Agentic AI Supply Chain
Charlie Madden@charliemadden
−6.5%
since Oct 3
The idea
I think consumer agentic AI is the iPhone moment, but the agents themselves will compete away their own margins chasing users. So I skip the agent makers and own what they all have to pay: inference compute first since agents burn tokens constantly, transaction rails for when agents start buying on our behalf, and power and data center capacity as the slower, steadier layer underneath it all. Compute gets the largest weight since it gets paid first and most, with a small cash buffer held back. I'd rethink this if agents stop needing constant inference or if the rails get disintermediated.
What it owns
- Agents running continuously throughout the day consume vastly more compute than traditional chatbots, and every agent company must pay for this infrastructure layer regardless of which competitor ultimately prevails. · NVIDIA CorporationIt supplies the chips that continuous agent workloads run on, and remains the dominant platform for inference compute as agent deployments scale., Taiwan Semiconductor ManufacturingIt manufactures the leading-edge chips for nearly every major AI accelerator, collecting revenue across the entire compute layer regardless of which chip designer gains share., BroadcomIt designs the custom silicon that data center operators commission to reduce their largest infrastructure costs, and wins contracts as buyers move away from general-purpose chips., Micron TechnologyIt manufactures high-bandwidth memory that every inference accelerator requires, and benefits as continuous agent workloads drive memory demand., Marvell TechnologyIt builds the interconnect silicon that moves data between accelerators in data centers, and gets deployed as always-on inference creates higher demand for chip-to-chip communication., Arm HoldingsIt collects royalties on every chip design that incorporates its instruction set, and benefits if agents migrate toward on-device execution on mobile and edge hardware., Astera LabsIt makes the connectivity components between accelerators in data center racks, and gets deployed as server builds accelerate and density increases.45%
- When agents make purchases on behalf of users, the payment and commerce infrastructure collects a fee on every transaction that flows through the system. · ShopifyIt provides the merchant infrastructure that agents can transact through, positioning it as the broadest way for agents to access storefronts at scale., VisaIt operates one of the two primary payment networks that agent transactions settle through, collecting a toll on commerce flowing through the system., MastercardIt operates the other primary payment network that agent transactions settle through, collecting settlement fees on commerce regardless of which checkout system agents use., CloudflareIt sits in front of large portions of the internet deciding which agent traffic reaches sites, and benefits if its tools for metering and charging machine traffic gain adoption.24%
- Running inference at scale requires electricity and physical facilities that take years to build, making them the slowest and most constrained part of the supply chain. · Vertiv HoldingsIt supplies cooling and power infrastructure that data centers require when running inference continuously, and sells to every facility buildout regardless of the accelerators inside., Constellation EnergyIt generates clean baseload power through nuclear generation and supplies long-term contracts to data center operators, benefiting as always-on inference demand outpaces grid capacity., GE VernovaIt manufactures turbines and grid equipment that transmission systems need to accommodate the power consumption from inference buildouts., Digital Realty TrustIt owns the data center buildings where inference infrastructure runs, and collects rent as tight vacancy rates and rising demand fill capacity.21%
- Cash10%