Is your AI startup building on quicksand or bedrock? Lucas Swisser, co-head of growth investing at CO2, cuts through the hype to show where the real money is made in AI. He's not just talking hypotheticals; CO2 is a major investor in frontier LLMs like Anthropic and OpenAI, alongside application companies like Harvey and Cursor (recently acquired by SpaceX).

Swisser reveals a stark reality: Anthropic alone reportedly pulls in over $50 billion in annual recurring revenue. That figure, he notes, is ten times larger than the sum of all independent AI application companies combined. This isn't a doomsday call for applications, but a clear signal: defensibility isn't automatic. You need a specific strategy to capture value against the raw power of foundational models.

Key Takeaways

  • Anthropic's widely publicized $50 billion ARR alone overshadows the total revenue of all independent AI application companies, showing the immense value being captured by frontier LLMs.
  • Value accrual in AI isn't a binary choice; it's a spectrum, with some application categories at high risk of direct LLM dominance.
  • Fields like coding are highly exposed to LLM disruption, as the application's output is often too close to the model's core function.
  • Application companies can build defensibility through unique data network effects, complex integrations, and the need for multi-model usage, particularly in regulated industries like legal.
  • Use Lucas Swisser's Framework for AI Application Vertical Viability to stress-test your startup's potential for sustainable value.

The Lucas Swisser's Framework for AI Application Vertical Viability

Lucas Swisser offered a direct framework to assess this market. Here’s his rubric:

Component 1: Proximity to the Model: How close is the application to the model itself? what categories are the application actually just the output of the model... in coding the output of the model the application is really tied very closely to the model itself in that part of the market you really think maybe the LLMs themselves will take the vast majority of the market.

Component 2: Integration Complexity & Unique Requirements: something else where maybe there's a lot of meat in between the application that you're trying to build and the model itself That may be that you have to do very deep integrations. It may be that there's a lot of compliance risk. It may be that there's unique data that you have to access.

Component 3: Need for Multi-Model Usage: Are you really in an independent situation where the the applicationoriented company or the application layer company is going to need to use multiple models at the end of the day... They want to be able to use multiple models. We've heard a lot about token maxing, right? And token minning over the last few weeks. They want to be able to control cost. They want to be able to control compliance. They want to be able to control security.

Component 4: Data as a Network Effect: Is the data itself a network effect and is the model actually improving itself based on the inputs... As you feed in more legal information, the models get better. Why? Because legal is game theory.

When This Works (and When It Doesn't)

Swisser's framework works best when evaluating specialized, enterprise-facing AI applications. It's especially useful for founders targeting industries with high compliance, proprietary data, or complex workflows. For example, developers, he notes, often care more about the frontier LLMs themselves. Simple tasks, like querying an HR database about PTO policy, might not demand a top-tier model like Fable 5; a lighter model like Haiku or Sonnet could do the job. The framework struggles when AI capabilities change rapidly, or when multi-model switching becomes a trivial feature of LLM APIs, eroding the value of abstracting away the underlying models.