Follow the Token: Where AI's Profits Actually Go - Part 1

The Paradox Start with a setup that shouldn't be possible. The hyperscalers are guiding to combined capital spending north of $650 billion in 2026, and that spending runs well ahead of what their AI revenue brings back. The frontier model companies burn more on compute than they earn from the tokens they sell. And the enterprises buying those tokens, the ones supposedly capturing all this productivity, mostly can't point to a return yet.

Read literally, that's a chain in which every visible participant loses money. Which raises a question the breathless coverage tends to skip. Money doesn't evaporate. If everyone you can see is underwater, the profit is pooling somewhere you're not looking. So let's follow one dollar of token spent down the stack, layer by layer, and watch where it stops getting recycled and starts getting kept. That last distinction turns out to be the whole story.

Layer One: The Accounting Winners

Drop down one level from the model companies and the picture changes immediately. The supply chain that builds and feeds the data centers is booking real margin, right now, on real shipments. Start with the obvious name. Nvidia booked roughly $216 billion in revenue for its fiscal year ending January 2026, and it has been guiding gross margins around 75%. Those aren't the numbers of a company waiting for a payoff; they're the numbers of a company being paid. But the more durable money sits behind Nvidia, at the choke points that get paid no matter which chip designer wins the cycle. TSMC makes the chips, for Nvidia and for everyone trying to compete with Nvidia, and it controls roughly two thirds of the global foundry market. One step further back sits ASML, whose lithography machines TSMC can't do without, and whose own management has described demand running ahead of supply. These are positions that don't depend on picking the right horse, because they sell to the whole field.

Then the layers most people hadn't priced two years ago. Memory, where the high bandwidth chips that sit inches from every accelerator have pushed Micron and SK Hynix into a supply-and-demand setup that looks nothing like the commodity cycle of old. Networking and custom silicon, where Broadcom, Marvell, and Arista sell the connective tissue of the cluster. And the physical layer, which has quietly become the binding constraint. Power, not chips, is now widely described as the thing the buildout runs short of, which has turned suppliers like Vistra and GE Vernova into AI names. Cooling and power distribution belong here too. Vertiv, which sells the liquid cooling and power gear inside the racks, has guided to a roughly 51% jump in adjusted earnings per share for 2026.

So one answer to where the profits go is already in hand. They go to the supply chain, at high margin, while the headline names burn cash. That's the picks and shovels story, and it isn't wrong. It is just incomplete.

Layer Two: The Catch

Here's where the tidy story gets complicated, and where it gets interesting. A good deal of that supply-chain revenue isn't end-customer money. It's recycled capital, looping through the system in a way that flatters every income statement it touches. The pattern isn't hard to trace. A model company raises billions from a hyperscaler, then spends much of it back on that same hyperscaler's cloud. The clouds and the labs sign large compute deals, then use the proceeds to buy hardware from the chipmaker, who in turn has been investing in the labs that buy its chips. In at least one reported case, the chip supplier is helping finance the very campus its largest customer plans to lease. None of that is fraud. The infrastructure is real, the shipments happen, the electricity gets consumed. But it means booked revenue overstates value captured, because a meaningful share of the dollars are the same dollars going around the loop more than once.

The scale of the gap is the part worth sitting with. By one estimate, roughly $202 billion went into AI infrastructure in 2025, against something closer to $12 billion of direct consumer spending on AI services. That's a ratio of about seventeen to one between what goes in and what comes back out through the front door. Even generous accounting for enterprise software and internal tooling doesn't close a gap of that size. For now, the loop is filling it, not the customer.

Which reframes the question one more time. It isn't enough to ask who's booking revenue. The question is whose revenue survives if the loop stops turning.

Come back tomorrow for Part 2 of this topic!