Aatish Nayak, an investor who writes about AI markets, argued in a thread posted on X that the idea of a moat-less application layer misreads what abundant model intelligence actually does to an economy. A narrative has taken hold as frontier labs post record revenue and pack more capability into each new base model: that this surge of intelligence leaves nothing defensible for anyone building on top of it. Nayak’s central claim runs the other way: intelligence behaves like a utility, and utilities create durable value for whoever builds the pipes that route them into specific, unglamorous jobs.
His central image reworks the flood metaphor. A moat cannot hold back a flood, he writes, but engineered channels, dams, canals, and irrigation systems, turn that same flood into water that reaches crops and reservoirs instead of destroying everything downstream. Applied to AI, raw model capability is what commoditizes fast; the systems that convert that capability into a completed insurance claim, a cleared legal review, or a closed sales contract are not commoditizing at anywhere near the same rate.
The strongest part of Nayak’s post leans on history rather than forecasting. Adoption and real economic payoff have never moved on the same clock, he argues: factories wired for electric power in the 1880s did not translate into measurable output gains for another four decades, arriving only after plants redesigned how they operated around the new source of power. Today’s version of that lag is simply compressed. Usable AI systems are spreading across companies in a matter of months, he says, while the surrounding institutional work, rebuilt workflows, new lines of accountability, retrained teams, still runs on a years-long timeline. That gap between fast adoption and slow institutional catch-up is, in his telling, the arbitrage window application-layer companies have before it shuts.
Seven tactics fill out how Nayak thinks that window gets exploited: building coordination layers across humans and agents, accumulating proprietary workflow data labs will never see in pretraining, letting customers configure their own automation rather than imposing it, establishing a credible narrative about an industry’s future, moving up the abstraction stack as lower layers get automated, pricing against business outcomes instead of seats, and making a company regulators or institutions cannot route around. Each tactic gets a named example. Harvey stands in for legal coordination, Applied Compute for customer-controlled model deployment, Factory for the move from individual coding agents to full software-production pipelines.
A real gap sits inside the argument that Nayak does not close. Every company he names is a well-funded, venture-backed startup already selected for survival, which makes it easy to build a framework around winners after the fact rather than before. Missing is any counterexample: a firm that tried one of these seven moves and still lost, which would tell us whether this is a genuine strategic filter or just a description of what already-successful application-layer firms happen to do. The piece also sidesteps the hardest case, whether a frontier lab’s own product suite, bundled straight into a model API, eventually absorbs the coordination and workflow-memory layers he calls defensible.
The most testable claim in the thread concerns timing, not strategy. Nayak expects that within roughly a year, AI spend turns into a line item every C-suite must defend against revenue or operating costs, which would push companies toward pricing tied to outcomes such as resolved tickets or closed claims instead of headcount. Should that shift land on schedule, it would validate his outcome-based pricing argument more convincingly than any single company he cites. Operators building on top of frontier models have roughly the next year to treat as the window Nayak describes: the stretch to lock in proprietary workflow data and outcome-based contracts before a lab ships a bundled version of the same layer.
Aatish Nayak (on X), August 25, 2026.
