AI Tollbooths, Real Cash
seansalmon 🤷♂️@seanaes
−11.1%
since Oct 7
The idea
I believe the real money in AI gets made by the picks-and-shovels businesses sitting underneath every chip architecture, not by whoever wins the GPU race. So this basket owns the tollbooths: foundry, EDA and inspection names that get paid on every design and wafer, networking and power gear that gets paid on every rack, plus a smaller slice of capital-light software and the one hyperscaler still funding its own buildout with operating cash. I kept about 8% in cash and spread bets across many names rather than leaning on one.
What it owns
- Every AI chip, GPU or custom, gets designed with EDA software, made at advanced foundries and checked with inspection tools, so these companies collect on whichever accelerator wins. · Taiwan Semiconductor ManufacturingIt makes the most advanced chips for NVIDIA, Broadcom's custom chip customers, Google's TPUs and AMD, so it gets paid whichever accelerator wins., KLA CorporationIt sells the inspection and measurement tools every chipmaker needs to get good yields at new chip sizes, no matter whose design is on the wafer., Cadence Design SystemsNearly every chip, GPU or custom ASIC gets designed on Cadence or Synopsys software, so it collects on design work whichever architecture wins., ASML HoldingIt is the only company that makes the EUV machines needed for leading-edge chips, so every advanced foundry has to buy from it., SynopsysIt is the other half of the chip design software duopoly, paid on every new GPU, ASIC and TPU design.36%
- Every accelerator cluster needs switches, custom silicon, connectors and power chips to work, and these suppliers do it without running costly factories of their own. · BroadcomIt builds custom AI chips for big cloud companies and its switching chips connect GPU clusters, so it wins whichever architecture gains ground., AmphenolIt makes the connectors and cables inside every AI rack, whatever accelerator sits in it., Arista NetworksIt sells Ethernet switches for AI clusters that work with any accelerator, so it doesn't depend on any one chip to win., Monolithic Power SystemsIt designs the power chips that feed accelerators on their boards, and being fabless means it outsources manufacturing and needs little capital.29%
- Software and cloud businesses that still turn most of their operating cash into free cash, adding AI upside without depending on the buildout itself. · MicrosoftIts operating cash still covers huge data center spending with real free cash left over, plus it earns from software that needs no new buildings., ServiceNowIt sells workflow software and AI agents on top of it without building data centers, so AI demand turns into subscription revenue and free cash., DatadogIt monitors cloud and AI applications, so more workloads of any kind mean more usage, with very little capital needed.14%
- Power is the physical bottleneck for every data center no matter what chips go inside, and these companies sell the gear that keeps them running. · EatonIt sells the switchgear, transformers and backup power that every data center needs regardless of the chips inside., VertivIt supplies the cooling and power gear for racks, and denser AI chips of any kind need more of it, especially liquid cooling., GE VernovaIt builds the gas turbines and grid equipment that power-hungry data centers are lining up for, a bottleneck no matter which chip wins.13%
- Cash8%