Agent Compute Buildout
Christopher Wallace@chris
−9.0%
since Sep 29
The idea
I believe AI agents that run in loops and call tools for minutes at a time will keep multiplying demand for inference compute over the next three to five years, so I built a basket around the whole stack. It leans into inference-heavy chipmakers and interconnect, holds the foundry and fab-tool chokepoints every chip must pass through, and balances that with the big clouds renting the compute by the hour. I keep a cash reserve to add on the sharp dips this sector tends to serve up, and I am holding for years, not trading the swings.
What it owns
- Agents require far more compute per task than chatbots, and inference chips are where that work runs. · NVIDIA CorporationIt supplies the dominant chip for running inference workloads, holding its position as long as its accelerators remain the default platform for agent deployments., BroadcomIt designs custom inference accelerators that large cloud providers manufacture to reduce their chip costs, and growing agent demand drives more custom silicon orders., Advanced Micro DevicesIt is the credible alternative source for inference chips that cloud providers want to reduce single-supplier risk, and its competitive position depends on continued wins in inference deployments., Micron TechnologyIt supplies high-bandwidth memory that stores context for long-running agents, benefiting from tight memory capacity as inference demand grows., Marvell TechnologyIt is the secondary supplier of custom inference chips to large cloud providers, and growth depends on cloud custom silicon programs expanding beyond what a single vendor can supply.30%
- Agents execute in cloud environments, so cloud providers collect revenue for every hour of compute that agents consume. · MicrosoftIt rents GPU compute hours to enterprises through Azure and benefits from its tie-up with OpenAI, collecting revenue for every hour of agent inference deployed on its cloud., AlphabetIt owns both its TPU inference chips and the cloud infrastructure they run on, capturing the full stack value if Google Cloud expands its share of agent inference workloads., AmazonIt operates the largest cloud compute rental business and builds its own Trainium and Inferentia chips, capturing a large share of inference hours billed to customers., OracleIt transitioned into supplying GPU cluster capacity for model companies, and revenue conversion depends on whether that backlog translates into sustained agent inference demand., CoreWeaveIt rents only GPU compute capacity, making it the purest play on whether inference demand is actually multiplying, and it remains viable only if utilization and contracts keep growing., Nebius GroupIt is a smaller GPU cloud provider that grows if agent workloads exceed what large cloud providers can supply, depending on continued customer wins and chip access.30%
- Agent inference requires many chips wired together into clusters, and the networking hardware that connects them is a less crowded way to benefit from the same buildout. · Arista NetworksIt supplies the Ethernet switches that move traffic between thousands of accelerators in agent clusters, growing as data centers build out larger inference deployments., Astera LabsIt makes the connectivity chips linking accelerators and memory within servers, and benefits directly as cluster sizes expand and chip density increases., CoherentIt supplies optical interconnect components that replace copper as data center clusters exceed copper's reach, growing if clusters continue scaling upward., Credo TechnologyIt sells high-speed cables and connectivity chips that wire together racks of accelerators, tracking the pace of cluster buildouts directly.16%
- Every advanced AI chip must be manufactured and inspected at a small number of factories and with specialized equipment, creating a stable chokepoint regardless of which chip design wins. · Taiwan Semiconductor ManufacturingIt manufactures nearly every advanced chip in this take, from training to custom inference accelerators, so capacity utilization remains high regardless of which chip architecture gains share., ASML HoldingIt is the sole manufacturer of the lithography machines required for advanced chip production, so every new fabrication plant in a buildout phase must purchase from it., KLA CorporationIt provides inspection tools that maintain yields on advanced chips, and its business expands as fab capacity additions accelerate to meet buildout demand.15%
- Cash9%