The Agent Compute Bill
vishesh@vishesh
−11.0%
since Oct 6
The idea
I believe that within about four years, personal AI agents start working for people around the clock, each paying its own compute bill before any profit reaches its human. That turns this into an inference story, so I built the basket around inference silicon as the core bet, with memory, networking, and data center power as the supporting picks, plus big clouds as steadier ballast since they collect agent compute spend regardless of which chips win. I kept a cash buffer and spread bets across many names rather than one winner, planning to hold for years while this plays out.
What it owns
- Every working agent runs on chips, and the companies designing and making them collect the first slice of every agent's bill. · NVIDIAIt is the default chip agents run on today, so most compute bills flow through its hardware first as long as it keeps its lead in inference., Taiwan SemiconductorWhoever wins the agent chip race, advanced AI chips get made almost entirely here, so this take pays either way as long as demand keeps growing., BroadcomIt builds the custom chips that big clouds design to run agents more cheaply, working as long as those orders keep expanding to reduce agent operating costs., Advanced Micro DevicesIf agents deduct their own compute costs, cheap inference matters, and it is the main challenger selling chips that run AI for less than rivals at the big clouds., Marvell TechnologyIt helps clouds design their own agent-running chips, the other big custom silicon shop, growing with Amazon, Microsoft, and others as agent workloads scale.33%
- The clouds rent out the compute and collect agent bills whichever chipmaker wins, steadying the take. · MicrosoftIf agents pay a constant compute bill, it collects a big share as the cloud where much of that work runs, whoever's chips are underneath., AlphabetIt both rents out agent compute and runs its own models on its own chips, so it keeps more of each agent's bill while both businesses keep growing., OracleIt has signed huge multi-year contracts to supply AI compute, a direct way to own the agent compute bill before it arrives if customers use that capacity., AmazonIt runs the biggest compute rental business, so if millions of agents need somewhere to live, many of them pay through it while it wins AI workloads.20%
- An always-on agent economy needs electricity and cooling that the grid doesn't have spare today. · Vertiv HoldingsAgents that work around the clock throw off heat around the clock, and it sells the cooling and power gear every AI data center needs as construction stays strong., EatonIt makes the electrical equipment that gets power from the grid into AI data centers, steadily owning the buildout as long as data center demand outpaces supply., Constellation EnergyAlways-on agents need always-on power, and its nuclear plants are signing long deals to supply data centers directly at good prices., GE VernovaIt sells the turbines and grid equipment needed to generate and move the extra electricity agents will burn while data center demand keeps the order book growing.16%
- Millions of agents talking to models and each other means the gear moving data inside data centers gets bought in bulk. · Arista NetworksAgents running at scale means data centers full of chips that have to talk to each other, and it sells the switches that connect them as long as big builders keep buying its gear., Credo TechnologyIt sells the high-speed cables linking racks of AI servers, a small company that grows directly with data center buildout by the big clouds., Astera LabsIt makes the small parts that let AI chips and memory connect inside a server, pure infrastructure for the compute agents would run on if server builds keep growing., CoherentAs agent traffic grows, data centers move data with light instead of copper, and it makes those optical parts while that demand keeps rising.14%
- Agents running all day need huge amounts of fast memory, which is the scarcest part of an inference machine. · Micron TechnologyEvery agent chip needs fast memory bolted to it, and it is one of three companies making that HBM memory while it stays scarce and priced well., Lam ResearchIt sells the machines that make the memory agents depend on, getting paid when memory makers expand capacity through the investment window.9%
- Cash8%