The moat is in the data
In a world where every AI trains on the same internet, the real advantage is data, taste, and knowledge that was never online to begin with.
Three stories this week. DoorDash paying couriers to film chores, tech workers competing on who burns the most AI tokens, and your sommelier losing to ChatGPT. The common thread: in a world where every AI trains on the same internet, the real advantage is data, taste, and knowledge that was never online to begin with.
***
Salt Bae does not have a Michelin star.
Worth remembering when Nvidia’s Jensen Huang tells the world that any employee earning $250,000 who is not burning $500,000 in AI tokens is probably bad at their job. The New York Times has a name for it now: “tokenmaxxing”, a status game among tech workers competing to consume as much AI as possible. To do what, exactly, nobody seems to know.
I loathe all the tech bros who keep saying how taste is important but then go on comparing their token usage like rednecks with the size of their trucks’ wheels. The hospitality allegory would be a chef whose signature dish is lobster stuffed with duck liver, topped with caviar and a splash of champagne. Is the dish expensive to produce? Yes. Is it good? No. Token usage has nothing to do with outcomes, the same way a steak covered in gold leaf has nothing to do with taste.
For hospitality operators already nervous about AI, this is dangerous framing. When the benchmark is twice an employee’s salary in tokens, it stops being advice. It is just tech people talking to other tech people.
If your pain point is marketing or staffing, AI can already help, and it does not require burning through your margins in tokens. Food cost analysis, shift scheduling, unglamorous work, but solvable with a well-written prompt and a basic automation script.
***
A few years ago, Google Maps did to taxi drivers what ChatGPT is now doing to your sommelier. The moment a passenger could track the route in real time, the local expert became a variable to monitor. Everything the driver knew, the shortcuts, the back streets, suddenly made irrelevant because a device knew better.
Now it is happening at the table.
A guest does not recognize a name on your wine list and reaches for their phone, fair enough, but here is what they miss.
Your sommelier knows your menu, they know which wine works with your fish and which one the chef hates. When a guest scans the list and asks ChatGPT, they are asking from an algorithm who has never set foot in your restaurant.
This is why I am so against kiosk ordering and QR code menus. Hospitality is first and foremost a human business, it is literally in the name (from hospes “host, guest”). The goal of AI should be to remove screens, not add them. That sommelier might not give the best recommendation, but if they are good at their job, they will give you the right one.
***
DoorDash is paying people to film themselves loading a washing machine.
Genius move. What they are buying is the data inside the video, physical-world data that has never been digitized. Every foundation model trains on the same internet. The difference is what you can feed it that did not come from the internet.
DoorDash figured that out. Their delivery network is not just moving things anymore, it is capturing what the physical world looks like, and that is the kind of data you cannot scrape from a website.
Now think about your restaurant.
There is a table in every restaurant that is considered the best, but it changes depending on the moment. The one that catches the light at 4 PM, the one near the wall where you can fit a high chair. It is never the same table for every guest, and listing all the edge cases and combinations manually would take forever. A short video or a few photos train the tool faster than any spreadsheet.
To some extent, Eli Feldman, the owner of Shy Bird in Boston, mentioned something similar in his newsletter this week. Toast IQ is a good example of the limitation. Useful, but it only works with Toast data. Any embedded AI is only as smart as the data it sits on. If you want AI making better decisions for your restaurant, you need a central repository first, and most restaurants do not have one yet.
Everyone’s talking about prompts and tools. You’re pointing at the part that actually compounds: what you know about your own operation that never made it online in the first place. The sommelier example is the one operators should pay attention to. It’s not that AI replaces them. It’s that guests now have a second opinion in their pocket. The job shifts from “knowing more than the guest” to “knowing the context better than the model.” Menu, timing, mood, table, history. That’s the layer AI doesn’t see. Same with your DoorDash point. Most operators are sitting on valuable data and don’t treat it like it matters. Where people sit. What they ask for. What slows service down. What gets sent back. It lives in staff heads or disappears at the end of a shift. The gap isn’t access to AI. It’s that most places haven’t captured their own reality in a way AI can use. That’s the moat. Not more tokens. Not louder tools. Better inputs grounded in how the place actually runs.