While the model labs burn capital to win users, the companies one layer down are booking the steadiest profit in the entire boom. The clouds that rent out compute and the firms that design the chips are not waiting for an agent market to appear. They are paid today, in advance, by everyone racing to build one. In a gold rush, the reliable money has always been in picks and shovels — and this rush is no different.
The shovels got repriced
A few years ago, selling servers and accelerators was a respectable, cyclical hardware business. The AI build-out turned it into the most strategic supply chain on earth. Demand outruns supply, buyers compete on allocation rather than price, and the sellers set terms. That is the opposite of the labs' situation, where buyers compete on price and the sellers eat the loss. The middle of the stack captured pricing power that the top of the stack gave away.
Why the middle is defensible — for now
The advantage runs deeper than scarcity. Leading-edge chips depend on a tiny number of fabs, packaging capacity, and design talent that cannot be conjured in a quarter. Clouds layer on switching costs: once a team's data, tooling, and pipelines live in one place, moving is painful. So the middle enjoys both a hard physical bottleneck and a soft lock-in. That is a rare and durable combination.
The risk to the middle
Durable is not permanent. The same scarcity that prints money is also the loudest possible signal to build competing supply, and capacity is being poured in across regions and vendors. If demand ever cools while that new supply lands, the pricing power that defines the middle today flips to oversupply fast — and a business that looked like a tollbooth starts to look like a commodity again.
Our read
The picks-and-shovels layer is the safest seat in the boom, but it is safe because of a shortage that everyone is racing to end. The winners will be the ones whose lock-in survives the moment compute stops being scarce. We covered why the labs above them are betting on a market that does not exist yet in Frontier Labs Are Betting on a Market That Doesn't Exist Yet. Next, we go up to the layer closest to users — and the hardest place to defend: Building on Someone Else's Model Is a Rented Moat.
◆
The Signal
AI-generated brief
Compute suppliers and cloud platforms capture the sector’s most stable margins today, but their profitability depends entirely on a temporary shortage that intense capacity investment aims to erase.
Stance · CautiousConfidence · Established
The analysis confirms strong near-term economics but frames them as structurally vulnerable to inevitable supply normalization and competitive replication.
Key takeaways
Infrastructure layers secure predictable revenue through upfront commitments and seller-driven pricing, contrasting sharply with the cash-burn strategies of model developers.
Defensibility combines hard physical bottlenecks in advanced manufacturing and packaging with soft lock-in from entrenched enterprise data pipelines and tooling.
Current pricing power carries inherent vulnerability, as rapid global capacity expansion threatens to normalize supply and revert services to commoditized pricing if demand decelerates.
Long-term viability requires transitioning from scarcity-based leverage to enduring ecosystem stickiness before compute becomes a standardized utility.
What to watch next
Global foundry capacity utilization and backlog clearance rates
Enterprise multi-cloud adoption and workload portability metrics
Shifts in committed versus spot compute pricing structures
