Why Vertical AI Startups Are Facing an Inevitable Downfall – The Bitter Lesson Revisited

01 – History Repeats Itself

Rich Sutton’s 2019 essay The Bitter Lesson observes a 70‑year pattern: approaches that rely purely on compute eventually dominate over hand‑crafted expert systems. The same dynamic is visible in today’s AI startup scene, where many YC Demo Day projects build vertical AI products by chaining fixed workflows and prompt engineering. These products can launch quickly but depend heavily on current model limitations and extensive engineering effort.

Key observations

General, compute‑driven methods have consistently won in AI research.

Founders are repeating the same mistake by over‑engineering domain‑specific solutions.

Stronger models will soon render workflow‑centric products obsolete.

The primary breakthrough comes from scaling compute, not from adding domain‑specific rules.

02 – Vertical AI Has No Cheap Advantage

General AI models will eventually surpass most vertical solutions. Users can integrate a general model via a simple API, incurring near‑zero migration cost—effectively hiring a remote colleague. Vertical AI products struggle to build durable moats because they cannot match the breadth, cost efficiency, and rapid iteration speed of a single, powerful model.

Applying Hamilton Helmer’s seven competitive advantages shows that vertical AI rarely secures scale economies, network effects, strong brand power, or low switching costs. The only plausible moat is an exclusive resource—unique data, patents, or infrastructure that cannot be replicated.

Product classification

AI products can be viewed along two axes:

Specialisation (Vertical vs General) : Vertical products focus on a narrow domain; general products aim to handle many tasks.

Control (Workflow vs Agent) : Workflow systems follow a fixed sequence of tools; agents decide autonomously which tools to use and when.

Examples:

Vertical workflow: a fixed pipeline that queries a database, summarises with a small LLM, refines with a larger LLM, validates, and generates a PPT.

Vertical agent: the same tools are available, but the LLM decides dynamically when to invoke each step.

General workflow: ChatGPT can perform part of the task but lacks domain‑specific constraints.

General agent: Claude Computer‑Use can operate desktop software directly, acting like a human collaborator.

03 – Six Predictions and Five Obstacles for AI Applications

By 2025‑2027, general agents are expected to become reliable enough to replace many workflow‑based solutions, reducing low‑skill hiring and shifting value toward general AI assistants. The timeline could be altered by the following risks:

Potential stagnation of model scaling.

Regulatory interventions after a major incident.

Trust issues caused by hallucinations and autonomous actions.

Reluctance of AI labs to open their most advanced models.

High inference costs that may keep some vertical solutions viable.

The value curve for AI‑application startups is likely “U‑shaped”: early engineering effort creates value, but as stronger models appear, that value erodes.

04 – Startup Strategies in a Shifting Landscape

Founders can consider four paths:

Build a vertical product only if they can lock down a defensive resource (exclusive data, patents, or infrastructure).

Wrap emerging general agents with an API layer—profitability is uncertain due to API costs.

Leverage open‑source models; they improve rapidly but still lag behind frontier labs on complex agent tasks.

Become a supplier or ecosystem contributor for AI labs (e.g., providing compute, data, or tooling).

Timing is critical: if general agents become competitive within a few years, vertical startups must pivot quickly or focus on truly exclusive resources.

05 – Conclusion

The “bitter lesson”—that raw compute eventually outperforms handcrafted solutions—appears to be resurfacing in AI entrepreneurship. Vertical AI products that rely on extensive engineering to compensate for current model limitations are likely to see diminishing returns as models improve. Sustainable advantage will require either exclusive resources or a shift toward building infrastructure that supports general AI agents rather than competing with them.

Code example

相关阅读：