Key Takeaways

  • Late-stage foundation model rounds require massive market capture assumptions that clash with near-zero switching costs across major LLMs.
  • Enterprise deployments remain trapped in proof-of-concept stages, with roughly 95 percent of corporate AI pilots failing to reach production.
  • The revenue base for foundation models remains heavily concentrated in software developers, a segment representing only a small slice of the world's 50 million software engineers.
  • Venture underwriting diverges sharply between growth funds modeling infinite agent TAMs and early-stage firms targeting proprietary data behind enterprise firewalls.
  • Growth investors underwrite multi-trillion-dollar upside scenarios using Swisher's Four S-Curves of AI Adoption.

Swisher's Four S-Curves of AI Adoption

David Weisburd outlines the bull case for foundation model valuations using an adoption model developed by Lucas Swisher, head of growth at Coatue. The model lays out four successive waves of demand:

  • First S-Curve: Developers: Early adopters driving initial LLM revenue through coding and development tools. Represents only a fraction of the world's 50 million developers out of 8 billion people.
  • Second S-Curve: Enterprises: Broader enterprise integration across business operations, currently constrained by high failure rates and early-stage infrastructure adoption.
  • Third S-Curve: Consumers: Mass-market consumer adoption reaching the wider global population of 8 billion people as tools become intuitive and omnipresent.
  • Fourth S-Curve: Autonomous Agents: An interconnected agent economy where autonomous AI agents communicate and transact with each other, expanding total addressable market far beyond the human population.

When This Works (and When It Doesn't)

Growth investors use this framework to justify multi-billion and multi-trillion dollar valuations for foundation models. It demonstrates why developer-driven API and seat revenues are merely the opening baseline for total addressable market sizing. If autonomous agents execute high-frequency transactions without human friction, traditional population caps on software spend disappear.

Where the framework fails is in its assumption that foundation model creators will capture that value. High TAM does not equal high margins. Studio VC's Joseph Coyne points out that the architectural reality of LLMs breaks the classic enterprise software lock-in thesis. “Switching costs are low,” (play from 34:03) Coyne observes. “I'm always chasing the best and fastest LLM for my use cases. And so what that results in potentially is that it's harder to monetize your huge user base because of competition.” (play from 34:04) When an engineering team can point its API routing layer from Anthropic to OpenAI or open-source weights with minimal friction, pricing power collapses. Furthermore, with 95 percent of enterprise pilots failing, the transition from developer experimentation to reliable corporate workflows faces severe delivery bottlenecks that delay the second S-curve.

Why It Matters

This debate highlights a widening split in private capital allocation between mega-cap model rounds and early-stage applications. Late-stage growth allocators must believe in the full realization of the agentic fourth S-curve to generate target returns on tens of billions in invested capital. Coyne takes the opposite side of the trade, arguing that commoditization risks and capital-rich tech giants turn late-stage model underwriting into a low-probability bet. Instead of backing capital-intensive frontier training runs with unstable customer retention, smart money is rotating toward companies that control proprietary datasets behind enterprise firewalls, where operational workflows build real defensibility.