The Real Moat in AI Is Not the Model. It Is Taste.

Taste is the moat. Not the model. Not the data. Not the GPU cluster. After three years of watching AI companies spend billions to be marginally better than the next lab, I am done pretending the frontier matters for anyone except the labs. The companies that will win the next decade are the ones with the better eye.

Models Are Commoditized

GPT-5.6, Claude Opus 4.8, Gemini 3 Pro, DeepSeek V4 Pro, GLM 5.2 — pick any two and the benchmark delta is in the noise. The leaderboard shuffles every six weeks. The API surface is near-identical. I have helped four companies swap their primary model in the last quarter. None noticed in the product, and two saw cost improvements. The model is the database. You pick one. You stop thinking about it. The moat is not the model.

Data Is a Temporary Moat

Your proprietary data is meaningful for about 18 months. After that the frontier labs have caught up — scraping the public parts, licensing the private parts, or training on synthetic data distilled from models that already saw your data. Distilabel, Glaive, the in-house generators at every major lab, the agentic-browser startups mining the web for click-through sequences — the shape of your data is in the next pretraining run. The actual rows are not. The judgment about which rows to keep, drop, weight, or ignore — that is taste.

Compute Is a Moat for Labs and a Bill for Everyone Else

The GPU cluster is a moat if you are selling tokens. If you are building a product on top of tokens, it is an operating expense that gets cheaper every quarter. CoreWeave, Lambda, the hyperscalers, and the neoclouds have made capacity abundant. HBM3E is shipping. Inference cost per million tokens fell another 30% in Q2. The labs are racing to spend capital their shareholders will eventually want back. That is not a moat. That is a treadmill.

Taste Is What Separates the Winners

Look at the products winning in 2026. Linear, Vercel, Stripe, Figma, Raycast, Granola, Perplexity, Apple Intelligence. They are not winning because their model is better. They are winning because a human — usually a small group — made a thousand decisions about what to keep, cut, refuse, polish, and ship ugly. The taste is in the empty space on the screen. The taste is in the words the assistant does not say. The taste is the difference between a search result that answers the question in two seconds and a search result that answers a related question in two paragraphs.

Taste is the thing that cannot be acquired by training another epoch. It cannot be copied from a public benchmark. It cannot be acquired by raising a $200M Series C. It can only be acquired by hiring people who have it, protecting them from the meeting culture that punishes it, and shipping the thing they want to ship.

The Counterargument I Respect

The honest reply: distribution is the moat, not taste. Google ships Gemini to two billion users. Apple ships Apple Intelligence to every iPhone. OpenAI ships ChatGPT to 800 million weekly actives. They do not need taste — they have a captive audience.

Agreed on the numbers. The take still holds. Distribution is a moat for the incumbents. Taste is the moat for everyone who has to earn their users. Linear did not beat Jira with distribution. Stripe did not beat PayPal with distribution. Perplexity did not beat Google with distribution. The take is not taste beats distribution. The take is taste is the moat for companies that do not already own an operating system, a search engine, or a chat client.

The Take

Stop chasing frontier model deals in 2026. Stop building strategy around which lab has the highest benchmark this quarter. Hire the people with the eye. Pay them more than you think is reasonable. Protect their time. Ship the thing they want to ship. In five years the only thing that mattered is whether you built something with a point of view.

The frontier models are a commodity. The frontier products are not.

— Mr. Technology