AI's New Value Frontier: Why Loops, Not Models, Build Moats
By Klaus Berger ·
The next wave of value capture in AI will shift from raw model intelligence to sophisticated application layers and automated business loops.
The recent episode of "a16z" tackling the state of AI—Models, Moats, and the Consumer Renaissance—is less a discussion of technological capability than it is an analysis of economic primitives. What struck me most was the clear delineation between intelligence as a primitive (like cloud computing before it) and how value capture will occur at the application layer through sophisticated "loops."
On the podcast, Anish noted the fierce competition among model developers, citing the rapid rise of XAI and OpenAI's performance. The co-host elaborated on this structure, arguing that while models are not commodities—there is specialization occurring (e.g., neurotic vs. open models)—the true value lies in aggregation. This perspective suggests that an application layer can achieve a "greater than sum of parts outcome" by running the same query through multiple specialized models. Furthermore, the discussion highlighted the evolution from simple prompting to complex "business loops," where agents automate entire processes, such as bug fixing or price optimization, rather than just generating text.
The Persistence of Structural Moats
The underlying economic argument presented is compelling: while AI provides an extraordinary primitive, it does not nullify traditional business moats. The consensus was that network effects, scale effects, and brand recognition remain critical. However, the co-host introduced a crucial vulnerability—the "Integration Moat"—arguing that coding agents pose an existential risk to older, complex systems (like SAP) that relied on sheer complexity for value extraction.
The strongest case against this structural shift is the belief that raw intelligence will inevitably commoditize all specialized knowledge, rendering proprietary data or unique integration difficult to maintain. To counter this, one must recognize that while AI lowers the marginal cost of distribution dramatically, it does not eliminate the high onboarding costs and complexity required for deep enterprise adoption. Moreover, the market's appetite is shifting away from purely technological spectacle toward "luxury software"—high-end, subscription services built around compounding value through memory (like Town), which promises sustained customer retention and pricing power.
From Tech Play to Economic Infrastructure
The most profound takeaway was the shift in capital allocation focus: investors are increasingly prioritizing product velocity and requiring live product demonstrations over mere pitch decks. This signals a maturation of venture funding, moving beyond pure research potential toward demonstrable market fit.
This is not unprecedented; every major technological wave—from the telephone to the internet—has required foundational platforms (the "Windows" moment) before mass consumer adoption occurs. The argument that AI is currently in a "DOS era" for consumers holds up because of infrastructural hurdles: there is no singular, established "AI native distribution channel," and consumers remain reluctant to pay for software despite the low marginal cost of delivery.
However, where this analysis excels is its structural view of capital allocation. Instead of viewing AI as simply another vertical market (e.g., a legal AI or an accounting AI), it must be viewed as a new industry that fundamentally changes how value can be packaged and consumed. The opportunity lies in the application layer—the ability to take raw primitives and customize them for unique operational goals, such as Harvey for law or tailored productization for credit unions.
The current moment is thus defined by an "embarrassment of riches," where capital allows founders to pursue a broader array of value propositions than ever before. This shift favors the young founder—the researcher who assumes everything is possible—over the senior executive whose ideas may be limited by past technological ceilings. The ultimate bet, therefore, is not on which model lab will win the race for compute dominance, but rather on which application layer can successfully build an original network effect through word-of-mouth and deliver highly specialized economic outcomes to specific, underserved customer segments.