@nayakkayak - Moats in the Age
@nayakkayak on X The author argues that as frontier labs like OpenAI and Anthropic absorb more capability, app-layer startups should stop chasing moats and instead build infrastructure that spreads intelligence through real workflows. The post lays out seven tactics for doing that, with examples from companies such as Harvey, Klarity, Applied Compute, and Factory.
- Intelligence is treated as a utility with unbounded demand, so the winners will be the companies that disperse it into workflows and convert tokens into outcomes, not the ones that just sell raw models or agents.
- Adoption lags capability because real work carries hidden state, exceptions, and human coordination costs; the author sees this gap as a short-lived arbitrage, comparing it to electricity taking decades to show up in productivity statistics.
- Durable advantage comes from owning coordination across people, agents, and organizations, accumulating proprietary process data, letting customers control their own AI deployments, and pricing against business outcomes rather than seats.
The core argument is that a flood of cheap intelligence from the frontier labs won't wipe out application companies, but it will make the ones that build distribution infrastructure extremely valuable. The author's analogy is that moats are useless against a flood, while dams, canals and pipes that route water to farms and homes are critical infrastructure. Intelligence, he argues, is a utility with unbounded demand, and the premium goes to companies that convert raw tokens into outcomes across industries. He points to a real gap: GPT-5.5-class and Opus 4.5-class models would have looked like AGI in 2019, yet the economy hasn't visibly transformed. The reasons he gives are context (real work carries more state, exceptions and history than any prompt holds), the stubbornly human nature of coordination, incentives and approvals, and the historical lag between adoption and productivity, using electrification as the example: wired into factories by the 1880s, but not showing up in productivity statistics until the 1920s. He calls this gap the biggest arbitrage available right now, and says it's open only for a limited window.
To exploit that window, he lays out a roadmap of eight tactics and says you need to do most of them over time, not just one. The first two are the ones he spends the most time on. Orchestrate a multiplayer network: target workflows where human coordination costs are highest and build products where people and agents collaborate end-to-end, so you accumulate a graph of individuals, agents, data and organizations that is hard to unseat. Harvey's work connecting Fortune 500 clients with their law firms is his main example. Accumulate workflow gravity: capture the proprietary data that general models never see in pretraining, including internal documents, corrections, decisions and exceptions, and use it to build memory and personalization. He argues a continual-learning layer may split off from the models themselves, since enterprises want to hedge against single providers. Within, which surfaces latent work to suggest automations, is the example here.
The remaining tactics are shorter but follow the same logic. Let customers own their transformation means treating tokens as a budgeted resource like headcount, with configurable agents, model choice, cost visibility and permissions, while avoiding giving so much control that customers simply vibe-code their own version (Applied Compute is the example). Tell your version of the future means offering a specific, credible picture of what an industry looks like in five years, and pairing it with a brand that takes a stance, since model access alone won't differentiate anyone. The last four, which include keep climbing the abstraction, sell what wasn't possible before, and make yourself a structural necessity, follow the same pattern of moving up from features toward outcomes that customers can't easily do without.