This is an opinion. It’s built on reasoning, not a dataset I can point to, so test it against your own deal instead of just adopting it.
The question is where the money actually piles up in the AI compute stack. Not where the activity is, or the headlines. Those are often different places. My answer: the stack is a barbell. Value pools at the top, in the chips and memory that are scarce, and at the bottom, in the orchestration and brokerage layer that sits closest to the customer and owns none of the depreciating hardware. The middle, the layer that owns the racks, is the hardest seat at the table. Understand why before you put capital into it.
The top of the barbell: scarcity priced into silicon
Chip and memory vendors sit at the top. Right now their advantage is structural, not cyclical. Demand for the highest-end accelerators, and the high-bandwidth memory that goes in them, is outrunning the industry’s ability to add capacity for either.
Memory shortages in particular tend to push prices up across a whole product line, not just the AI parts, because every other memory buyer is fighting over the same fab capacity.
A vendor sitting on a real supply bottleneck gets to price to the value its customers capture, not to its own cost of production. That gap is where the margin is. This is the easy end of the barbell to understand, because markets have priced it for a long time: when a necessary input is scarce, whoever controls it takes an outsized share of the value created downstream.
The bottom of the barbell: owning the relationship, not the metal
At the other end is the orchestration, software and brokerage layer. These companies route each unit of work to whatever capacity is the cheapest fit at that moment. They package that routing into something a customer can buy without understanding the hardware, and they handle the contracts and billing on top.
This layer owns almost no depreciating hardware. Its assets are the routing logic, the customer relationship and the contracts. Those depreciate far more slowly than silicon, if at all.
A brokerage or orchestration business that gets routing right earns a margin on every unit of compute flowing through it, without ever carrying the balance-sheet risk of the hardware underneath. That’s a very different risk profile from owning a rack of GPUs. It’s why this layer can make money even while the hardware owners below it fight for margin.
The middle: project finance wearing a GPU costume
The seat in between, whoever actually buys, owns and runs the hardware, is the hardest one. The reason isn’t operational. It’s financial.
Buy racks of accelerators at today’s prices, finance that against a customer’s prepayment or a multi-year offtake, and what you have underneath the GPU branding is the same structure as project finance for a power plant or a toll road. A big upfront capital outlay. Recovered over a multi-year contracted revenue stream. Financed with debt priced against the credit of whoever signed that revenue stream.
What makes this version harder than the power plant: the chip depreciates and goes obsolete far faster than a turbine or a stretch of road, while the financing wrapped around it still wants a multi-year horizon to make the numbers work.
That mismatch is the whole problem.
A lender financing this middle layer is underwriting two risks at once. Will the customer behind the offtake actually keep paying for the full term? And will the hardware still be worth running, not quietly obsolete, when that term ends? Get either one wrong and the economics fall apart. Not gradually, either. They fail the way financed assets fail: debt service doesn’t pause because the asset stopped earning what the model said it would.
What makes an offtake bankable
In this framing, the quality of the customer behind the offtake is worth more than almost anything about the hardware. The cost-of-capital arithmetic in what it actually costs to run a GPU cluster gets to the same place from a different direction. Moving the embedded cost of capital on a financed cluster by a few points changes the breakeven by more than most of the haggling over the hardware does, because a lender is pricing the strength of the revenue behind the deal, not the chip.
A bankable offtake has a few concrete features:
- A counterparty whose credit can actually carry a multi-year commitment.
- A term long enough for the hardware to pay for itself, but not so long it outlives the hardware’s useful economic life.
- Language that survives a change of ownership on either side, so the lender isn’t holding collateral against a contract that evaporates if either party gets acquired or restructured.
An offtake missing any of those three is weaker collateral than it looks on a term sheet. A lender who knows this asset class will price it that way. That flows straight into the rate the middle-layer owner pays, and so into whether owning the middle layer makes sense at all.
Depreciation versus scarcity
Here’s the tension that makes the middle seat structurally hard, not just hard right now.
The silicon in a financed cluster today is scarce, and it commands a premium because it’s scarce. But scarcity in this market has a short half-life. New fab capacity gets built. New chip generations arrive. A part that was scarce and expensive eighteen months ago is often a commodity at a fraction of the price, well before the debt against it is paid down.
So the middle-layer owner is betting that contracted revenue pays off the debt faster than the scarcity premium erodes. Win that race and the middle layer looks like a good business. Lose it and you’re holding a depreciating, increasingly commoditized asset against a debt schedule that assumed it would stay valuable longer.
That race, not any single operational decision, is the defining risk of this seat. It’s also why the basis risk in can you resell GPU compute you already bought isn’t a side concern for a middle-layer owner. It’s close to the main one. Being long a depreciating asset with no good exit if the market moves is exactly where this layer sits.
Where to put capital
None of this means the middle layer is uninvestable. It means the bar for getting in correctly is specific, and people skip it.
Only go in with an offtake that’s bankable by the standard above, financing priced against that strength rather than the hardware alone, and a term matched to the hardware’s realistic economic life, not its marketing life. The vetting questions for anyone operating in this layer, about leverage, loan-to-value, and what happens if a lender forecloses on the hardware, are in how do you vet a GPU cloud provider before you wire the deposit. They apply to your own position in this layer just as much as to someone else’s.
If a deal can’t clear that bar, the more defensible money is at the ends of the barbell. Direct exposure to the scarcity at the top, or ownership of the routing and customer relationship at the bottom. Neither one carries the mismatch between a fast-depreciating asset and a multi-year financing structure that makes the middle the hardest seat in the room.