Taste as a Moat

Stable Diffusion is open source. You can download it for free, today, and generate images at industrial volume. By the arithmetic of competitive strategy, Midjourney should have no defensible position — its underlying diffusion architecture is the same one its rivals use, often available without a license.

And yet Midjourney reached roughly $500 million in revenue in 2025 with a team of fewer than 200 and zero venture capital. Users recognize its output instantly. Not because the algorithm is secret, but because the curation behind it isn't reproducible. The product intuition lives in what David Holz and his team chose to emphasize, suppress, and refine across years of iteration. No open-source license distributes that.

That gap — between the model anyone can copy and the judgment no one can — is the most important competitive fact of the AI-Born era. It has a name: Taste as a Moat.

Why the obvious moats stopped working

For most of business history, defensibility was something you could point to. Land, in the agricultural economy. Factories, patents, and proprietary processes in the industrial age. Proprietary data and network effects in the information economy. The railroad barons who controlled routing through a mountain pass held an advantage that needed no fence — geography was the moat.

The reasonable executive's instinct is to look for the AI-era equivalent: the proprietary model, the secret training data, the technical edge. Steel-manning that instinct, it served well for two centuries. The trouble is that the central asset of this era refuses to behave like the old ones. Frontier model costs roughly doubled each year between 2020 and 2023, keeping the leading edge expensive — but accessible capability expanded just as fast. Harvey runs on commercially available LLMs; so does every BigLaw competitor willing to write the API call. When the engine is a commodity, owning a better engine isn't a moat. Everyone can rent the same one.

The reframe: judgment is the scarce input

Here's the inversion. When execution becomes cheap and models commoditize, the bottleneck moves to the one input that doesn't commoditize — the human capacity to distinguish good from merely adequate.

Taste, at its highest form, isn't pattern matching over past data. It's cultural foresight: knowing three months before the data shows it that maximalist layering has tipped from fresh to overdone; understanding why minimalism that elevates a luxury brand feels cold in a children's toy catalog; perceiving the gestalt of a design where curve, material, proportion, and context combine into a quality no checklist can decompose. Agents detect violations of existing aesthetic rules with ruthless precision. Only humans sense when the rules themselves need revision.

Figure: The moat isn't the generator anyone can run — it's the filter, the accumulated judgment about which outputs are actually good.

The mechanism: taste, scaled and written down

Taste only becomes a moat when you can scale it without diluting it. The architecture for that is a filter, not a generator.

A generative swarm produces a thousand variations. A Force Multiplier — the Player-Coach role in the The New Triumvirate: The Three Roles That Survive — trains "evaluator agents" on curated examples that encode her judgment, then reviews the cases the evaluators flag. Each judgment becomes training data. The system surfaces the top hundred; the human selects the best; the loop tightens. One person's aesthetic sense now governs outputs she never directly touches.

What makes this durable is that it compounds and gets recorded. Put the whole defensibility layer in one line and it reads simply: intent becomes the moat. The durable advantage is the accumulated record of judgment a company has written down — every edge case resolved, every charter refined, every decision logged on the audit spine. That's the moat now: time, written down. It's the one asset a competitor with identical models can't buy, because it was earned one closed loop at a time.

Tesla shows the same compounding in a different domain — by 2024, Full Self-Driving had logged over a billion miles, each generating edge cases rivals couldn't replicate without equivalent fleet scale. The mechanism is identical: not better engineers, but more accumulated, recorded judgment that competitors would have to live through to match.

The risk worth naming

I want to be honest about the assumption buried in all of this. The claim that taste stays durably beyond AI rests on a distinction holding over time — that anticipatory judgment about where human sensibility is heading requires cultural embeddedness, not just pattern sophistication. That's philosophically defensible today. But AI systems already pass aesthetic tests experts assumed required human sensibility. If genuine cultural foresight emerges in machines within ten to fifteen years, the moat framing shifts.

Two things survive that scenario. First, accountability: someone still has to answer for the brand. Second, and more interesting, pedagogy. Even if AI can exercise taste, it can't transmit taste through the collaborative doing that builds craft in another person. The Force Multiplier's deepest product isn't the artifact — it's the team that can build the next one without her. That's irreducibly social, and it doesn't depend on a capability gap.

What to do about it

1. Stop hunting for a model moat. If your edge is the model, you don't have one. Audit which of your advantages a competitor could rent tomorrow.
2. Build the Guardian and Force Multiplier as first-class functions. Taste is only a moat if you've institutionalized the people who hold it — not bolted on as an afterthought.
3. Write your judgment down. Every resolved edge case, refined charter, and logged decision is a deposit in an account competitors can't access. Treat the audit spine as a strategic asset, not compliance overhead.
4. Protect the conditions for transmission. Taste convergence is real — humans working with AI internalize its patterns within weeks, rejection rates falling as their distinctiveness erodes. Rotate people through non-AI environments; preserve the eye the moat depends on.

The principle

When everyone can rent the same engine, the race stops being about horsepower and becomes about steering — about the accumulated, recorded, transmittable judgment of which direction is worth going. Midjourney's competitors have the architecture. What they don't have is the years of choices about what to emphasize and suppress, written into a product people recognize on sight. That is the moat now. It's invisible on a balance sheet, and it's the hardest thing in the world to copy.