Labor to Compute Bottlenecks
Bryce Bjork@bryce
−14.6%
since Oct 7
The idea
I believe the next five to ten years sees spending shift from payroll to compute, and the real winners are the chokepoints everyone has to pay to build capacity, not just the headline chip names. This basket spreads across memory and packaging, fab equipment, networking and optics, power and cooling gear, and data center landlords, pairing steadier large companies with tighter supply-squeeze plays. It keeps a cash buffer and avoids concentrating in any single name, holding for years while the bottleneck thesis plays out.
What it owns
- Electricity and heat, not chips, now decide how fast AI capacity gets built, and this gear is what's scarce. · Vertiv HoldingsIt leads in liquid cooling, which denser AI racks increasingly cannot run without., EatonIt sells the electrical gear every new data center needs before it can turn on, and power equipment is one of the gates on how fast capacity comes online., GE VernovaIt sells the turbines and grid equipment that generate and deliver the power all this compute needs., Quanta ServicesIt builds the actual grid hookups, the queue everyone is stuck in, and gets paid to shorten that line., nVent ElectricIt makes the rack-level power and cooling hardware inside every data center build, a smaller play on the same gate as the larger power suppliers.24%
- Tying thousands of chips into one computer takes switches and fiber optics that are now among the most supply-constrained parts. · Arista NetworksIt sells the switches that tie AI clusters together, one of the networking bottlenecks where lead times have extended significantly., BroadcomIt makes the switch silicon inside most data center networking plus custom chips for the biggest compute buyers, so it sits at two chokepoints at once., Marvell TechnologyIt sells the interconnect silicon that moves data between chips in a cluster, which gets more valuable as clusters get bigger., CoherentIt makes the fiber optic transceivers that are one of the parts with blown-out lead times right now., Credo TechnologyIt is a smaller, faster-growing bet on the cables and links inside clusters, the tightest end of the networking squeeze.23%
- Every new chip factory, whoever owns it, buys from the same handful of toolmakers. · ASML HoldingIt is the only company that makes the EUV machines advanced fabs need, so every dollar that moves from labor to compute eventually buys one of its tools., Lam ResearchIt makes the tools memory fabs buy when they expand, and with memory sold out, expansion is the only way out., Applied MaterialsIt is the broadest fab toolmaker, so it gets paid on almost any chip capacity being built, memory or logic., KLA CorporationIt inspects chips for defects, and when capacity is this scarce, no fab can afford bad yield, so its tools get bought alongside every expansion.18%
- AI data centers are on track to consume most of the world's memory output, and the few companies that make it set the price. · Micron TechnologyIt makes the high-bandwidth memory that every AI system needs more of, and memory is sold out through the coming years., Taiwan SemiconductorNearly every advanced AI chip has to pass through its advanced packaging lines, which makes it a chokepoint no buyer can route around., Amkor TechnologyIt sells the overflow packaging capacity chipmakers turn to when their primary suppliers' lines are full.16%
- Whoever rents finished, powered space collects from every tenant in the compute scramble. · Digital Realty TrustIt rents out powered data center space, and powered space is exactly what the grid queue makes scarce., EquinixIt owns the interconnection hubs where networks meet, so it collects rent on the traffic the compute shift creates regardless of which chipmaker wins.9%
- Cash10%