.png)
The Data Moat 2.0: Escaping LLM Commoditisation by Owning the Endless Feedback Loop
- Didier Vila
- Nov 19
- 3 min read
By Didier Vila, PhD. Founder and MD of Alpha Matica.
The foundational shift in the AI economy has already happened: raw intelligence is commoditising fast. Frontier labs are releasing models that are “good enough” for most tasks, open-source alternatives are catching up within months, and inference prices are collapsing.[1]
For any company that wants to build a durable AI business (or defend an existing one from the “LLM tsunami”), the battleground has permanently shifted upward: from the model layer to the application layer and, above all, to the data-feedback layer.[2]
The new defensible moat is no longer a static pile of pre-training data. It is a living, proprietary feedback loop that continuously captures, refines, and redeploys human judgment at scale. Call it **Data Moat 2.0**.[3]
The Rise of Dynamic, Proprietary Data Flywheels
General internet-scale data is abundant and increasingly shared. Specialised, high-signal data is not—and it is the only kind that still moves the needle in vertical applications.[4]
The highest-quality data is generated in real time inside closed-loop systems through structured Human-in-the-Loop (HITL) workflows. Every meaningful user interaction becomes a proprietary training signal.[5] These traces are invisible to the outside world. When systematically captured, labeled, and fed back into domain-specific fine-tunes or preference models, they teach the system nuances that no public dataset can. The result is a classic flywheel: more usage → more proprietary data → better performance → stickier product → even more usage. After 12–24 months of compounding, the gap often becomes uncrossable.[6]
The Alignment Moat: Turning Human Judgment into Process Power
Statistical accuracy is table stakes. In enterprise settings, the AI is frequently the public face of the organisation. Here, alignment with brand voice, regulatory constraints, and ethical red lines is non-negotiable.[7]
Reinforcement Learning from Human Feedback (RLHF)—and its newer variants (RLO, DPO, etc.)—is the primary tool for baking that alignment in. But the moat is not the technique itself; it is ownership of the feedback dataset and the operational discipline to keep refreshing it.[8]
Companies that treat alignment as a one-time post-training step get left behind. Leaders treat it as a continuous risk-management process, creating an **Alignment Moat** that generic providers cannot replicate without violating their own customers’ privacy or neutrality promises.[9] In regulated industries, this moat doubles as a compliance moat.[10]
Operationalising the Flywheel—and Weaponising Switching Costs
Building a Data Moat 2.0 is an engineering and organisational challenge, not a research one. Winning companies treat it like core infrastructure from day one.[11] This discipline is expensive upfront but pays off asymmetrically. After a critical mass of proprietary data (typically 50k–500k high-quality interactions, depending on domain), the tailored model materially outperforms any generalist alternative. Switching providers then becomes effectively prohibitive.[12]
Conclusion
The LLM era is not ending; the base-model arms race is. The winners of the next decade will be the organisations that transform their daily operations into proprietary data factories—owning living datasets no one else can access and performance curves that keep steepening while everyone else’s flatten.
In short: the new moat is not the model.
It is the closed-loop ownership of human preference, at scale.
References
[1] Epoch AI – “LLM inference prices have fallen rapidly but unequally” (March 2025)
[2] Martin Casado (a16z) – “Who Owns the Generative AI Platform?” (January 2023)
[3] Alexandr Wang (Scale AI) – original “The Data Moat” post (2021)
[4] Epoch AI – “Will we run out of data? Limits of LLM scaling based on human-generated data” (2024)
[5] GitHub Engineering – ongoing posts on Copilot’s proprietary feedback loops
[6] GitHub Copilot Workspace launch and flywheel documentation (2024–2025)
[7] Anthropic Economic Index – September 2025 Report
[8] Paul F. Christiano et al. – “Deep reinforcement learning from human preferences” (2017)
[9] Anthropic – “Claude’s Constitution” and Constitutional AI updates (2023–2025)
[10] Hippocratic AI, Nabla, Abridge, and similar healthcare AI providers – industry reports & public positioning on auditable human oversight (2024–2025)
[11] Jerry Chen (Greylock) – “The New New Moats” (2024)
[12] GitHub Copilot enterprise case studies: after 12–24 months of accumulated interaction data inside customer codebases, internally fine-tuned instances consistently outperform any publicly available model, making migration practically infeasible (Microsoft & GitHub reports, 2024–2025)
