Who Eats? Profit and Power, From the Cloud Kitchen to the AI Stack

Picture a plate of biryani at someone’s door. Four parties touched it: a supplier sold the rice and spices, a cook made the food, a cloud-kitchen operator owned the brand and the building, and an app put it in front of a hungry stranger and took a cut. Everyone insists they’re essential. Only some get rich.

The question isn’t who works hardest. It’s who captures the value, and why. Answer it for food and you have the lens for AI, where the same four roles exist: infrastructure selling raw compute, the model lab doing the cooking, the platform, and the apps serving the end user. The two chains look identical. One of them is upside down. That’s the story.

The food chain

Start at the bottom. The supplier hands over raw items and earns the least, always has: commodity inputs are interchangeable, so the buyer switches sources tomorrow. Even at scale, food distribution runs low-single-digit margins (Eternal’s own supply arm, Hyperpure, sits thin to negative). The cook is labour: skilled, essential, paid a wage, and capturing none of the enterprise’s value unless the cook becomes a brand.

The cloud-kitchen operator looks clever (no dining room, ten virtual brands from one cheap kitchen) and still bleeds. Rebel Foods lost roughly ₹337 crore in FY25 on ₹1,617 crore of revenue, with EBITDA margins near minus 8 to minus 10 percent. It is squeezed from both sides: rising material costs below, platforms skimming 15 to 30 percent commission above.

The aggregator, Zomato (Eternal) and Swiggy, is the structural winner, but the win is thin and took a decade: blended net margins of 2 to 3 percent, food-delivery EBITDA around 5 to 6 percent, Swiggy still loss-making, and five rivals (Foodpanda, Uber Eats, Amazon Food, TinyOwl) already dead. It wins for one reason: it owns the customer, and everyone else pays to reach through it. The chain’s shape is textbook: value concentrates at the demand layer.

The principle

Three old ideas explain why, and predict what AI does next.

Christensen’s Law of Conservation of Attractive Profits: when a layer commoditizes, profit migrates to the adjacent layer that is still scarce.

Aggregation Theory: value accrues to whoever owns the end customer and can commoditize suppliers.

And the instinct to commoditize your complement: every layer wants the one below it cheap and abundant. Hold those three. In AI, for now, all three appear to break.

The AI chain, inverted

Mapping it: infrastructure (Nvidia’s chips, the clouds, the power) is the supplier; the model labs (OpenAI, Anthropic, Google) are the cooks; the apps are the kitchens and brands. In food the supplier loses. In AI the supplier is king. Nvidia’s FY2026 ran roughly 75 percent gross and 55 percent net margins on about $216B of revenue and $116–120B of net income. A chipmaker at 75 percent gross is a level almost nothing in hardware history has touched.

The cooks pay the supplier and bleed. OpenAI runs about a 33 percent gross margin and a projected ~$14B loss in 2026, turning cash-flow positive only around 2030, with its compute spend flowing straight back to Nvidia and the clouds. Anthropic is healthier (similar revenue, a quarter of the training spend, break-even near 2028), but the rule holds: the ingredient supplier keeps the profit. The app layer rhymes with the cloud kitchen, buying tokens the way Rebel Foods buys commissioned reach. A thin wrapper with no data, no lock-in, and no brand is a ghost kitchen.

Why is it inverted? Because in AI the raw ingredient isn’t yet a commodity. Compute is scarce, proprietary, and supply-constrained; Nvidia sells as fast as it builds, CUDA is a moat, and there is no second source. By Christensen’s law, profit pools at the bottleneck. It is the oldest pattern in capitalism: in the 1849 gold rush the miners went broke and the shovel-sellers got rich. AI is in its shovel-selling phase. But bottlenecks move, and this one is already loosening.

Where this is heading

Compute commoditizes slowly, and the next scarcity is power. AMD and the hyperscalers’ custom silicon chip away at Nvidia’s rent; the binding constraint is shifting from chips to electricity and data-center capacity. The next rent may belong to whoever controls power, not silicon.

Models commoditize fast, and the cook layer hollows out. Open-weight models now match the frontier for the cost of running them, triggering token price wars. For most tasks one model is interchangeable with another, a utility. The premium survives only on what can’t be copied: proprietary data, post-training, and distribution.

Value climbs to whoever owns demand. As the layers below commoditize, the durable margin moves up to whoever owns the customer and a hard workflow. The wrapper dies; the workflow app inherits the margin. The chart below is the prediction: the profit pool slides from infrastructure toward demand as the stack matures.

The layers blur. AI permits vertical integration that food never did. The lab becomes the front door (OpenAI is both the model and ChatGPT’s ~800M weekly users); the chipmaker funds the labs (Nvidia’s ~$100B into OpenAI, money that flows back into chips). Whoever owns a scarce layer and the customer escapes the squeeze.

Expect an overbuild, then a feast. Every boom overbuilds (1840s railways, 2000 fiber). Over $1 trillion in AI infrastructure is committed, much of it circulating between the same few companies. If revenue lags those commitments, infrastructure corrects first. History is oddly reassuring: after the fiber bust, the spoils went not to the firms that laid the cable but to the internet companies that feasted on the cheap, stranded capacity. If AI overbuilds, the long-run winners are the demand-side players who get abundant compute at a discount.

The bottom line

One rule governs both chains: profit goes to the layer that is scarce and owns the customer. Today in AI those are two different layers (compute is scarce; ChatGPT owns demand), which is why it looks chaotic; they converge as the market matures. If you build above the infrastructure line, the takeaway is exact: don’t be the cook reheating someone else’s ingredients. Own a bottleneck, or own the customer. In a gold rush you can rent a shovel and do fine. When it ends, only the people who own land keep eating.