network economics, whereas Vestis shows that recurring contracts and route scale can coexist with weak profitability old-dominion-freight-line-and-the-ltl-network-that-compounds-by-refusing-bad-freight vestis-vsts. The newest evidence sharpens the model: compounding depends less on possessing nominal scale than on controlling the bottleneck that converts accumulated assets into reliable output—clearing in markets, verification and memory in agents, or service execution in route businesses cme-group-cme the-agent-bottleneck-is-moving-outside-the-model vestis-vsts.

A practical test is to name four things: the state retained after each cycle, how it improves the next cycle, what leaks or resets it, and which constraint becomes binding once the current one is removed. If more volume produces congestion, errors, or capital needs faster than learning and cash, the system is merely filling available capacity. If every completed cycle lowers cost, improves judgment, strengthens trust, or creates a reusable artifact, the system can compound; public work adds distribution and reputation to that loop 2026-05-07-building-in-public. Update the thesis whenever the bottleneck moves: as model capability becomes cheaper, evaluation quality, selective memory, and deployment realism become the constraints, so additional model scale may have lower returns than improving the surrounding control system the-agent-bottleneck-is-moving-outside-the-model.