AI Buildout Chokepoints
Rebecca
@rebecca
· 4
+18.5%
since Jul 30
The idea
I believe AI will reshape the knowledge economy the way machinery reshaped industry, and in this buildout phase the value accrues to whoever controls the physical bottlenecks, not just the headline chip names. So this portfolio spreads across five choke points: power and grid equipment, foundry capacity, networking, memory, and cooling, tilted heaviest toward the constraints that take years to expand. It holds 16 positions with no single one dominating, plus a modest cash reserve to add on the sharp drawdowns this sector tends to serve up.
What it owns
- Data centers need staggering amounts of electricity, and the grid equipment and generation to supply it takes years to build, making it the slowest bottleneck to fix. · GE VernovaIt dominates gas turbine and grid equipment making with a multi-year order backlog driven by datacenter power demand, creating a textbook chokepoint., VistraIt operates nuclear and gas generation fleets selling into markets where datacenter demand is tightening supply, benefiting directly from rising power prices., EatonIt makes the switchgear, transformers, and power distribution equipment every datacenter needs, with long lead times providing pricing power., Constellation EnergyIt operates the largest US nuclear fleet, signing premium long-term power deals directly with hyperscalers who need clean, always-on electricity that cannot be quickly replicated., Quanta ServicesIt is the largest specialty contractor building transmission lines and grid upgrades, owning the largest pool of skilled electrical labor as that becomes a bottleneck.27%
- Every AI chip, whoever designs it, must be manufactured in a handful of advanced fabs using machines only a few companies on Earth can make. · Taiwan SemiconductorIt manufactures essentially every leading-edge AI chip regardless of who designs it, making it an agnostic chokepoint across the entire industry., ASML HoldingIt is the only company on Earth that makes the EUV lithography machines required for advanced chip manufacturing, a literal monopoly., Applied MaterialsIt is the broadest supplier of chip manufacturing equipment, selling into every new fab built worldwide as fab construction booms globally., Lam ResearchIt leads in etch and deposition equipment, especially exposed to the memory capacity expansion that high-bandwidth memory demand is forcing.24%
- AI clusters are only as fast as the connections between chips, making switches and optical links a quieter but real chokepoint. · BroadcomIt makes the networking chips inside AI datacenters and designs custom AI accelerators for hyperscalers, combining two chokepoints in one company., Arista NetworksIt dominates high-speed Ethernet switches wiring together AI clusters at hyperscalers, with networking spend growing faster than compute spend as clusters scale., CoherentIt is a key maker of optical transceivers carrying data between AI racks, addressing an underappreciated bottleneck as copper hits physical limits at AI speeds.17%
- AI models are starved for high-bandwidth memory, which is in structural shortage, plus a core position in the compute leader itself. · Micron TechnologyIt is one of only three makers of high-bandwidth memory that AI accelerators consume in huge quantities, with supply sold out well into the future., NVIDIAIt sets the pace of the entire buildout, and its roadmap drives demand for every chokepoint in this basket.13%
- AI racks run so hot that liquid cooling and specialized construction are now mandatory, not optional, and few firms can deliver at scale. · Vertiv HoldingsIt leads in datacenter power and cooling infrastructure, with liquid cooling now mandatory for dense AI racks and order backlogs compounding across every new facility., EMCOR GroupIt is a major electrical and mechanical contractor building datacenters themselves, with skilled construction capacity scarce and backlogs reflecting multi-year buildout runways.11%
- Cash8%