A new investigation into 200 funded AI startups has found that 73% do not run their own models, despite publicly claiming to have proprietary systems. The researcher analyzed network traffic, decompiled code, and traced API calls, uncovering widespread reliance on OpenAI, Anthropic’s Claude, and lightly edited open-source models from platforms like Hugging Face.
Many startups were found to be calling ChatGPT directly, wrapping Claude with template prompt libraries, or using off-the-shelf models with minimal adjustments while positioning their products as breakthrough AI. Some even masked basic GPT-4 outputs behind system prompts and marketing language.
The findings highlight a widening gap between investor-facing narratives and the technical reality of how AI products are built today. Analysts say the trend reflects rising pressure to scale quickly, as well as the high cost of building proprietary models. The report suggests that differentiation across the ecosystem is thinner than it appears – which may pose risks for customers, investors, and the long-term credibility of AI startups.
