73% of AI startups are secretly just OpenAI wrappers, according to a viral reverse-engineering study
Nov 24, 2025
Key Points
- A reverse-engineering study claims 73% of AI startups misrepresent their technology, wrapping OpenAI or Anthropic APIs while marketing proprietary models, though the methodology's technical credibility remains disputed.
- Wrapper startups charge 75x to 1000x markups over API costs, making the business model economically viable but raising fraud concerns if investors are pitched on custom model development.
- The core issue is transparency: founders privately acknowledged the proprietary framing felt disingenuous, and VCs worry undisclosed API dependency could constitute securities fraud.
Summary
A reverse-engineering study claimed that 73% of AI startups misrepresent their technology as proprietary when they actually call OpenAI or Anthropic APIs directly. An anonymous author analyzed 200 startups by examining network traffic, API fingerprints, and rate-limiting patterns. Of those 200 companies, 54 made accurate technical claims or were transparent about third-party dependencies. The remaining 146 claimed proprietary AI but were found to be using OpenAI or Anthropic under the hood.
The author identified OpenAI's exponential backoff pattern (1 minute, 2 minutes, 4 minutes, 8 minutes) as a fingerprint. The breakdown showed 19% of companies claimed in-house models but used fine-tuned public models instead, and 8% had custom ML pipelines that wired together standard cloud services. The author acknowledged that training private language models would be unrealistic and economically irrational for most startups.
The economics of wrapper businesses are stark. GPT-4 API costs 3¢ per query, but one startup charged $3 or $300 per month for 200 queries, a 75x markup. Another case showed margins closer to 1000x when including Pine Cone embeddings. These margins favor the wrapper companies themselves.
A semantic dispute emerged over fine-tuning. The author found 45% of companies claiming fine-tuned models were actually using OpenAI's fine-tuning API. Critics pushed back that if a startup fine-tunes OpenAI's model and produces a unique version, that is legitimately their artifact even if they did not train the base model. The analogy offered was that customizing a store-bought sandwich still makes it yours.
Seven founders reached out privately after the post. Some were defensive. Others asked for help reframing their marketing from proprietary AI to built with best-in-class APIs, acknowledging the proprietary tagline felt disingenuous. A VC also contacted the author, saying they had been told they were investing in companies training their own AI. If misled, that could constitute securities fraud. The core issue is transparency with investors and customers about the actual tech stack. A startup pitching investors on proprietary models while wrapping OpenAI is a real problem. A customer who believes they are paying for custom AI scientists but receiving repackaged ChatGPT might simply buy ChatGPT directly from OpenAI instead.
The methodology itself raised skepticism. The author claimed to observe backend calls to api.openai.com/v1/chat/completions from the frontend via Chrome inspect tools, which would normally expose API keys as a security vulnerability. Standard secure architecture has the frontend call the startup's backend, which then calls OpenAI. How this observation was actually made remains unclear, and the credibility of specific technical claims stays somewhat opaque.
The broader insight holds: many startups wrap third-party APIs and use proprietary language in marketing and pitch decks. Whether that is inherently fraudulent or simply the standard way the AI services market works, where most value comes from UI, workflow, and distribution rather than model training, remains contested.