Will OpenAI Swallow the Application Layer? a16z Says Real Opportunities Lie Outside General Models
As large language models (LLMs) from companies like OpenAI and Anthropic become more powerful, many fear they will dominate the AI application layer, leaving no room for startups. However, this article argues that the real opportunity lies not on the "Yellow Brick Road"—the high-profile, general-purpose tasks like code and text generation that model labs are directly pursuing—but in the "rest of Oz": complex, vertical-specific applications.
On the Yellow Brick Road, model companies have inherent advantages: control over the model, better margins, pricing power, and strong distribution. Startups building generic, horizontal "co-pilot" tools for standard tasks are competing directly on this path and are vulnerable.
True defensibility and value are found in specialized, vertical applications. These involve deep integration into messy, multi-step business workflows (e.g., sales, insurance, legal), handling legacy systems, data quality issues, compliance, and governance. The "scaffolding" around the model—the specialized tools, automations, workflows, and industry knowledge—becomes more critical than the raw model power itself.
Vertical AI companies can build defensible moats through:
* **Data & Learning Flywheels:** Capturing unwritten industry practices and specific customer feedback not found in public training data.
* **Managing Model Complexity:** Routinely evaluating and routing queries across multiple models (including open-source) to optimize for performance and cost, and absorbing the migration burden of model upgrades for clients.
* **Cost Optimization:** Using cheaper, fine-tuned models for specific sub-tasks instead of always calling the most expensive, general-purpose model.
* **Governance & Compliance:** Providing the control plane for permissions, auditing, and ensuring compliance with industry-specific regulations (e.g., HIPAA, FINRA).
Examples from sales (11x) and insurance (FurtherAI) illustrate that clients pay for systems that drive specific business outcomes (e.g., sales pipeline, policy underwriting), not for generic intelligence. These systems become the "operational memory" of a business, a layer that is hard to replace, even as the underlying LLMs commoditize and improve.
To test if a startup is building in the "rest of Oz," it should pass checks like the **Tool & Steps Test** (requires complex, multi-step workflows), the **System Test** (owns the end-to-end workflow, not just a tool on top), and the **Hedge Fund / P&L Test** (measured by client business outcomes, not benchmark scores).
Both model labs and vertical application companies will win. The next generation of enterprise software will be built in the specialized, complex, and high-value territory beyond the Yellow Brick Road.
marsbit1 h fa
