Local AI, Memory and Silicon
Jaap Vergote@jaapvergote
−13.2%
since Oct 3
The idea
I believe AI is moving out of the cloud and onto our own devices, updated like an OS rather than queried over the internet. To express that, this basket leans into edge silicon and the memory that local models need to run, with foundry and fab equipment underneath and a power-efficiency sleeve since low-watt inference is the whole point. I spread bets across each layer rather than picking single winners, keep a small cash buffer, and plan to hold for years unless it becomes clear inference is staying centralized in the cloud after all.
What it owns
- AI processors built directly into phones and PCs let models run locally instead of in the cloud, and these chip designers are racing to own those sockets. · QualcommIts neural processing units power most high-end Android phones, so it wins when on-device AI becomes a reason people upgrade their devices., AppleIt designs its own chips, memory architecture and operating system, making it the purest bet that inference moves onto custom silicon instead of cloud rental., Arm HoldingsNearly every phone and edge chip starts from its designs, so it collects royalties on the entire local AI buildout regardless of which chipmaker wins., Advanced Micro DevicesIt puts neural processing units into laptop and desktop chips, so it wins when the AI PC becomes a real market and people replace machines to run models locally.31%
- Running models on a device requires far more RAM and storage than cloud computing, so these companies benefit when every phone and PC ships with larger memory capacity. · Micron TechnologyIt is one of three DRAM makers globally, so it benefits when every phone and PC ships with much larger memory configurations to hold models in silicon., SanDiskLocal models live in flash storage between updates, so it benefits when model downloads push devices toward much larger NAND capacity., RambusIt designs interfaces that speed data movement between memory and processors, addressing the bandwidth bottleneck that slows local inference.20%
- Every edge chip and memory die gets made by a small number of manufacturers and equipment makers, so they win regardless of which chip design wins the race. · Taiwan SemiconductorIt manufactures nearly every chip in this basket, from Apple's to Qualcomm's to AMD's, so it wins whenever leading-edge edge silicon demand grows., Lam ResearchIt sells equipment to memory makers building DRAM and NAND, so it wins when devices need far more storage and RAM for local models., ASML HoldingIt is the only maker of lithography machines needed for the most advanced chips, so it wins whenever leading-edge edge silicon keeps advancing.20%
- Smaller companies built around low-power inference in cameras, sensors and embedded devices benefit when intelligence spreads beyond phones to billions of everyday objects. · NXP SemiconductorsIt builds small chips with embedded processors for cars, factories and appliances, so it wins when local AI spreads beyond phones to billions of overlooked devices., AmbarellaIt makes chips that run vision AI on cameras and in cars locally without the cloud, exactly matching the inference-moves-to-device world., Lattice SemiconductorIt sells tiny reprogrammable chips that run small AI tasks on very little power, enabling compute on cameras, laptops and industrial devices without battery drain.12%
- Running AI models on batteries makes power management and sensing silicon essential, so these companies win when every device needs to stretch its power budget. · Monolithic Power SystemsIt makes power management chips that stretch battery life, so it benefits when running models locally raises the power management content in each device., Analog DevicesIt makes sensing and low-power chips that feed data into local AI without draining the battery, so it wins when intelligence truly moves to the edge.9%
- Cash8%