Last year, launching an AI startup meant building models. Now it often means connecting an OpenAI API key to a Next.js frontend and adding Stripe billing. The product is the markup.
This criticism has been building for months across the developer community, and it cuts at a real tension in the current AI landscape. Companies like OpenAI, Anthropic, and Google are building the actual AI. A layer of startups sits on top, reselling access to those models with a UI and a brand.
The math is brutal. GPT-4o costs roughly $2.50 per million input tokens through the API. A wrapper startup charges $20-50/month per user, handles maybe a few hundred thousand tokens per user, and pockets the difference. The value proposition isn't the AI - it's the interface, the workflow integration, or the specific prompt templates baked in.
Some wrappers genuinely earn their margin. Cursor takes base models and builds a deeply integrated coding environment around them. Jasper tailors the experience for marketing teams with brand voice controls and campaign workflows. These products would take real engineering effort to replicate.
But a significant number of AI startups amount to a chat interface with a system prompt and a credit card form. When the underlying model improves or the API provider ships a competing feature, these companies have nothing to fall back on. OpenAI adding custom GPTs wiped out dozens of thin wrappers overnight. Google adding Gems did the same.
The defensibility question isn't new to software - plenty of companies have been built on top of AWS or Twilio APIs. The difference is that AI model providers are unusually aggressive about expanding into their customers' use cases. OpenAI doesn't just sell the API; it also runs ChatGPT, a consumer product that competes directly with many of its API customers.
For anyone evaluating AI tools, the wrapper question is a practical one: if this startup disappeared tomorrow, could you get 90% of the same result by using ChatGPT or Claude directly? If yes, you're paying for convenience, not capability.