Building a Startup with Large Models Is Like Ordering a Coke at McDonald's: The Crispness Is the Ice, Not Your Recipe

Greg Isenberg's tweet compares building an AI startup with OpenAI, Anthropic, Google, or Grok models to ordering a crisp, refreshing Diet Coke at McDonald's. The video shows only a hand and a McDonald's cup against a light gray background—the point is clear enough.

The interesting part of this analogy isn't the Coke—it's the “crispness.” McDonald's Coke tastes good because the ice, pressure, syrup ratio, and straw diameter have all been worked out. But the cup has an M on it, and the recipe isn't yours to control. One commenter piled on: it's still Diet Coke, the cup says McDonald's, and you can't choose the syrup ratio. Another asked: if the model is the syrup and your product is the fountain, is that enough of a moat?

Greg's response was blunt: “No moat.” He then added that distribution, network effects, and data will still be moats. That answer is more worth remembering than the analogy. The more standardized the model layer becomes, the harder it is for the application layer to build a barrier by saying “we integrated the strongest model.” What users taste is the ice and pressure; what they remember is the cup. If a product doesn't have its own use case, workflow, data feedback loop, and distribution, then you can swap out the syrup and the cup is still there.

One commenter said AI is an amplifier: it amplifies capabilities, and it amplifies flaws. That holds here. A model can amplify a clear product process, but it can also amplify the fact that you're just “calling an API.” Another commenter put it even more sharply: renting Coke is easy; owning the shelf is the business. For people building AI products, that sentence is closer to reality than “the model is the capability.” The model is the ingredient, the fountain is the product, and the shelf is the channel. Ingredients get cheaper, fountains become commoditized, and shelves have to be occupied inch by inch on your own.

Of course, this Coke really is crisp. Greg himself stressed “extra crispy” in a reply. Don't miss that point: standardized ingredients let products get off the ground faster and raise the baseline experience. Ordinary users don't need to understand model parameters to feel that something is smooth. Developers can validate use cases faster. The question is where to go after validation. Stopping at “we used a certain company's model” is like treating the McDonald's cup as your own brand.

Finally, here's an observation from one reply: Is everyone drinking the same Coke? If the answer is yes, then the business isn't in the Coke. It's in who owns the fridge, who decides the shelf, and who controls which app users open first the next time they want an ice-cold Coke. This analogy doesn't solve the startup problem, but it puts the problem in the right place.

Original tweet: https://x.com/gregisenberg/status/2115608906760135136