Moat

Definition

A sustainable competitive advantage that prevents rivals from replicating your position. In AI, moats can come from proprietary data, distribution, fine-tuned models, vertical expertise, or switching costs, but raw model capability is rarely a moat.

Why it matters

The 'moat' question is existential for every AI company. When the underlying models are commoditizing rapidly and the biggest tech companies are giving away free AI features, what makes your AI business defensible? The answer is almost never the model itself. Google's leaked internal memo 'We Have No Moat' (May 2023) argued that neither Google nor OpenAI has a durable model-quality moat. The real moats in AI are: proprietary data (what can you train on that others cannot), workflow integration (how deeply are you embedded in the user's process), network effects (does each new user make the product better), and domain expertise (do you understand the domain better than generalists).

In practice

Bloomberg's financial data moat justified training BloombergGPT. Tesla's driving data moat from millions of vehicles is irreplaceable. GitHub Copilot's moat is distribution: it is integrated into the IDE where 100M+ developers already work. Stripe's payment data moat enables fraud detection that competitors cannot match. OpenAI's moat is brand and developer ecosystem, not model quality (which competitors routinely match). The companies losing moat battles are the ones who relied on model access or thin UI layers: when ChatGPT launched, dozens of GPT wrapper startups saw their value proposition collapse overnight.