−11.8%
since Oct 8
The idea
I believe the AI buildout is really a physical infrastructure race, and the real money goes to whoever controls the parts that can't be expanded quickly, not just the headline chip names. So I built this around five bottlenecks: power and grid, foundry and fab tooling, networking, memory and compute, and cooling plus datacenter construction, weighted most heavily toward power and grid since that's the slowest constraint to fix. No single holding dominates, and I keep 8% in cash because this space tends to sell off hard on sentiment, and those drops are when I want to add.
What it owns
- New datacenters require massive amounts of electricity, and the power plants, grid infrastructure, and distribution equipment needed to supply them face years-long backlogs. · GE VernovaIt manufactures the gas turbines and grid equipment that new datacenter power demand requires, with its order book sold out years in advance., VistraIt owns generation capacity in regions facing the tightest datacenter power demand, where new capacity cannot be built quickly., EatonIt supplies the switchgear and power distribution equipment every new datacenter needs, with lead times extending years into the future., Quanta ServicesIt builds transmission lines and grid connections connecting power plants to datacenters, work that requires permits and crews that cannot be assembled quickly., Constellation EnergyIt operates the largest US nuclear fleet, providing clean always-on power that hyperscalers are securing through long-term agreements.27%
- Advanced AI chips require cutting-edge semiconductor manufacturing capacity that takes years and billions to build, creating a structural supply constraint. · Taiwan SemiconductorIt manufactures nearly every advanced AI chip, and building competing leading-edge capacity takes half a decade and tens of billions of dollars., ASML HoldingIt is the only manufacturer of EUV lithography machines required to print advanced chips, making it the singular bottleneck in cutting-edge semiconductor production., Applied MaterialsIt sells the deposition and etch tools that every new fabrication facility needs regardless of which chips it manufactures., Lam ResearchIt dominates the etch equipment critical for three-dimensional chip stacking, the direction AI memory production is heading.24%
- Training massive AI systems requires thousands of GPUs connected by specialized networking hardware and optical components that form critical supply chokepoints. · BroadcomIt designs both the networking silicon inside AI datacenters and custom accelerators for hyperscalers, occupying two critical chokepoints simultaneously., Arista NetworksIt provides the switches that wire together the tens of thousands of GPUs needed for AI training clusters at major cloud providers., CoherentIt manufactures the transceivers and lasers that enable optical links between GPUs, a component supply facing strain from AI datacenter demand.17%
- AI accelerators are bottlenecked by the memory systems feeding them, with only a handful of suppliers capable of producing the specialized components required. · Micron TechnologyIt makes the high-bandwidth memory that AI accelerators depend on, with only three companies globally capable of production at this scale., NVIDIAIts chips and CUDA software remain the default platform for building AI systems, making it the demand engine that drives the rest of this basket.13%
- AI workloads generate extreme heat that forces datacenters to adopt liquid cooling and specialized infrastructure, creating demand for specialized equipment and construction services. · Vertiv HoldingsIt provides the liquid cooling and power equipment that datacenters require when air cooling becomes insufficient for AI workloads., EMCOR GroupIt contracts for the physical wiring and plumbing of new datacenters, with skilled electrical labor as constrained as any component.11%
- Cash8%