In most industries, the company closest to the customer keeps the biggest margin. In AI, Apollo's chief economist says it's exactly backward.
Torsten Slok grouped the AI value chain into four layers: energy and grid suppliers, chip and equipment makers, cloud and compute infrastructure, and the models and applications layer where OpenAI and Anthropic actually sell AI products to customers. Margins rise steadily the further a company sits from that last layer.
Where the Money Actually Goes
The upstream layers are genuinely profitable. Energy and grid companies like Constellation, NextEra and Vertiv, chipmakers like Nvidia and Micron, and cloud infrastructure providers like Equinix and CoreWeave are all capturing real margin from the AI buildout.
The layer selling the actual product to end customers is a different story. Based on Apollo's estimates for the second quarter of 2026, OpenAI and Anthropic are running deeply negative margins, funded by capital raised from investors rather than revenue earned from customers.
Those specific figures are estimates, not audited results. Apollo sourced OpenAI's number from PitchBook and Anthropic's from the Financial Times, both third-party approximations rather than numbers either company has confirmed itself. The direction of the finding, upstream layers profitable and the model layer not, is the more durable takeaway than the exact percentages.
"The upstream margins are real, but they are paid for out of capital raised by the layer losing money, not out of cash generated by end demand," Slok wrote.
Why This Is Backwards
Slok's phrase for it is blunt: the 41% depends on the -59%. The most profitable part of the chain only exists because the least profitable part keeps raising enough money to keep buying from it.
That's not necessarily a crisis on its own. Capital can bridge a profitability gap for a while.
Slok's point is that it can't do so forever. At some point the model layer's ability to keep raising the capital that funds everyone upstream of it becomes the binding constraint on the whole industry's growth, not the technology itself.
The Question Nobody's Answered Yet
The whole structure depends on one unresolved question: will AI's actual return on investment show up fast enough for end customers to justify the spending that's currently propping up those upstream margins.
If it does, the current structure is just an early, capital-intensive phase of a normal technology buildout. If it doesn't, every upstream layer currently profiting from AI spending is exposed to a slowdown at the one layer paying for all of it.
That framing maps directly onto stories MAIN has already covered this year. Broadcom's roughly $100 billion debt raise financing Anthropic's chip supply, Nscale's push toward a $3 billion IPO built on data center contracts, and Nvidia's own price hikes passed through to customers are all upstream-layer activity, precisely the part of the chain Slok's data says is currently profitable because the model layer keeps financing it.
What This Means for Miami
This is the clearest single articulation yet of the concentration risk J.P. Morgan's Bill Eigen has spent this year urging Miami investors to take seriously, now with a specific mechanism attached rather than a general warning.
Apollo's own framing makes the stakes concrete: an investor holding upstream AI infrastructure exposure, chips, data centers, power generation, is making a bet that is only as good as OpenAI and Anthropic's ability to keep raising capital. That's a different risk profile than owning a profitable business outright, and it's worth Miami's family offices and institutional investors treating it as exactly that when they weigh AI infrastructure bets against more conventional ones.