Cheap AI, Scarce Hardware
David Greenstein@davidgreenstein
−12.9%
since Oct 9
The idea
I believe cheaper AI means much more usage, and that usage is colliding with real physical shortages in memory, storage, chips, and power. So this take leans hardest into memory and storage, where the shortage looks most durable, then spreads the rest across chipmakers, power and cooling suppliers, the tool makers they depend on, and steadier data center landlords. No single holding dominates, a slice stays in cash, and I'd rethink this if cheaper AI starts meaning less memory per answer rather than more usage overall.
What it owns
- High-speed memory and storage are the tightest bottlenecks in AI infrastructure, with the best profit margins and lowest valuations relative to earnings. · Micron TechnologyIt makes the high-speed memory that AI chips depend on, and most of its near-term output is already sold to customers under contract, meaning the shortage persists as long as orders stay committed., SK hynixIt supplies most of the high-bandwidth memory stacked directly next to major AI chips, covering the memory market with Micron regardless of which processor wins., SandiskIt makes flash storage for the data that AI systems generate and store, which grows with each new user and each new application., Seagate TechnologyIt makes high-capacity hard drives that store most of the data in cloud centers and shares that market with one other major competitor., Western DigitalIt makes hard drives for bulk cloud storage and together with one peer controls the duopoly, capturing pricing power as long as supply stays tight.35%
- Every AI system runs on these chips, so this layer captures value regardless of which architecture or company wins the competition. · NVIDIAIt makes the default chips for running AI systems, and each one shipped pulls in high-speed memory from other members of this take., Taiwan Semiconductor ManufacturingIt manufactures nearly all advanced AI chips for every major designer, so it gets paid whichever company leads the market., BroadcomIt designs custom AI chips for major cloud companies to reduce the cost of running AI, the exact trend this take targets., Arista NetworksIt makes networking switches that connect thousands of AI servers inside data centers, and that need grows with every new cluster built.23%
- Data centers cannot operate without turbines, electrical infrastructure and cooling systems, all of which face multiyear backlogs. · GE VernovaIt makes gas turbines that power new data centers and reports being sold out through the early part of the next decade with agreements signed years beyond that., Vertiv HoldingsIt sells cooling and power equipment inside data centers that must handle the rising heat from faster AI servers., EatonIt makes transformers, switches and electrical gear that connect data centers to power networks, all in short supply with long lead times., Quanta ServicesIt builds the power lines and grid connections that new data centers need before they can turn on.18%
- Memory makers can only increase output by buying equipment from these suppliers, so their growth depends on continued factory expansion. · Lam ResearchIt sells the equipment that memory makers buy to build new factories, so it captures growth from efforts to expand capacity and ease shortages., ASML HoldingIt makes the only machines capable of printing the most advanced chips and memory, which every new leading-edge factory needs to add capacity.8%
- These companies collect rising rents from growing data center occupancy as AI usage expands, providing steadier returns than hardware makers. · EquinixIt rents data center space where companies connect directly to cloud providers, collecting rising rents as occupancy and demand grow., Digital Realty TrustIt rents large data center buildings to cloud and AI companies, collecting rent from whoever wins while demand for new space keeps rising.7%
- Cash9%