−9.3%
since Sep 30
The idea
I think the AI buildout is a physical infrastructure race, and the real money goes to whoever controls the parts that can't be scaled quickly, not just the chips everyone talks about. So this basket leans into the slowest bottlenecks: power and grid capacity, foundry tooling, networking, memory, and cooling and construction, with the heaviest weight on power since that's the hardest to fix. No single holding dominates, and I keep 8% in cash on purpose, since this space swings hard on sentiment and those drops tend to be the entry points.
What it owns
- AI datacenters consume massive amounts of electricity, and building generation capacity, transmission lines, and distribution equipment creates multiyear backlogs. · GE VernovaIt manufactures gas turbines and grid equipment for new datacenter power generation, with order books sold out years in advance., VistraIt operates nuclear and gas generation in regions where datacenter demand is most concentrated, and new capacity takes years to build., EatonIt makes switchgear, transformers, and power distribution equipment that every new datacenter requires, with lead times extending multiple years., Quanta ServicesIt constructs transmission lines and grid connections between power plants and datacenters, and that infrastructure cannot be built quickly., Constellation EnergyIt operates the largest US nuclear fleet, providing the clean baseload power that hyperscalers are locking in through long-term agreements.27%
- Advanced chip manufacturing requires specialized equipment and process nodes that take years and billions to replicate, creating lasting supply constraints. · Taiwan SemiconductorIt manufactures nearly every advanced AI chip using process nodes that competitors cannot replicate for years, making fab capacity the limiting factor in chip supply., ASML HoldingIt is the sole maker of EUV lithography machines required to print the most advanced chips, creating a singular bottleneck in the semiconductor supply chain., Applied MaterialsIt sells deposition and etch tools that every semiconductor fab needs regardless of the chips being made, winning on total fab construction volume., Lam ResearchIt dominates etch equipment critical for stacking chips and memory in three dimensions, which is the direction AI memory architecture is moving.24%
- Training large AI models requires connecting thousands of processors with specialized networking hardware and optical components that face supply constraints. · Arista NetworksIt makes the networking switches that connect tens of thousands of processors in AI training clusters for major cloud providers., BroadcomIt designs both the networking silicon for AI datacenters and custom accelerators for major hyperscalers, capturing two critical components in the buildout., CoherentIt manufactures optical transceivers and lasers that move data between GPUs, and demand from AI infrastructure is outpacing supply.17%
- AI systems are constrained by memory bandwidth and processor performance, both of which require specialized components from a limited set of suppliers. · Micron TechnologyIt makes high-bandwidth memory that AI accelerators depend on, and only a handful of companies can produce the specialized modules that data throughput requires., NVIDIAIt makes the processors and software stack that drive AI training, and its dominance in both hardware and the CUDA ecosystem remains unchallenged.13%
- Modern AI datacenters generate extreme heat and require physical infrastructure that takes time and specialized labor to construct and install. · Vertiv HoldingsIt supplies liquid cooling and power infrastructure for datacenters where air cooling becomes insufficient, and adoption is accelerating across the industry., EMCOR GroupIt performs the electrical and mechanical contracting work to physically build and outfit datacenters, with skilled labor remaining constrained.11%
- Cash8%