{
  "id": 15725400,
  "title": "When the model takes away the moat",
  "url": "https://urgent.news/2026/10/20/when-the-model-takes-away-the-moat",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-20T08:58:44.000Z",
  "source": {
    "name": "TechRadar",
    "slug": "techradar",
    "url": "https://www.techradar.com/pro/when-the-model-takes-away-the-moat"
  },
  "original_language": "en",
  "account": "The AI landscape has undergone a surge in the past four years, with numerous products emerging from startups, enterprise software firms, and even established corporations. These AI tools typically rely on APIs from leading AI labs or open-weight alternatives. While this approach has fostered a proliferation of AI offerings, it has also introduced structural vulnerabilities that could be addressed through the democratization of model creation and training.\n\nAn AI wrapper is a product that provides an interface atop a third-party foundation model's API, with minimal proprietary technology beneath it. An example is Jasper, which peaked at a $1.5 billion valuation before its revenue plummeted from $120 million in 2023 to $55 million in 2024 due to ChatGPT's free writing features. Many AI startups that ceased operations in 2025 were wrappers and applications designed around LLM front-ends rather than proprietary data or infrastructure.\n\nNot all wrappers are created equal. Thin prompt-and-display wrappers differ significantly from products that incorporate heavy retrieval, tool orchestration, agentic workflows, or domain-specific fine-tuning. These latter products may possess genuine defensibility even when built atop a third-party base model. Some companies employ multiple foundation models or fine-tune/distill their own smaller models for specific tasks after prototyping on a frontier API. Open-weight models also diversify the wrapper landscape, as not all wrappers call closed-source frontier models.\n\nThe wrapper problem persists regardless of an AI product's complexity. When a company's core value proposition hinges solely on an API call to a third party, it lacks control over pricing changes, rate limits, deprecations, model behavior shifts, or the provider's potential direct competition in the same vertical. This risk is not theoretical, as companies' vulnerabilities depend on how much of their value resides above the model layer. The defensibility of an AI wrapper often hinges on intellectual property. If a wrapper's distinctiveness is observable and replicable by a competent competitor within a limited time, it is likely to be emulated, leading to a fleeting competitive advantage.\n\nAcquiring model ownership is one potential solution, yet it remains elusive for many companies. The high costs associated with training frontier-competitive models—enormous compute budgets, specialized talent, proprietary or licensed data pipelines, and infrastructure for reinforcement learning—render model creation inaccessible to all but the most well-funded entities. While open-weight models and democratized hosting and fine-tuning are encouraging trends, they still necessitate technical expertise and compute infrastructure that most startups lack. True democratization of model ownership extends beyond training to encompass the entire training layer, akin to the accessibility AI has attained for end users. Democratizing model training and ownership could enable founders to build businesses that resist being absorbed by model providers as mere features. Ultimately, the central issue lies in making the ability to train AI models widely accessible, ensuring value, control, and responsibility are distributed rather than centralized.",
  "summary": "The ability to train an AI model should be fully democratized.",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}