The Mechanics of the ‘Wrapper Death Spiral’
The ‘wrapper death spiral’ is not merely a market sentiment; it is a structural failure mode inherent to startups that lack infrastructure independence. It occurs when a startup’s core value proposition is subsumed by a feature update from a foundational model provider. For instance, startups that built businesses solely around ‘chatting with PDFs’ saw their valuations collapse overnight when OpenAI integrated native file analysis.
Investors are now assigning lower valuation ceilings to agents that lack infrastructure independence. If a startup’s roadmap can be broken by an API update from a single provider, it possesses zero technical leverage. The market has shifted toward valuing ‘Deep Tech’ AI, defined as companies that build proprietary orchestration layers, maintain multi-model compatibility, and own their data feedback loops.
If a roadmap can be broken by a single API update, the startup has zero technical leverage.
Valuation Divergence: Wrappers vs. Deep Tech
The divergence in valuation multiples between wrappers and deep tech AI is widening. Seattle-based AI investors, for example, have noted that pre-seed rounds for genuine deep tech AI startups now command premiums of 40% over non-AI or wrapper companies, driven by the recognition of legitimate compute costs and the scarcity of engineering talent capable of building proprietary models.
The most valuable data is not the document, but the log of how a human expert corrects the AI.
The Investor’s Lens: Wrapper vs. Deep Tech Indicators
| Evaluation Criteria | “Wrapper” Startup (High Risk) | Deep Tech / Defensible Startup (Investable) |
|---|---|---|
| Model Dependency | Hardcoded to a single provider (e.g., OpenAI API). Vulnerable to vendor lock-in and price hikes. | Model-agnostic orchestration layer; capability to route tasks between models (e.g., GPT-4 for reasoning, Phi-4 for speed). |
| Data Strategy | Relies on user prompts and base model knowledge. No proprietary retention. | Proprietary data lake; continuous fine-tuning pipelines; Reinforcement Learning from Human Feedback (RLHF) loops. |
| Customisation | System prompts (text instructions) only. | Fine-tuned weights (LoRA/QLoRA); RAG with proprietary vector indexing and re-ranking logic. |
| Workflow Depth | Single-turn Q&A or simple summarisation. | Multi-step agentic workflows; deep integration with legacy ERP/Banking systems; State management. |
| Defensibility | User Experience (easily copied). | Workflow stickiness; Volume of proprietary interaction data (The “Data Moat”). |
The ‘Data Moat’ Reality Check
While ‘data is the moat’ has become a venture cliché, the mechanical reality of how data constitutes a defence has evolved. In 2025, access to public data is meaningless; the Common Crawl datasets used to train base models are available to everyone. The competitive advantage lies in the proprietary data pipelines that feed these models, specifically ‘data loops’ that capture user interaction to refine model behaviour over time.
The Three Pillars of a Modern Data Moat:
1. Private User Interaction Logs: The most valuable data in 2025 is not the document itself, but how a human expert interacts with it. For a Fintech startup, the ‘moat’ is not the access to SEC filings (which are public), but the logs of how a senior credit analyst corrects the AI’s initial summary of those filings. This correction data serves as the ‘Gold Standard’ for fine-tuning and RLHF, creating a compounding advantage that a generic model cannot replicate.
2. Domain-Specific Transaction Data: High-frequency, proprietary feeds of financial data that are absent from public training sets. Startups that have negotiated exclusive access to ‘dark data’, such as private equity transaction details or unlisted real estate financials, possess a defensible asset that inoculates them against commoditisation.
3. Edge-Case Error Repositories: Generalist models perform well on averages but fail on edge cases. A startup that has systematically collected thousands of failure modes specific to a vertical (e.g., detecting a specific type of synthetic identity fraud in neo-banking) can train a specialised model that outperforms GPT-4 in that narrow but critical slice.
Auditing the Team: Engineering Maturity Metrics
A critical, often overlooked component of technical due diligence is assessing the maturity of the engineering team. In an era where GitHub Copilot can generate boilerplate code, the ability to build a demo is democratised. However, the ability to build a scalable, compliant system remains rare.
The ‘Code-Level’ Competency Check: Investors must look for specific tooling in the startup’s stack that indicates a mature understanding of AI risks. AI-Code Detection: The presence of tools like span-detect-1 implies the team is aware of the risks of ‘shadow AI code’, which is code generated by AI that may contain vulnerabilities or licensing issues. Mature teams use these tools to track the provenance of their codebase. MLOps Rigour: Does the team use version control for data as well as code? Tools like DVC (Data Version Control) are non-negotiable for reproducibility. If a startup cannot reproduce a specific model version from six months ago because they overwrote the training data, they are failing a basic engineering competency test.
Red Flags in Team Dynamics: Lack of Evaluation Pipelines: If the CTO cannot quantitatively demonstrate how their model has improved over the last quarter using a consistent benchmark (e.g., ‘Our F1 score on fraud detection improved by 4%’), they are operating on intuition rather than engineering rigour. Ignorance of Technical Debt: Founders who cannot articulate their strategy for managing ‘glue code’ and ‘hidden feedback loops’ are likely building fragile systems that will collapse under scale. The ‘buy now, pay later’ nature of Machine Learning technical debt is a primary cause of post-Series A failure.
Conclusion: The Investor’s Mandate
The era of investing in ‘AI for X’ is over. The mandate for 2025 is to invest in ‘Architecture for X.’ Investors must peel back the UI layer to inspect the plumbing. Is there a proprietary orchestration engine? Is there a data loop that gets smarter with every user interaction? Is the team disciplined in MLOps? Startups that answer ‘yes’ are building the industrial infrastructure of the future. Those that answer ‘no’ are merely renting intelligence from tech giants, destined for the wrapper death spiral. For the sophisticated investor, the alpha lies in the ability to tell the difference.
