Follow the Token: Where AI's Profits Actually Go - Part 2

If you haven't read yesterday's blog post yet, please do so before continuing here.

Layer Three: Who Actually Keeps It

This is the layer the headline coverage rarely reaches, and it's the one that matters most for anyone allocating capital rather than just reading about the trade. Separate the money that gets recycled back into the ecosystem from the money that comes off the table for good. By that test, a few groups stand out.

First, the choke-point monopolies. TSMC and ASML sell to every side of the contest and depend on none of them individually. If the question is which positions are hardest to disintermediate when the cycle cools, names the whole field has to buy from tend to screen well on that pattern.

Second, the financiers. The buildout has moved off balance sheets and into debt markets, with AI-related debt issuance set to roughly double toward $570 billion in 2026. Lenders earn their yield on the schedule, regardless of whether the models ever clear their hurdle rate. The caution worth flagging is the collateral. Some of this debt is secured against hardware whose rental rates have already fallen sharply, so the quality of the security behind the yield is itself a condition to watch.

Third, and least intuitively, the hyperscalers themselves. They look like the biggest losers in the opening paradox, spending far ahead of what they earn back. But spending isn't the same as losing if you end up owning the asset. Custom chip shipments from the cloud providers are projected to grow far faster than merchant GPU shipments in 2026, which is the statistical fingerprint of the buyers becoming the builders. In the original gold rush, the miners and the shovel sellers were different people. In this one, the miners are increasingly making their own shovels, which changes who holds the durable asset when the dust settles.

The Winner Off the Value Chain

There's one more group, and it doesn't appear anywhere on the diagram of chips, power, and clouds. The most cleanly realized money this cycle is being made in the capital markets, by selling the story rather than the product. When a marquee private company can raise a record sum while still running losses, the lesson is that a strong narrative can outrun profitability for a long time, and that selling equity into that enthusiasm is its own form of getting paid. Founders, early backers, and the bankers who run the books realize cash today against profits that may or may not arrive later. It isn't part of the value chain in any operational sense. It's arguably the most reliable profit center in the whole picture.

Where the Profits Actually Go

Pull the threads together and an uncomfortable pattern emerges. The further a name sits from the end customer's wallet, the more reliably it's been paid. Near-term accounting profit concentrates in the supply chain and the physical layer. More durable profit gathers at the choke-point monopolies, the financiers holding the paper, and the integrated hyperscalers who end up owning the infrastructure. And the most cleanly realized cash sits with the people monetizing the story in the capital markets.

Put plainly, the profits are flowing away from where the value is supposedly being created, the helpful model answering the prompt, and toward the infrastructure underneath it and the people financing that infrastructure. The token goes in at the top. The profit settles at the bottom.

Whether that arrangement holds depends on a single condition, and it's worth stating without dressing it up as a forecast. At some point the end customer has to start closing that seventeen-to-one gap, or the recycled dollars stop recycling. The pattern described here is what the money is doing now. It isn't a claim about what it has to do next.

The Allocator's Lens

For anyone watching this as a practitioner rather than a spectator, the useful shift is in the question itself. Not which AI company wins, which is a bet on an outcome nobody can price yet, but which layer keeps the cash even if customer returns never quite show up. A few illustrative things to track, offered as observations rather than recommendations:

• Margin durability at the choke points, since that's where pricing power would show up first.
  
• The ratio of recycled to end-customer revenue, which is the health check on the whole structure.
  
• The quality of the collateral behind the debt, which is where stress would surface early.
  
• Any sign the loop is tightening, money circulating faster among the same handful of players, which historically rhymes with the late stages of a buildout rather than the early ones.

Follow the token far enough down the stack and the profits turn up where the recycling stops. That isn't where most of the attention is pointed. Which, for anyone doing the following, is rather the point.