73% of AI startups are just prompt engineering (pub.towardsai.net)

An independent reverse-engineering investigation of 200 funded AI startups found that 73% (146 companies) are essentially relabeling third‑party LLMs — mostly OpenAI’s models and Anthropic’s Claude — behind new UIs and prompts. The reporter traced network traffic, decompiled code, and followed API calls, uncovering frequent, direct calls to external LLM APIs even from companies that pitched proprietary model infrastructure to investors. The discovery began casually while debugging a webhook and snowballed into a broad pattern of “model-as-a-service + UX” businesses rather than new model engineering. This matters because it reframes where value and risk live in the ecosystem: most startups’ IP is prompt engineering, UX, data pipelines, and fine‑tuning choices, not novel model architectures or weights. That concentration creates vendor lock‑in, supply‑chain and privacy risks (sensitive data being routed to third‑party APIs), and thinner defensibility for high valuations. For builders and investors it underscores the need for clearer disclosure, robust data governance, cost/latency analysis, and genuine differentiation (e.g., proprietary data, fine‑tuned models, on‑prem deployments, or vertical specialization) if a startup hopes to escape the “repackaged LLM” bucket.