The AI infrastructure buildout gets the headlines, but the software layer is where margins and durability live. We map the AI software stack and identify where value is most likely to accrue over the next 3-5 years.
The AI infrastructure buildout gets the headlines, but the software layer is where margins and durability live. We map the AI software stack and identify where value is most likely to accrue over the next 3-5 years. The narrative around AI investing has been dominated by infrastructure — GPUs, data centers, power. And for good reason: the capex numbers are staggering, and the picks-and-shovels thesis has been the dominant investment framework. But the software layer is where the long-term value creation will happen. Infrastructure is necessary but insufficient. The companies that build durable moats will do so through software, data, and workflow integration — not hardware. ## The AI Software Stack We break the AI software stack into five layers, each with distinct competitive dynamics and value capture potential: Layer 1: Foundation Models — The model providers (OpenAI, Anthropic, Google, Meta). This layer has the most capital deployed and the most uncertain economics. Models are increasingly commoditized, differentiation is temporary, and the capital requirements are enormous. Layer 2: Model Infrastructure — The tooling for training, fine-tuning, deploying, and monitoring models (Weights & Biases, Hugging Face, Anyscale, Modal). This layer benefits from the complexity of production ML systems. Switching costs are moderate but growing as teams build workflows around specific platforms. Layer 3: AI-Native Applications — Applications built from the ground up around AI capabilities (Harvey for legal, Cursor for coding, Glean for enterprise search). This is where the most interesting value capture dynamics exist today. Layer 4: AI-Enhanced Incumbents — Existing software companies adding AI features (Salesforce, Adobe, Microsoft). These companies have distribution and data advantages but face organizational challenges in fully leveraging AI. Layer 5: Orchestration & Integration — The middleware connecting models to enterprise data and workflows (LangChain, LlamaIndex, various agent frameworks). This layer is critical but faces