Guillermo Rauch on the Economics of the AI Supercycle and Agentic Infrastructure

Guillermo Rauch on the Economics of the AI Supercycle and Agentic Infrastructure

The Shift to Agentic Infrastructure

The rise of AI coding agents is creating the largest expansion in the total addressable market for software creation in history. While traditional software development was limited by the number of human programmers, AI agents can now write, repair, and secure software at a scale and velocity previously impossible. This shift is moving the cloud from a model of "Amazon Web Services" (AWS) toward what Guillermo Rauch describes as "Amazon Agent Services" (AAS).

From Pages to Agents

Historically, the internet was built on instantaneous request-response cycles focused on delivering pages and pixels. The "agentic cloud" introduces a different paradigm:

  • Long-thinking streams: Unlike traditional web pages, agents may take seconds, minutes, hours, or even days to complete a task (e.g., generating a comprehensive report or building software).
  • Compute Inversion: The focus is shifting from hosting static or dynamic pages to hosting autonomous agents that can maintain, advertise, and scale companies behind the scenes.
  • Token-based Economics: Intelligence is now streamed as tokens rather than pixels, leading to a transition from seat-based SaaS pricing to token-based pricing that measures intelligence consumption.

The "Block Economy" and Agentic Ergonomics

Software success in the AI era depends on "composability" and "agentic ergonomics"—creating building blocks that AI agents can easily utilize. Rauch refers to this as the "block economy," where agents do not reinvent the universe for every task but instead rely on existing, high-quality infrastructure blocks.

Why Agents Prefer Certain Tools

Coding agents exhibit a bias toward tools with "local reasoning" properties. For example, Rauch notes that technologies like Tailwind CSS, Next.js, and React are highly favored by agents because:

  • Local Reasoning: Components can be reasoned about in isolation, making them future-proof and easier to insert into different environments without breaking.
  • Context Window Efficiency: Because LLMs have limited context windows, code that allows for local reasoning is more efficient for the model to process.
  • Training Data Prevalence: The vast amount of open-source documentation and existing code for these frameworks provides a strong grounding for agents.

Core Components of Agentic Infrastructure

To make AI models useful, they require a full-stack environment that mimics the tools provided to human knowledge workers. Rauch identifies three critical layers of infrastructure required for the agentic era:

  1. The Sandbox (Compute): Just as a human employee is given a laptop, an agent needs a secure, ephemeral computer (a sandbox) to execute code. This increases the "IQ points" extracted from a model by allowing it to interact with a real environment.
  2. The AI Gateway (Token CDN): Tokens require the same scaling, acceleration, and security that pixels once did. An AI gateway acts as a "CDN for tokens," providing:
    • Observation and Failover: Monitoring token streams and switching providers if one fails.
    • Semantic Caching: Reducing costs by using smaller models for simple responses (e.g., responding "You're welcome" to a "Thanks") while reserving large models for complex tasks.
  3. Self-Driving Cloud (Automation): The future of cloud management is a "self-driving car" model where agents automatically configure, optimize, and ship performance improvements via Pull Requests (PRs) without human intervention.

The Impact on the SaaS Business Model

The velocity of AI software generation is making traditional, "lowest common denominator" SaaS products obsolete, while favoring systems of record with open interfaces.

The Rise of "Vibe Coding" and Tailored Software

Rauch observes a trend where users "vibe code" highly tailored tools to replace generic off-the-shelf software. He cites an example of a CEO replacing corporate parking lot management software with a custom tool live-coded in V0.

Value Accrual in the AI Stack

  • Short on: Static data/content businesses (e.g., Stack Overflow) and "training wheel" drag-and-drop builders that are too opinionated and constraining.
  • Long on: Companies moving at the "speed of tokens," offering consumption-based pricing, instant sign-up, and raw API access.
  • The System of Record: While the presentation layer of software is becoming plastic and malleable, the underlying system of record (databases, ACLs, and access control) remains essential. SaaS companies that expose these via agentic interfaces (like MCP or CLIs) will survive and thrive.

Operational Shifts for the AI Era

The scale of agent-driven demand is forcing a rethink of traditional engineering constraints, specifically regarding rate limiting.

Because agents can trigger deployments at a volume humans never could, Rauch has pushed his team to move away from rigid rate limits. Instead, the focus has shifted to abuse prevention and KYC (Know Your Customer) to prevent supercomputer-backed actors from incurring massive costs instantly, while allowing legitimate agentic demand to scale without artificial ceilings.

요약: Vercel 창업자 Guillermo Rauch는 페이지 중심 웹에서 에이전시 클라우드로의 전환을 논의합니다. AI 코딩 에이전트가 소프트웨어 생성의 대규모 확장을 주도하고, 이를 지원하기 위한 특화된 배포 인프라가 필요해지고 있습니다.

제목: Guillermo Rauch on the Economics of the AI Supercycle and Agentic Infrastructure

Sources