I was at Terminal 1 of Bengaluru airport this week, watching a crowd swarm The Rameshwaram Cafe. Three deep at the counter, waiting eagerly for idli, dosa, and filter coffee. The same items sold at a thousand Darshinis across the city. Food made in every South Indian home. The barrier to cooking one good idli is close to zero.
And yet Rameshwaram prints money.
A few kilometres away, Burma Burma did something stranger. In a market drowning in pan-Asian meat and standard Indian fare, they opened a premium, exclusively vegetarian Burmese restaurant, and turned it into a profitable, fast-expanding business in one of the most brutal industries on earth.
It is tempting to turn this into a neat parable for the AI gold rush: the model is commoditised, just like idli batter, so win on packaging and positioning. That parable is half right. The half that is wrong will quietly kill your company, because it rests on the one thing a restaurant has that your AI product never will.
You cannot download a dosa. You can download your entire product before lunch.
So instead of admiring the analogy, let me push on it until it tells the truth. Here is where the restaurant lesson holds, where it shatters, and what an AI founder should actually do about it.
The Moat You Can Smell
Rameshwaram's real moat is not branding. It is physical.
The customer travels to the cafe. The airport lease is scarce and contracted. The density of outlets compounds familiarity. The queue itself is social proof you can see and smell. None of this can be cloned over a weekend by someone with a clever idea, because the clone would need the location, the supply chain, and the years of repetition that built the operation.
Your AI application has none of these. It is software: infinitely copyable at zero marginal cost, globally visible the day it ships, readable from your own landing page. The most powerful moat a great restaurant has, physical non-replicability, is exactly the moat you cannot inherit.
This is not a reason for despair. It is a filter. Every time you admire a restaurant's defensibility, ask the harder question: what is the software equivalent, and does it survive the fact that my product can be copied instantly and my own supplier can become my competitor? Most of the time, that question rewrites the lesson completely.
Let me show you what survives.
The Real Lesson of the Idli
Everyone reads Rameshwaram as "commodity product, brilliant packaging." That reading is wrong, and it sends founders sprinting toward the surface while the real moat goes unbuilt.
Making one good idli is trivial. Serving thousands of consistently excellent idlis a day, at airport speed, across multiple outlets, with zero quality drift, is extraordinarily hard. That gap, between "can make one" and "can reliably make ten thousand," is the entire moat. It is an operational system that took years to build and cannot be reverse-engineered from a photo.
The translation to AI is exact, and almost nobody wants to hear it. Anyone can demo an agent. That is why your feed is buried in them. Almost no one can run an agent that performs reliably across thousands of messy, edge-case-ridden production interactions, every day, without drifting into nonsense. That gap is your moat: the evaluation harness, the error handling, the unglamorous reliability engineering that turns a viral demo into a system a business will actually bet its operations on.
Build the machine behind the counter. The demo is the dosa held up for the photo. The machine is why people come back.
Rank Your Moats Honestly
Founders collect moats like badges: brand, trust, data, switching costs. Elite operators rank them, because in AI they have wildly different shelf lives. Here they are, weakest to strongest.
Trust is table stakes, not a moat. It stops customers leaving. It rarely wins a deal against a rival who also has it. The day your competitor ships comparable guardrails and source citations, trust parity arrives and the advantage evaporates. And read the fine print on "if it hallucinates, the customer churns": that is a warning about entering high-stakes verticals like law and medicine, not an invitation. The stakes that make a vertical lucrative are the same stakes that make one failure fatal.
Proprietary data is a moat only under strict conditions. "We fine-tuned on 10,000 contracts" sounds defensible and often is not. Base models keep getting better at zero-shot reasoning, and a competitor on a newer model with good retrieval frequently matches your fine-tuned older one. Most "proprietary" data is not exclusive at all. Data becomes a real moat only when it is genuinely exclusive, continuously refreshed so no snapshot can catch up, and wired into a loop where usage generates more data that makes the product better that drives more usage. No loop, no moat. Just a head start.
Switching costs are strong, but they cost you something back. Embed your agent into a company's daily workflow inside the tools they already live in, and ripping it out becomes painful no matter how cheap the alternative. This is the best of the common moats. But embedding has a price nobody mentions: the platform you embed in is the one most able and most motivated to replace you. Live inside Salesforce and you live alongside Agentforce. Live inside Zendesk and you live alongside its native AI. Embedding is a bet that you can entrench faster than the platform can absorb you. Make it with open eyes.
Distribution is the most underrated moat. Rameshwaram's airport location is distribution, and it may explain more of its success than the food does. When everyone can build a comparable product, the founder with the cheapest, most durable path to customers wins. A community you own. A content engine that compounds. A partnership that puts you in front of buyers your rivals cannot reach. In commoditised software, distribution is often the only moat that holds, and it is the one founders neglect because building is more fun than selling.
When Your Supplier Becomes Your Competitor
Here is the part that breaks the metaphor entirely, and it is the most important paragraph I can give you.
Rameshwaram buys rice and lentils from suppliers who will never open a competing cafe next door using Rameshwaram's recipe. The supply chain is safe.
Yours is not. OpenAI, Anthropic, and Google are not passive batter vendors. They are climbing the stack into applications, agents, and entire verticals. The deadliest threat to a "GPT-wrapper" was never another wrapper built over the weekend. It is the model provider deciding to build your product directly, or making it a free default feature shipped to every developer, or releasing one capability that erases the exact gap your company exists to fill.
So the real strategic question is not "how do I avoid being cloned by a startup." It is "what am I building that survives my own supplier choosing to compete with me." The answer is almost never the model interaction itself, because that is the layer the provider controls. It lives in the layers they will never want to own: the proprietary data loop, the deep workflow integration, the regulatory specificity of a vertical, the distribution, the operational reliability tuned to one industry's ugliest edge cases.
Build on the ground your supplier will never want to take. Everything else is rented land owned by your most capable competitor.
Burma Burma's Real Secret
Burma Burma is held up as proof that focus wins. It is. But focus alone is a trap, and the trap has three jaws.
Narrow your niche and you cut direct competition. Good. You also cap your market: "AI for dental clinic patient reactivation" might have zero rivals and a ceiling too low to build anything that matters. And the third jaw is the one that draws blood: the narrower and more software-specific your niche, the easier it is for the incumbent who already owns the customer to bolt your feature onto their product and reach your entire market overnight, with distribution you will never match.
Burma Burma is safe because authentic Burmese cuisine is genuinely hard to fake, and the pan-Asian restaurant next door cannot credibly bolt a Burmese menu onto its kitchen. That is the test. Pick a niche narrow enough to win, large enough to matter, and shaped so the obvious incumbent either cannot or will not absorb you. If they can add your feature in a single sprint, you have not found a moat. You have found their next release note.
The Economics Will Ambush You
One last thing, because it is where good founders get blindsided.
A restaurant is capital-heavy, low-margin, and hard to scale, but a dosa costs roughly the same to make every single time. Your AI business is the mirror image: light on capital, beautiful gross margins in theory, trivial to scale, and carrying inference as a real cost that grows with every query. Price on a flat subscription while your power users run expensive queries all day, and you will discover your best customers are your least profitable, and that growth makes your margins worse, not better. There is no dosa equivalent of this. Model your unit economics per query, not per seat, or the cash cow you think you are building will quietly bleed out.
The Takeaway
A commoditised core, whether fermented rice batter or a foundation model, is not a death sentence for margins. It is the starting line. But the lessons people draw from Rameshwaram and Burma Burma are usually the wrong ones.
The moat at Rameshwaram is not packaging. It is an operational machine no one else can build. The moat at Burma Burma is not focus. It is a niche no incumbent can swallow. And neither restaurant has to survive the founder's nightmare you live with every day: a supplier who sells you the core ingredient this morning and may sell your customers the finished dish tonight.
So if you want to build the Rameshwaram Cafe of AI, do not start with the brand. Rank your moats honestly. Treat trust as the price of admission. Build a data loop, not a dataset. Embed for switching costs with your eyes open about platform dependency. Invest in distribution as if it were the product, because in a crowded market it usually is. And plant yourself in the layers your model provider will never want to own.
The next great AI company may well be doing the tech equivalent of selling a basic dosa. It will win the way Rameshwaram actually wins: not because the dosa is special, but because the machine behind the counter is one no one else can build, in a spot no one else can take.
You cannot download that. And that is the whole point.