· 3
+7.7%
since Sep 1
The idea
I think the AI buildout is really a construction problem, and the money goes to whoever controls the parts that can't be scaled quickly: grid capacity, fab tooling, networking, memory, and cooling. So I've spread across those five choke points, tilted heaviest toward power and grid since that's the slowest to fix, with no single holding dominating. I keep 8% in cash because this space can drop hard on sentiment, and I'd rather have room to buy those dips than sit fully invested through them.
What it owns
- AI datacenters require unprecedented amounts of electricity, and the power plants, transmission infrastructure, and grid equipment needed to supply them face multi-year lead times. · GE VernovaIt manufactures gas turbines and grid equipment needed for new datacenter power, with turbine orders sold out years in advance., VistraIt owns gas and nuclear generation in regions where datacenter power demand is most concentrated, and new generation cannot be built quickly., EatonIt supplies the switchgear, transformers, and power distribution equipment that every new datacenter requires, with lead times stretched to years., Quanta ServicesIt builds the transmission lines and grid connections that deliver power from plants to datacenters, with crews and permits difficult to acquire quickly., Constellation EnergyIt owns the largest nuclear fleet in the US, providing clean and always-on power that hyperscalers are locking up in long-term deals.27%
- Building leading-edge semiconductor capacity requires specialized equipment and fabs that take years and billions to construct, creating sustained demand for the companies that supply them. · Taiwan SemiconductorIt manufactures nearly every advanced AI chip, and building competing leading-edge capacity requires half a decade and tens of billions of dollars., ASML HoldingIt is the only company that makes the EUV lithography machines required to print advanced chips, making it a singular bottleneck in chip production., Applied MaterialsIt supplies deposition and etch tools that every new semiconductor fab needs regardless of the chips being made, benefiting from global fab expansion., Lam ResearchIt dominates etch equipment critical for three-dimensional stacking in chips and memory, which is the direction AI memory production is heading.24%
- Training AI systems requires thousands of processors wired together with specialized networking hardware and optical components that face supply constraints during rapid buildout. · BroadcomIt makes both the networking silicon inside AI datacenters and custom AI accelerator chips for major hyperscalers, sitting on two supply constraints simultaneously., Arista NetworksIt supplies the networking switches that wire together the tens of thousands of GPUs used in AI training clusters at major cloud providers., CoherentIt makes the optical transceivers and lasers required for data links between GPUs, facing demand that outstrips supply during rapid datacenter expansion.17%
- AI systems are bottlenecked by the speed at which memory can feed data to processors, concentrating demand among the few suppliers qualified to provide it. · Micron TechnologyIt is one of just three companies that can make the high-bandwidth memory required to feed data to AI accelerators at the speeds they need., NVIDIAIt is the primary processor and software standard for building AI systems, making it the demand engine that drives the rest of the infrastructure buildout.13%
- Modern AI datacenters generate extreme heat that requires specialized cooling systems and infrastructure, with physical construction timelines that cannot be compressed. · Vertiv HoldingsIt supplies the liquid cooling and power equipment that datacenters require to handle the extreme heat from AI workloads., EMCOR GroupIt provides the physical installation and construction services for datacenters, with skilled electrical labor as scarce as any component.11%
- Cash8%