Software, the perennial disruptor, is now being disrupted. That’s the core seismic shift explored in a conversation with a16z’s infrastructure experts, who frame the current AI revolution as the emergence of a “fourth pillar” of infrastructure, joining compute, storage, and networking. This new pillar of intelligence is fundamentally changing how software is built and who can build it, turning developers into kids in a candy store with powerful new tools while simultaneously challenging the very nature of programming. The discussion traces the evolution of infrastructure investing from the cloud transition to the present AI boom, examining how this wave differs from past cycles and what it means for the future of technology creation.

The core of the conversation revolves around defining infrastructure not by product category, but by its technical buyer—developers, data scientists, and engineers who use tools to build other things. In this light, AI models and their surrounding tooling become a foundational layer. This shift is profound because, for the first time, software is abdicating core logic to AI models, forcing a rethinking of programming itself. The panel draws parallels to past disruptions like the internet, noting that such expansive phases create massive new markets and behaviors, offering white space for new companies rather than just replacing old ones.

A significant thread is the evolving role of the developer in this new stack. Far from being replaced, developers are becoming more powerful and numerous, empowered by AI coding assistants and agents that act as “thought partners.” This is leading to a “horseshoe theory” where marketing to developers now resembles consumer marketing, and the line between infrastructure and application is blurring in the early stages of the cycle. The discussion concludes by pragmatically assessing the current state of AI, arguing that while “agents” show immense promise in constrained environments like coding, the real near-term work lies in “context engineering”—the complex systems work of providing models with the right data and tools to function reliably.

Surprising Insights

• The median pull request changes just two lines of code. This statistic underscores that the primary challenge in software isn’t the brute-force act of coding, but the nuanced work of understanding business needs and user requirements.
• Infrastructure layers don’t get commoditized away; they consolidate into oligopolies that retain value. Contrary to fears of zero-sum competition, history shows that each infrastructure layer (cloud, chips, databases) tends to consolidate into a few dominant players that maintain healthy margins.
• The most common uses for generative AI today are “top of the pyramid” human needs like companionship, therapy, and schedule management, not the enterprise workflow automations many initially predicted.
• Low-code’s ultimate form is natural language. The long-promised vision of low-code/no-code platforms is finally being realized, not through simplified GUIs, but through the ability to prototype and build using plain English instructions to AI.
• During market expansion phases, zero-sum thinking is “deadly.” In a rapid-growth wave like AI, there’s enough new value and territory being created that companies are not yet fighting over slices of a fixed pie, making aggressive investment and building the correct stance.

Practical Takeaways

• For builders: Focus on “context engineering” over simple prompt engineering. The next frontier is building the reliable data pipelines and tool-integration systems that provide AI models with the correct, up-to-date context to perform useful work.
• For investors and strategists: In an expansionary super-cycle like AI, avoid zero-sum mental models. Look for where new behaviors are being created and new TAM is being expanded, rather than focusing narrowly on direct competition within existing layers.
• Embrace the role of the “professional” in the AI era. The need for deep, formal understanding of system specifications and product design won’t disappear; it will shift. Value will accrue to those who can articulate problems and guide AI systems to solve them correctly.
• Evaluate AI tools, especially coding agents, by their capacity for error correction. Agents work best in domains like code generation where there are built-in linters, compilers, and tests to catch and correct mistakes within the iterative loop.
• Think of developers as a growing, empowered consumer base. Their adoption patterns and preferences are central to infrastructure now, and winning their attention requires strategies reminiscent of consumer products.