Agent Compute Stack
sunny@phunnysunny
−7.8%
since Oct 2
Practiced to −0.3% over 1 day before going real.
Practice chapter: simulated results; no money was invested.
The idea
I believe AI agents running nonstop will turn compute into something bought like electricity, with demand hitting the whole stack over the next year or two. So this owns the chips doing the inference work, the clouds renting capacity, the power and electrical gear feeding data centers, and the networking moving agent traffic between them, picking one highest-conviction name per layer rather than spreading thin. I keep about 10% in cash to add when this volatile group sells off, and I'd rethink the idea if agent demand fails to show up broadly across these layers within that window.
What it owns
- Inference work from always-on agents lands first on the processors that execute it, making this where new demand for computing power shows up most directly. · NVIDIAIt manufactures the inference processors where most agent workloads run, holding share as long as cloud operators treat its chips as the default compute platform., BroadcomIt builds the custom inference chips that major cloud companies design to reduce computing costs, winning regardless of which silicon brand ends up in the data center., Taiwan SemiconductorNearly every processor in this take gets manufactured by it, so it captures payment on all compute demand across the entire stack if agent workloads grow as expected., Advanced Micro DevicesIt sells both processors and graphics chips into data centers, winning as cloud operators seek alternative sources for the processors powering agent workloads., Arm HoldingsAmazon, Microsoft and Google all build their server processors on its chip designs and pay it royalties on each one shipped as clouds shift their CPU fleets toward its architecture.30%
- Agents run on rented compute infrastructure, so the companies operating those data centers collect recurring revenue from the capacity agents consume. · MicrosoftIt rents compute capacity to agents through Azure and sells agent software directly through Copilot, sitting on both sides of growing AI infrastructure demand., AlphabetIt rents AI capacity from Google Cloud while running agents on its own processors, capturing cloud demand without paying full cost for external silicon., AmazonAWS is the largest cloud operator, so it captures the bulk of compute spending if agents make processing power bought like a utility service., OracleIt has signed large contracts to build and rent AI compute infrastructure, making it a direct bet on those committed capacity orders turning into actual revenue., CoreWeaveIt rents nothing but AI compute capacity, making it a direct bet on agents driving demand for processing power sold like a utility service.26%
- Data centers full of constantly active agents draw far more electricity and require far more electrical infrastructure than the existing grid was designed to handle. · GE VernovaIt sells the power generation and grid equipment that data centers must install before agent infrastructure can run, with a multi-year order backlog showing sustained buildout demand., EatonIt makes the electrical equipment inside data centers that distributes and controls power, with its orders growing as facility buildouts continue., VistraIt sells electricity directly, and data centers running constantly-active agents draw power around the clock, driving up consumption and pricing., Constellation EnergyIt operates nuclear plants that cloud companies are signing long-term power contracts with because agents need electricity that runs without interruption.17%
- Agents communicating with models, tools and each other create traffic flowing through switches, optical connections and server interconnects. · Arista NetworksIt makes the switches that move agent traffic inside AI data centers, with major cloud providers as its largest customers as networking demand for AI grows., Marvell TechnologyIt makes the chips connecting AI compute together and builds custom silicon for major cloud operators, growing with their interconnect and custom processor orders., Astera LabsIt sells the connectivity chips that let processors and memory communicate within AI servers, growing with every new computing rack that gets built., CoherentIt makes optical components that carry agent traffic as light between chips and across data centers, growing as AI facility buildouts drive faster link demand., Credo TechnologyIt sells high-speed cables and chips linking AI servers together, mostly to major cloud operators who need to keep expanding their agent computing capacity.17%
- This bucket contains positions that provide supporting infrastructure and services across the agent compute ecosystem. · Memory constraints have emerged as a genuine bottleneck in AI infrastructure buildouts, and this company supplies the memory chips that enable the compute stack to function at scale.3%
- Cash7%