Key Takeaways

  • Anthropic reportedly cut its target public valuation from $2 trillion to $1.5 trillion as percentage growth cools ahead of an IPO.
  • Rory O'Driscoll argues that reaching a $20 billion run rate and adding $5 billion per quarter yields $100 billion to $110 billion in forward revenue, justifying a 12x to 13x multiple.
  • Jason Lemkin counters that foundation model revenue lacks the retention of traditional enterprise software because developers route calls strictly on price and speed.
  • Replit proved this vulnerability by dropping Claude as its default model for OpenAI simply because OpenAI was cheaper and performed just as well.

The Disagreement

Can an AI foundation model command a public software multiple if its buyers feel zero platform loyalty?

Rory O'Driscoll looks at the raw math and sees a path to a historic public offering. If Anthropic enters next year at a $20 billion run rate and continues compounding, the revenue scale justifies an eye-popping valuation without requiring heroic assumptions. As O'Driscoll puts it: “If you start next year which is when you go public, if you're at a $20 billion run rate adding five a quarter, then you do 100, 105, 110, and now you're saying it's 12, 13 times forward revenues for the most exciting company in the world. It doesn't feel crazy. It doesn't take a lot of squinting to make that work.”

Jason Lemkin sees a fragile house of cards built on rented volume rather than committed enterprise software contracts. Traditional software revenue repeats because ripping out Workday or Salesforce carries unbearable switching costs. In contrast, calling an LLM API requires zero emotional or technical commitment. Builders route to whichever endpoint delivers the best balance of cost and latency today.

Lemkin warns that treating model usage like annual recurring revenue misreads the entire market: “There's only so much predictability in this revenue. There's only so much R in the ARR here. It wouldn't surprise me if at the end of next year, Anthropic's market share is down 10% from where it is today.”

He points to developer tools switching vendors overnight as proof. When developer platforms like Replit swap underlying providers, they do not hesitate. Lemkin highlights the cold calculation behind the switch: “He said we're going from Claude to OpenAI because it's just as good and materially cheaper. That's all he cares about.”

Who's Right (and When They're Wrong)

O'Driscoll is right only as long as aggregate enterprise spending on generative AI expands faster than model providers cut prices. When overall demand doubles every six months, every competent model provider looks like a compounding software giant. Capacity shortages hide commoditization.

Lemkin is right the second capacity catches up to demand. Foundation models sell compute-heavy commodity tokens. Unlike databases or ERP systems, an API call creates zero workflow lock-in. When Replit flipped its default away from Claude, it took engineering hours, not a multi-year migration project. If OpenAI, Google, or Meta drops pricing or matches benchmark output on code generation, developer traffic shifts over a weekend.

Anthropic can only defend traditional software multiples if it converts raw API volume into sticky workflow applications and proprietary enterprise knowledge graphs that competitors cannot easily match.

What to Do With This

Audit your model vendor dependencies before this Friday. If your product relies on a single proprietary model API without a routing layer, write a fallback wrapper that lets you switch between Claude, OpenAI, and open-source models with an environment variable. When your provider tries to hike rates or suffers an outage, you should be able to redirect your token traffic in under ten minutes.