If you're building on top of GPT or Claude without doing your own AI research, you're not really building a company. You're building a feature request.

The dynamic is straightforward but brutal. A startup may build some clever AI application, maybe it helps people write emails, or generates marketing copy, or summarizes documents. It gets traction, users love it, revenue grows, they raise their series A at breakneck speed. But all this startup has actually done is prove to OpenAI that there's demand for this specific use case. And OpenAI has 100 million people using ChatGPT. The startup has maybe 10,000 paying customers.

Who wins when OpenAI decides to add email writing native in ChatGPT as a chrome extension?

This pattern has repeated so many times it's becoming predictable. Fine-tuned models replaced by GPT's marketplace. Prompt optimization shipped natively in Claude. OpenAI deploying their own consultants. Each time, the platform absorbed what had been a separate category.

Why do the platforms always win these battles?

Distribution is the most obvious answer. When OpenAI ships a feature to ChatGPT, it reaches 100 million users instantly.

But there's a second factor that's harder to see: authority. When OpenAI says they can do something, the people that matter believe them. They built GPT. To enterprises, OpenAI understands how to use ChatGPT better than anyone else.

When a startup claims they can do the same thing, there's skepticism. Are these people really better at AI than the company that created the underlying model? Probably not.

And this authority gap is almost impossible to overcome through better UX or marketing. It's like a third party windows app trying to compete with a feature Microsoft decides to build into the operating system. Even if the third-party app is superior, most users will default to whatever Microsoft ships, because whatever Microsoft ships comes with the computer when you buy it.

The irony is that success makes the problem worse. The more revenue a startup generates from their AI application, the more they validate the market opportunity for the platform providers. They're essentially conducting free market research for their biggest potential competitor.

And the platforms can see everything. API usage patterns, revenue growth, customer adoption. It's all visible to the companies providing the underlying models. They know exactly which applications are working and how much money is involved.

So what's the solution?

The companies that are surviving this dynamic aren't the ones with better interfaces or better prompt optimization, or shipping someone else’s tool in a new form. They're the ones building their own models, like Midjourney. They remain relevant in image generation, even in the face of Chinese Minimax and HaiLuo because they control their own models. They own their stack.

But the argument is not that companies need to build GPT-5 to have a defensible position. Startups need to build something that gives them capabilities incumbents can't easily replicate. Maybe that's specialized models for a particular domain, or new model architectures, or something else entirely.

The crucial element is doing research, not just engineering.

This is difficult advice to follow, because research is expensive and uncertain. It's much simpler to build an elegant interface around existing APIs and focus on user acquisition. But simple doesn't mean sustainable.

There's a historical parallel worth considering. In the early days of mobile computing, many companies built applications that provided features later absorbed into iOS and Android. The ones that survived either operated in niches too small for the platform providers to notice, niches too degenerate or inaccessible to incumbents, or they developed technical capabilities that were uniquely difficult to replicate.

The same dynamic is playing out in AI, but faster and with higher stakes.

The companies that will thrive in the next phase are those that start investing in AI research before they're forced to. Waiting until ChatGPT copies your core feature is waiting too long.

This creates an uncomfortable reality for AI entrepreneurs. Building applications on existing models is the fastest path to initial traction, but it's also the most dangerous long-term strategy. The platforms have overwhelming advantages in distribution and authority, plus complete visibility into what's working.

Which means if you want to build an AI company, you better start learning how to build models. Because that's what you're going to end up doing anyway.

7/30/2025