Ghost kitchen: the mistakes that burn cash vs the right method

Chapter 1 — Why does a ghost kitchen with full orders still go broke?

A ghost kitchen doesn't go broke for lack of orders: it goes broke because nobody corrected the price for the aggregator's commission.

On DoorDash's restaurant plans that commission reaches 15%, 25% or 30%, per CloudKitchens (2024). Demand isn't the problem, there's plenty of it: 65% of limited-service operators already deliver, per the National Restaurant Association (2025), and by 2023 close to 15% of U.S. foodservice delivery sales were already billed by a brand that doesn't exist outside an app, per Statista. The problem was never selling. It was the arithmetic of copying the dine-in price and letting a third party keep a third of the ticket without that cut ever entering the formula. I see it constantly in the Masterestaurant framework: businesses that bill well and still decapitalize, because the owner only spots the leak when the bank balance doesn't add up at month's end and two months of rent are already owed.

Chapter 2 — The cost of the error: replicating physical costing in the app

If a dish with 30% food cost is listed on Rappi, iFood or DiDi Food at the dine-in price, and the aggregator charges 25% commission, the net contribution margin falls below 20% before payroll or rent, which are paid separately, even factor in. Replicating physical costing in the app decapitalizes the business order by order, and the channel's scale multiplies the mistake: DiDi Food delivered more than 360 million orders in Mexico over five years, per DiDi Food's own figures (2024), and DoorDash closed 685 million orders in the fourth quarter of 2024 alone, 19% more than a year earlier, according to its financial results. Every one of those mispriced orders repeats the same loss thousands of times over. In the Masterestaurant method we recalculate the app's list price to absorb the commission without cannibalizing the margin, dish by dish, before the menu goes live.

Chapter 3 — Profit on net-of-commission margin, not on gross sales

On a USD 20 sale with 30% food cost (USD 6) and 25% aggregator commission (USD 5), USD 9 of contribution remains: a 45% that looks healthy until packaging, waste and the cost of an in-house rider get loaded on. That's the structural difference between the two models: one looks at total sales and calls it a day; the other subtracts the commission before drawing any conclusions, dish by dish and channel by channel. Meal delivery penetration reached 27.5% of users in 2024 and projects 29.2% for 2026 per Statista, so volume keeps climbing and the mistake, left uncorrected, scales right along with it. I insist, inside the Masterestaurant framework, on measuring each SKU's margin after commission, never the average ticket: a dine-in star dish can be exactly the one burning the most cash in the app, and only the per-channel math reveals it.

Chapter 4 — Commission is not a necessary evil: it is a variable to optimize

Treating the aggregator's commission as a necessary evil means resigning yourself to losing 15% of every sale at best, and up to 30% at worst, per the plans CloudKitchens documents (2024). But that commission is just another variable, one you can optimize by migrating volume to direct ordering: close to six in ten customers already order through their own channel before an aggregator, 58% exactly per NCR Voyix (2024). Direct demand isn't marginal. Operators know it well: 63% planned to invest in digital marketing during 2024, according to the National Restaurant Association. In the Masterestaurant framework we frame the aggregator as an acquisition channel, never as the owner of the relationship: the app captures the new customer, and a channel-owned repurchase strategy lowers the weighted-average commission cost. Every migrated point falls straight to EBITDA. Three points of input leakage, whether waste, petty theft or unstandardized portions, eat between a fifth and a quarter of operating profit when the net-of-commission margin runs a mere 12% to 18%.

Chapter 5 — Weekly food cost variance: the deviation that erases EBITDA

Measuring food cost once at launch and filing it away feels sufficient, until it isn't: the recipe card nobody revisits weekly ends up funding the leak. The discipline of comparing theoretical against actual, more than any single recipe, separates the operator building an asset from the one just putting out fires. Not for nothing did 48% of operators prioritize point-of-sale technology in 2024, per the National Restaurant Association: the POS is where the real cost lives. If you don't close that gap every Monday, you don't have a business, you have a leak with a logo on it. Variance gets controlled, or it controls your margin. If the net-of-commission margin survives 5% input inflation but turns negative at 12%, scaling volume fixes nothing: it multiplies the loss on every new order, and that's exactly where the blind model runs into the wall.

Chapter 6 — Stress-test inputs before investing CapEx in scale

The one that works validates before committing capital: it runs the same math with inputs 5%, 12% and 20% more expensive, and only then decides whether to equip another kitchen. The macro backdrop isn't kind either: agrifoodtech investment hit USD 16 billion in 2024, barely 5.5% of global venture capital dollars, per AgFunder (2024), a sign that capital for the sector is tightening and a costing mistake now costs more than ever. In the Masterestaurant framework we require validating break-even under stress before signing off on any investment in ovens, riders or a second location. You scale a margin already proven, never an expectation: scaling before validating is betting the whole cash box. A ghost kitchen becomes a profitable asset the day its pricing, its channel mix and its cost control stop being governed by dine-in instinct and start being governed by numbers.

Chapter 7 — From blind brand to asset: the arithmetic that does build cash

The path that builds cash combines four decisions, not one: cost per channel after commission, migrate volume to the direct ordering 58% of customers prefer per NCR Voyix (2024), measure food cost variance weekly, and stress-test inputs before scaling. The market, on top of that, rewards whoever has that arithmetic solved: 29.2% user penetration in meal delivery is projected for 2026, and 2.6 billion users by 2031, per Statista. This document doesn't replace an audit of your own operation: it's an expert synthesis of verifiable public data, read through my judgment and the Masterestaurant framework, not primary research. Its goal is simple: help an owner tell apart, with cash figures instead of optimism, the path that burns money from the one that builds a business worth owning. The model that fails calculates profitability on gross sales. The one that works calculates it on net-of-commission contribution margin, dish by dish and channel by channel: the only figure that actually matters.

Chapter 8 — What truly separates one model from the other

Direct-order preference is already the majority position: NCR Voyix (2024) measures 58% of customers choosing the restaurant's own app or website over an aggregator. That number changes how you read the commission, 15% to 30% per CloudKitchens (2024): it stops being a necessary evil and becomes the variable you optimize by migrating volume to the direct channel. A sustained three-point drift in food cost erases a ghost kitchen's EBITDA, which is why measuring it once when the business opens is not enough. The method requires comparing theoretical cost against actual cost every week. More orders don't fix an unvalidated margin: that's a bet that yesterday's percentage keeps holding tomorrow. The method runs stress simulations, at 5%, 12% and 20% input inflation, before committing CapEx to a second unit.