How to start a dark kitchen from scratch: the mistakes that burn CapEx and the method that defends margin

Chapter 1 — How much margin does a dark kitchen order actually leave?

A USD 14 digital ticket leaves roughly USD 4.60 in contribution margin, and that figure —not the lease— governs the entire build. The arithmetic is dry:

a 27% aggregator commission takes USD 3.78, packaging eats USD 1.10, food cost at 32% adds USD 4.48, and whatever survives is the only money paying for kitchen labor, energy and debt. With Uber Eats at 26.1% of the US market and Grubhub at 6.3% as 2024 closed, per Earnest Analytics (2024), nobody negotiates rates on goodwill. Here sits the tension almost no operator resolves: the channel bringing the volume is the same one cutting unit revenue, and the bridge is not fighting the commission, it is raising the ticket and shrinking packaging until margin per order can carry the 180 daily orders your rent demands. Below USD 500 thousand a year, the line that sinks the P&L is not payroll: it is packaging and transit waste.

Chapter 2 — Under 500 thousand in annual revenue: the band where packaging decides

We are describing an operation of 90 to 130 orders per day on a single heavy shift, where USD 0.40 of difference in a cardboard container is worth between USD 13,000 and USD 19,000 annually, more than a second line cook costs. Independents held 61.7% of cloud kitchen revenue in 2025 (Grand View Research, Cloud Kitchen Market), so this band is the sector norm rather than the exception. My recommendation at this scale stands untouched: lease shared square meters, hold the menu to eight dishes coming off one mise en place, and hire no in-house marketing until contribution margin per order clears USD 5. Between USD 500 thousand and USD 1 million in revenue a new cost appears that the previous band never paid: buying visibility inside the aggregator, which typically runs another 4% to 7% on channel sales. Stacked on the base commission, that layer pushes total extraction past 30 points of the ticket, and the road forks there.

Chapter 3 — From 500 thousand to 1 million: when acquisition cost enters the plate cost

Mexico already runs more than 1,200 active dark kitchens, up 40% from 2023 according to CANIRAC (2025), and that density raises the price of every click in saturated delivery zones. What I advise in this stretch is hard to swallow: open a second brand inside the same kitchen before you open a second kitchen, because the new brand buys you a second storefront at zero marginal rent and barely 6% to 9% more operating cost. Past the million-dollar mark, your largest exposure stopped being the kitchen and became platform concentration. iFood handles 87% of e-food bookings in Brazil (Statista, 2024) and reports more than 380,000 partner establishments across over 1,500 cities (iFood, 2024), with a single month of 100 million orders in August 2024; in Spain, Glovo sits near 31% and Just Eat near 26% (Ken Research, 2025). Consider the counterfactual: if the aggregator supplying 70% of your sales lifts commission by two points and you bill USD 1.4 million, you lose USD 19,600 of EBITDA in one quarter without a single change inside your operation.

Chapter 4 — Above 1 million: territory risk outweighs food cost

That is why at this scale I demand an owned channel worth at least 18% of sales, even though its ticket converts worse for the first six months. Above USD 5 million you enter branded operations with a face —the television chef, the licensed themed format— and three costs arrive that smaller bands never see: brand royalty between 3% and 6% of gross sales, presentation standards that double packaging cost to USD 2.20 per order, and a quality control team nobody budgets until the first viral review lands. North America holds 40.8% of the kitchen robotics market (Grand View Research, Food Robotics Market), and in this band automation genuinely pays: at 600 daily orders, a frying arm amortizes in eighteen to twenty-four months. What does not pay is an inflated ticket without backup kitchen capacity; the famous brand multiplies launch demand and then punishes dispatch time by the third weekend.

Chapter 5 — Above 10 million: a network of nodes, not scattered kitchens

Past USD 10 million you no longer manage kitchens, you manage coverage density, and the marginal cost of node number twelve decides your consolidated EBITDA. The reference model lives in quick commerce: Blinkit operates close to 2,100 dark stores and plans 900 more by March 2027 according to Storyboard18 (2025), a logic of small nearby nodes that transfers cleanly to prepared food. In this band logistics cost per order falls from 11% to 7% of sales once the average delivery radius drops from 4.5 to 2.8 kilometers, and those four points on USD 12 million equal USD 480,000 a year. Grab sustains 53.9% of Southeast Asian food delivery (Momentum Works, 2024), evidence that scale is won through coverage rather than menu. Diego F. Parra and Masterestaurant sequence a dark kitchen build the opposite way the market does: plate arithmetic first, square meters afterward.

Chapter 6 — The correct order of the build decisions

An operator who starts from the property ends up justifying a USD 2,800 monthly lease with an order forecast that never materialized; one who starts from the number knows that, at USD 4.60 of margin per ticket, that lease requires 61 orders daily just to cover rent, before payroll. Latin America's meal delivery segment will pass USD 39 billion in 2027 according to Statista (2024), and that growth is precisely what deceives: demand climbs while commission stays put. For years I treated commission as a marketing expense, and I was wrong: it subtracts from unit revenue and belongs in the plate cost ahead of every other line. The first automation that pays inside a dark kitchen is not the cooking, it is order capture and routing. White Castle rolled voice AI into more than 100 drive-thrus by late 2024 and reports a 90% order completion rate at roughly 60 seconds per transaction (SoundHound via Restaurant Dive, 2024); Wendy's took FreshAI to between 500 and 600 US locations by the end of 2025 per CNBC (2024).

Chapter 7 — Channel technology: what to automate, and in what order

Translated into your P&L: every minute shaved off ticket time frees capacity without one extra square meter of kitchen, and during peak hour that buys 12 to 18 additional orders. Autonomous delivery is not your lever yet —Starship completed 5.8 million deliveries in 2024 (Forbes, 2025) and still covers narrow geography— so spend where the order comes in, not where it goes out. The order of decisions. The improvised route starts with the property and ends with the number; the method starts with the number. Once you work out that a USD 14 ticket carrying 27% commission, USD 1,10 of packaging and 32% food cost leaves roughly USD 4,60 of contribution margin, the lease stops being an aesthetic choice and becomes a hard ceiling on orders per day. How commission is treated. Aggregator commission is not marketing spend; it is a direct reduction of revenue per unit sold, which is why it enters plate costing ahead of every other line.

Chapter 8 — Six differences that decide margin

With Uber Eats at 26,1% of the US market and Grubhub at 6,3% (Earnest Analytics, 2024), the fee conversation goes to whoever arrives with measured volume and live alternatives. Menu complexity. Every extra reference multiplies inventory, waste and error probability, and in delivery an error cannot be repaired with floor hospitality because there is no floor. Eight SKUs sharing inputs hold ticket time at 11-14 minutes; forty references push it past 22, platform rating drops, visibility follows and sales go with it. The dependency structure. An operator with 90% of revenue on one platform carries territory risk equivalent to a restaurant with a single meat supplier. In Brazil, where iFood moved 100 million orders in a single month (iFood via Statista, 2024), that dependency is paid in fees; the method builds the owned channel in week four, not during the crisis. The role of automation.

Chapter 9 — Six differences that decide margin — in practice

Here is where the Masterestaurant foodtech angle contributes: AI does not replace cooks in a 30-square-metre kitchen, it replaces the administrative work that consumes the owner's afternoons —consolidating orders from three tablets, reconciling inventory, catching food cost variance before the month closes—. The maturity benchmark exists: White Castle passed 100 AI-voice drive-thrus by late 2024 (Restaurant Dive, 2024) and Wendy's projected FreshAI across 500-600 US locations by the end of 2025 (CNBC, 2024). Expansion discipline. A dark kitchen that opens brand two before stabilising brand one does not diversify, it fragments. The criterion is arithmetic and allows no enthusiasm: three consecutive months with net-of-commission contribution margin above 18% and errors under 2%, or brand two waits.