−11.6%
since Oct 6
The idea
I believe AI is moving out of the cloud and onto the devices we already own, as open models get good enough to run on local chips for privacy, speed, and cost. To play that, I built a basket across four layers: the device makers who sell the hardware, the edge silicon that powers on-device inference, the memory and storage that local models lean on, and a small slice of the open-model publishers and the foundry behind nearly all these chips. I'm keeping 10% in cash to add on dips, and I'm treating this as a two-year view rather than something to hold forever.
What it owns
- Companies that sell end-user devices and systems benefit when AI inference moves onto the hardware customers own rather than relying on cloud services. · AppleIt runs AI inference on its own devices rather than in the cloud, leveraging its neural silicon and installed base of iPhones and Macs., NvidiaConsumer RTX cards are the standard hardware for running open models locally, making demand dependent on sustained home inference rather than data center sales alone., Dell TechnologiesIt sells the corporate laptops and workstations that enter refresh cycles when companies decide their fleets need AI capabilities on-device., HP Inc.As one of the largest PC sellers globally, it benefits when AI-driven refresh cycles lift unit volumes and device pricing., Super Micro ComputerCompact servers for on-premises AI workloads capture demand from businesses preferring local and edge inference over cloud services., Corsair GamingMemory, power supplies, and cases for DIY AI PC builds represent the home-enthusiast angle in local inference infrastructure.33%
- Chips designed to run AI workloads locally on phones, laptops, cameras, and industrial devices create value as processing shifts away from centralized data centers. · QualcommSnapdragon chips with dedicated AI processors are winning sockets in AI-enabled phones and Windows laptops, capturing the mobile edge shift., Advanced Micro DevicesRyzen AI processors with built-in NPUs compete in the AI PC market while consumer GPUs support local model inference., BroadcomConnectivity and networking chips enable edge devices to operate, while custom silicon design services capture demand from companies building proprietary AI processors., Arm HoldingsNearly every phone and growing share of laptops run on its architecture, so edge AI adoption generates royalties on power-efficient on-device inference., Lattice SemiconductorTiny, low-power programmable chips handle AI tasks in cameras, laptops, and industrial devices where full processors are inefficient., AmbarellaChips that run computer vision AI inside cameras and vehicles enable inference at the edge by design.27%
- Local AI models require more memory and storage per device than previous generations, creating sustained demand for DRAM, flash, and drive capacity as AI adoption spreads. · Micron TechnologyLocal AI models demand materially more DRAM than prior generations, and memory pricing staying firm as AI devices scale supports margins., Western DigitalModel files and local AI data live on its drives, creating storage capacity demand as AI devices and home setups expand., SandiskNAND flash in SSDs across AI PCs and phones benefits as devices carry larger, faster storage for models and inference data., Seagate TechnologyIts drives store data generated and consumed by local AI systems at the edge and in on-premises servers.19%
- Open-source AI models make local inference practical and economical, enabling the broader edge compute shift while creating royalty and manufacturing opportunities. · Meta PlatformsLlama open models enable capable AI to run on consumer hardware, making local inference practical while serving Meta's own infrastructure interests., Taiwan SemiconductorIt manufactures chips across the entire edge AI stack regardless of which designer wins, concentrating the take's manufacturing exposure in one foundry.11%
- Cash10%