Key Takeaways

  • Cursor's revenue from Anthropic once hit a staggering 40-50%, demonstrating extreme platform dependency risk for companies building atop foundational AI models.
  • Anthropic allegedly described its 'Claude Code' as a mere "research effort" before making it a product, highlighting the volatile and often competitive nature of relationships with upstream AI partners.
  • Every hyperscaler and SaaS product is developing its own Large Language Models, turning AI model providers into direct competitors for companies operating in the "token path."
  • To build a defensible business in the AI era, founders must own the end customer relationship and develop proprietary value beyond the base capabilities of AI models.

The AI Gold Rush: Who Owns the Mine, Who Just Digs?

Imagine building a product so successful it makes up nearly half of your foundational technology provider's revenue. That was Cursor, the AI coding assistant, and its relationship with Anthropic. John Coogan, host of TBPN, dropped a bomb: “Interestingly, Curser once made up 40 to 50% of Anthropic's revenue. That is a crazy girration in the market. Just shows you how quickly things are changing.” A massive win, right? Not exactly. This kind of dependency became a critical vulnerability when Anthropic's own strategy evolved.

The real sting came when Anthropic, a key partner, pivoted its own efforts. Coogan reflected on the tension: “Nothing tells that more than Anthropic telling Cursor that Claude Code was just a research effort. This goes back to the Dylan field like were they consistently candid or maybe they really just did think hey yeah this is just a research effort like you do you and then realized later like whoa whoa we need to be a player in this we cannot be in the middle of this situation so we got to do our own thing.” Whether by design or necessity, the message was clear: what's a research project today can be a direct competitor tomorrow. For any founder building on a platform – especially one as rapidly evolving as AI – this story is a cold shower.

Your Product Is Their Feature

The Cursor-Anthropic saga isn't an isolated incident; it's a stark preview of the AI market's core challenge for ambitious builders. Coogan posed the central question for anyone relying on external models: “Do not build a company that depends on the models getting better, but if the models get really really good, like what can't they do?” The fear is real: your innovative product today could become a trivial feature for an underlying model tomorrow. Every major tech player, from hyperscalers to established SaaS giants, is now integrating and developing their own LLMs. “Every company competing with every other company as you know every hyperscaler has an LLM at this point and every SAS product has other products because they can launch features faster,” Coogan observed. This means the 'token path' – the direct reliance on general-use AI models for your core value – is inherently unstable.

The only durable defense, as Jordi Hays pointed out, is “owning the end customer relationship.” In a world where AI models are rapidly commoditizing capabilities, your true moat isn't just about using the best model. It's about building unique interfaces, proprietary data, community, or workflow integrations that create an unbreakable bond with your users. If your value proposition is simply 'Model X, but slightly better UI,' prepare to be eaten alive.

What to Do With This

This week, immediately audit your reliance on foundational AI models. Map out which core features and revenue streams are directly tied to a single LLM provider's API. Ask: What percentage of your perceived value could be replicated if that provider built a similar feature directly into their offering? Next, identify three distinct areas where you can build proprietary value: unique data sets, exclusive UI/UX that isn't easily copied, or a community that keeps users sticky. Pick one to prototype within the next 30 days. Finally, prioritize diversifying your LLM dependencies. Set a technical goal to integrate and test at least two alternative models for your core functions. This isn't just a technical task; it's a critical hedge against existential platform risk.