Who Actually Makes Money in AI? The Value Stack, Explained

We've followed the money down through the whole machine — chips, training runs, tokens, buildings and power — and watched prices collapse even as spending explodes. One question remains, and it's the one investors, founders, and every curious observer keeps asking: after all this, who actually makes money?

The honest answer has a shape, and the shape is a stack.

The value stack

Every dollar you pay an AI product starts a journey downward through five layers:

• Fabs — the manufacturers, dominated by TSMC, whose leading-edge factories physically make the chips.
• Chip designers — overwhelmingly NVIDIA, which designs the GPUs and owns the CUDA software moat.
• Clouds — AWS, Microsoft Azure, Google Cloud, plus a new breed of GPU-specialist "neoclouds" — which buy chips by the hundred thousand, pour the data centres, and rent capacity out.
• Model labs — OpenAI, Anthropic, Google DeepMind and peers — which spend fortunes training models and sell access by the token and by subscription.
• Apps — everything built on top: coding assistants, writing tools, customer-service bots, and the AI features inside software you already use.

Profit is not distributed evenly across this stack. Not remotely.

The bottom of the stack: selling shovels

The oldest line in business journalism applies perfectly: in a gold rush, sell shovels. The most unambiguous profits in AI sit at the bottom of the stack:

• NVIDIA became one of the most valuable companies on Earth on the strength of AI demand, with reported gross margins on data-centre chips that most industries would consider a typo. Every lab's arms-race training budget is, from NVIDIA's side of the table, revenue. Its position isn't unassailable — every big customer is designing in-house chips to escape it — but the CUDA moat has held far longer than sceptics predicted.
• TSMC manufactures for everyone — NVIDIA and the in-house challengers alike. Whoever wins the chip-design fight, the fab gets paid.
• The clouds occupy the same happy position one layer up: they rent compute to every lab, so they profit from the competition itself rather than betting on a winner. Their AI capex is enormous, but it's backed by the most reliable revenue in the stack — and partly funded by their older, wildly profitable businesses.

The labs: rich, growing, and loss-making on purpose

The frontier labs are the strangest economic objects in the stack. Their revenues are real and growing at historic speed — public reporting puts the top labs in the tens of billions per year, from API tokens and subscriptions, up from single-digit billions barely two years earlier. And yet, by that same reporting, the top labs spend more than they make, with gaps that have themselves reached the tens of billions.

You already know why, because this guide has been building the answer:

• Frontier training is a subscription, not a purchase — stop spending and you fall off the frontier within a year.
• Inference costs scale with success — every new user arrives with a GPU bill attached.
• Competition compresses prices toward cost while open-weight models set a floor under the market.

The losses are a strategy: buy the frontier now, on the belief that leading models compound advantages — the best models attract the most users, whose usage funds the next model — and that the eventual market for machine intelligence is vast enough to repay any plausible buildout. It's the same wager as the data-centre boom, made one layer up. It might be right. It is not yet proven.

The top of the stack: wrappers vs workflows

At the app layer, the economics split into two very different fates:

• Thin wrappers — products that mostly forward your prompt to a model API and mark up the tokens — have a brutal problem: their feature is one model release away from being free. When the underlying model learns to do natively what the wrapper charged for, the wrapper evaporates. Many have.
• Workflow owners — products that hold something the model can't replicate — keep pricing power. That something is usually proprietary data (your company's documents, wired in via RAG), distribution (millions of users already inside the product), or deep workflow integration (the AI is embedded where the work actually happens, with context and tools — the pattern behind agents). For these products, the model is an ingredient, not the recipe — and collapsing token prices are a gift: their costs fall while their prices hold.

Where does the value settle?

Step back and the whole guide compresses into one tension. Making intelligence is getting more expensive; intelligence itself is getting cheaper. If good-enough models become interchangeable commodities — and chapter six's forces push that way — then history's infrastructure booms suggest the ending: railways, telecoms, and the dot-com fibre glut all made society rich while many of the builders went broke, and durable profits went to those who owned the customer relationship on top and the scarce inputs at the bottom.

That's the current shape of AI: comfortable at the bottom (chips, fabs, clouds), contested in the middle (labs betting billions on winner-take-much), and split at the top (wrappers dying, workflow owners thriving). Whether the middle's bet pays out is the biggest open question in the industry — and now you have every piece needed to follow it like an insider rather than a spectator.

Recap

• AI is a five-layer value stack — fabs → chip designers → clouds → model labs → apps — and profit is wildly uneven across it.
• The bottom layers profit most: NVIDIA's chips and TSMC's fabs earn from every competitor, and the clouds get paid whichever lab wins. Neutrality pays.
• The labs have fast-growing billion-scale revenue but spend more — deliberately — because frontier position must be re-bought every generation.
• At the app layer, thin wrappers get absorbed by the next model release; products owning data, distribution, or workflow keep pricing power and benefit from falling token prices.
• If intelligence commoditises, history's playbook says value settles at the scarce bottom and the customer-owning top — the middle is the bet.

That completes the guide — from a single token's price to the industry's balance sheet. To go deeper on what all this money is buying, see How Large Language Models Work for the technology, AI Agents Explained for where token demand is heading, and How Generative AI Actually Works for the no-math foundations.