Hundreds of billions of dollars are flowing into artificial intelligence. Far less clear is who is actually making money. Strip away the excitement and the AI economy resembles a layered stack — and today, profits are concentrated heavily at the bottom of it, in the least glamorous layers.

Layer one: the picks and shovels

The oldest observation about gold rushes applies here. The most reliable profits so far have gone to the companies selling the equipment.

Chipmakers sit at the top of this layer. NVIDIA built a dominant position in AI accelerators, and competitors are now investing heavily to challenge it — as seen in the AMD–Anthropic partnership covering up to two gigawatts of hardware. Beneath the chip designers sit memory manufacturers, networking suppliers, and the specialist firms that actually fabricate the silicon. All of them are selling into demand that currently exceeds supply.

Layer two: infrastructure and power

Next come the companies that house and feed the machines. Cloud providers are expanding aggressively, and the numbers are striking — Alphabet reported that Google Cloud grew 82% in a single quarter, and Meta is reported to be building a business to sell AI compute, as we covered in our look at the compute land grab.

Behind them sit the least visible winners: data-centre builders, cooling specialists, and energy suppliers. When AI infrastructure deals are measured in gigawatts, electricity providers become strategic partners rather than utilities — the dynamic explained in our piece on why AI needs so much power.

Layer three: the model builders

This is where the story gets more complicated. The frontier AI labs command enormous attention and enormous valuations — but they are also spending extraordinary sums. OpenAI has been reported to be planning compute expenditure through 2030 running to several hundred billion dollars.

Two forces squeeze this layer. First, training and serving models is genuinely expensive. Second, competition is driving prices down: as we documented in the 2026 AI price war, per-token costs have fallen sharply while capable open-weight models set a low price floor. Revenue is growing quickly, but so is the cost of staying at the frontier.

Layer four: the application builders

Companies building products on top of models face the opposite situation: low capital requirements, but thin defensibility. If your product is a modest layer over an API that competitors can also call, your advantage has to come from somewhere else — proprietary data, workflow integration, distribution, or brand.

The good news for this layer is that falling model prices directly improve margins. What cost a fortune to run last year may be routine this year, which quietly makes many applications viable that previously were not.

The layer nobody counts: the buyers

The largest economic effect may not appear as anyone's AI revenue at all. When a company uses AI to reduce support costs or accelerate engineering work, the benefit shows up as improved margins in an ordinary business, not as a line item in the AI industry. That value is real but diffuse — and probably underestimated in most analyses of who is winning.

The honest uncertainty

A necessary caveat: much of the current spending is a bet on future demand. Gigawatt commitments and multi-year data-centre projects assume growth that has not yet happened. If AI adoption continues climbing, today's infrastructure build-out looks prescient. If it plateaus, some of these commitments will look expensive. Anyone claiming certainty about which outcome arrives is guessing.

Why it matters

Understanding the stack explains industry behaviour that otherwise looks strange. It clarifies why chip supply drives so much strategy, why model prices keep falling even as capability rises, and why application companies are racing to find defensible niches. For anyone building a business or a career in this field, knowing which layer you occupy is the first step to understanding your economics.

Key takeaways

  • AI profits currently concentrate in hardware and infrastructure — the "picks and shovels" layers.
  • Cloud and power providers are major beneficiaries as deals move to gigawatt scale.
  • Frontier model builders earn large revenue but face enormous costs and falling prices.
  • Application builders enjoy low costs but must find defensibility beyond the model itself.
  • Much of the real value accrues invisibly to companies that simply use AI to work more efficiently.

The bottom line

Today, the surest money in AI is being made by companies selling the equipment rather than the intelligence. Whether that stays true depends on a question no one can yet answer: how much demand ultimately arrives to fill all the capacity now being built.