The Billion Agent Buildout
Kevin Li@liveink
−8.4%
since Sep 29
The idea
I believe personal AI agents are going to multiply until everyone is running many of them, and that means inference demand becomes permanent, always-on load rather than a one-time training spike. So instead of picking a winning agent platform, I built a basket around the toll roads underneath: chips and foundry capacity, memory and bandwidth, networking, and the power and data centers that keep it all running. It leans on large steady names for ballast, keeps a modest cash reserve to add on dips, and is meant to be held for a decade or more as this buildout plays out.
What it owns
- The chips agents run on, and the one foundry nearly all of them come from. · NVIDIAMost background thinking from always-on agents runs on its chips first, as long as it remains the default inference choice at scale., Taiwan SemiconductorNearly every serious AI chip gets manufactured by it, so the billion-agent buildout pays a toll to it no matter which chip designer wins., BroadcomIt manufactures the custom inference chips that hyperscalers design to reduce their compute costs., Advanced Micro DevicesBuyers running always-on agents won't want to depend on one chip supplier, so it benefits as a credible second source if it wins real inference deployments.27%
- Constant inference load needs electricity and cooling around the clock, and the grid is the slowest thing to build. · Vertiv HoldingsIt supplies the cooling and power infrastructure that keeps chips running continuously under permanent inference loads., EatonEvery new data center needs the electrical gear and distribution it supplies, and it is established enough to anchor a long-term data center buildout bet., GE VernovaIt builds the turbines and grid equipment that utilities and developers need to generate and deliver the new power capacity that sustained AI inference requires., Constellation EnergyAlways-on agents need always-on carbon-free power, and its nuclear fleet is what big AI buyers are signing long-term data center power deals for., VistraIt sells the round-the-clock electricity that permanent inference load consumes, with nuclear and gas plants near where data centers are being built.24%
- Billions of agents talking to models and each other means data center networks get rebuilt bigger and faster. · Arista NetworksAgents in data centers need fast switching between machines, and the big AI builders standardize on its networking gear., Marvell TechnologyIt makes the chips that move data between processors and across data centers, and a decade of inference growth means a decade of interconnect demand., CoherentAt AI scale, data moves between racks over light through its optical parts, and more agents means more traffic between machines., Credo TechnologyInside the data center, its cables and chips carry data at speeds copper barely handles, and always-on inference needs ever more short, fast connections., Astera LabsIt makes the connectivity chips that let processors, memory, and accelerators talk to each other inside AI servers designed for agent workloads.22%
- Always-on agents are fed by memory as much as by processors, and high-bandwidth memory is scarce. · Micron TechnologyAgents thinking constantly pull data from memory all day, and high-bandwidth memory demand from AI keeps outrunning supply., RambusIt collects royalties on interface chips between processors and memory, and gets paid on the upgrade cycle as inference pushes memory speeds higher.10%
- Someone owns the buildings all this compute sits in and collects rent no matter whose agents win. · EquinixIt houses all this compute worldwide and collects rent from it, keeping pricing power if agent inference stays permanently scarce., Digital RealtyIt leases data center real estate where hyperscalers deploy their inference hardware, and capacity constraints support sustained rental pricing.9%
- Cash8%