AI Bottleneck Bets
Sparsh@sparshmehta19
−13.4%
since Oct 6
The idea
I believe the AI buildout stays supply-constrained for the next two to three years, so instead of owning the obvious chip names everyone already holds, I'm betting on the bottlenecks behind them: power and cooling, chipmaking equipment, and the networking and optics inside the clusters. The portfolio is concentrated in eight names sized to matter, with some cash held back since a basket this focused will swing hard in both directions. I'd rethink it if the buildout itself slows or these layers stop being the constraint.
What it owns
- Data centers are hitting the limits of electrical grid capacity, making transformers, turbines, cooling systems and installation crews the tightest constraint on AI expansion. · Vertiv HoldingsIt leads the market in liquid cooling and power systems that data centers need as AI racks exceed what air cooling can handle., GE VernovaIt sells gas turbines and grid equipment with multi-year backlogs to provide the power that current grids cannot deliver to data centers., Eaton CorporationIt makes the switchgear, transformers and power distribution equipment every data center requires, with lead times stretched to years due to supply constraints., Quanta ServicesIt owns the skilled crews that physically build the substations and string the wires connecting data centers to power sources, and those crews are the scarcest part of the job.45%
- Only a handful of companies manufacture the machines that produce AI chips, making them essential suppliers as new fabrication plants are built globally. · ASML HoldingIt makes the EUV machines that print the most advanced AI chips, so every new fabrication plant in the buildout must source from it., Lam ResearchIt makes the etch and deposition tools that every advanced chip and high-bandwidth memory require, and memory makers are racing to add AI capacity.23%
- AI clusters move enormous amounts of data between chips, and the switches and optical links required to do so are in short supply. · Arista NetworksIt sells the switches that connect chips inside AI clusters, and the speed of those connections determines how fast the cluster can run., CoherentIt makes some of the fastest optical transceivers at scale, and every GPU-to-GPU connection in an AI cluster runs over them.20%
- Cash12%