A developer who examined 200 funded AI startups reported that 73% of them rely on external models while marketing proprietary systems.

Sample and approach

The researcher reviewed product integrations, API usage, and deployment details to determine whether companies used internally trained models or third-party services.

Coverage focused on funded ventures, with technical traces and behavioral tests employed to identify instances of model reselling and light modification.

Common implementation patterns

Surveyed projects typically repackaged public or commercial models rather than delivering independently trained foundational systems to end customers.

• Direct wrappers around ChatGPT-style APIs, presented under the startup’s brand.
• Claude-based solutions combined with curated prompt libraries and task-specific templates.
• Lightly adapted open models sourced via Hugging Face, with minimal fine-tuning applied.
• GPT-4 access concealed behind system prompts and bespoke interfaces to imply proprietary capability.

Observations and reporting

The analysis highlights a frequent gap between product messaging and underlying model provenance, particularly among recently funded startups and early-stage offerings.

Documentation and marketing often emphasize unique algorithms while technical indicators point to third-party inference or modest model adjustments.

What the study states

According to the developer, many services deliver user-facing novelty through prompts, UI changes, and integrations rather than through independently trained model weights.

The report presents counts and examples from the examined sample without assigning commercial or legal judgments about those implementation choices.