Small AI Goes On-Device
camille@camille
−13.6%
since Oct 7
The idea
I believe a growing share of everyday AI shifts from the data center onto phones and PCs over the next couple of years, driven by inference cost, privacy, and latency. To play that, I own the chip designers building bigger NPUs, the memory makers selling more DRAM and NAND per device, the IP licensors collecting royalties on AI-capable chips, a few edge-AI specialists, and the main foundry behind nearly all of it. I'm avoiding single-name concentration and keeping 11% cash to add on dips over the two-year window.
What it owns
- Companies that design NPU-equipped chips for AI phones and PCs capture the upgrade cycle directly as everyday tasks move from cloud to device. · QualcommSnapdragon chips are the primary way AI runs on Android phones, so if everyday AI tasks keep moving on-device, Snapdragon adoption and pricing per phone both rise., Advanced Micro DevicesRyzen AI chips put an NPU in every laptop, capturing the higher-spec PC cycle as AI workloads move from cloud to local devices., AppleApple designs its own silicon to run more AI on-device across a billion iPhones, capturing upgrade demand directly if Apple Intelligence features drive replacement cycles., IntelIntel ships most PC processors and Core Ultra embeds its AI bet, so it benefits if the refresh cycle materializes but weakens fastest if the AI PC cycle stalls., NXP SemiconductorsNXP puts low-power AI into cars, factories and smart devices, winning as inference spreads beyond phones and PCs into embedded applications.30%
- Running models locally on devices requires more memory per unit, so DRAM and storage content per phone and PC rises even if total unit sales stay flat. · Micron TechnologyRunning models locally requires more DRAM per device, so AI phones and PCs ship with higher memory content and Micron sells more DRAM per unit., SandiskModels and their data live on the device, requiring more NAND flash per phone and laptop, making storage content per unit the core bet., Western DigitalWestern Digital sells storage that holds on-device models and their outputs, earning on higher storage capacity per device as local AI spreads., Silicon MotionSilicon Motion makes the controller chips inside nearly every NAND device, earning on storage volume regardless of which memory maker supplies the chips.21%
- These companies collect royalties or license fees on nearly every AI-capable chip, earning regardless of which chipmaker wins market share. · Arm HoldingsArm collects royalties on nearly every phone chip and growing share of PC chips, with AI-capable designs commanding higher rates per license., RambusRambus licenses technology that moves data between chips and memory fast, earning more as devices carry larger memory for local AI workloads., SynopsysEvery company designing an NPU or AI chip uses Synopsys tools and IP blocks, so more on-device AI projects mean more Synopsys revenue regardless of chip winner., Ceva IncCeva licenses NPU designs to chipmakers who don't build their own, making it the most direct small-cap bet on AI spreading across many edge devices.20%
- Smaller companies whose core business is low-power inference in cameras, wearables, cars and embedded devices win as AI moves to the edge for speed and battery life. · AmbarellaAmbarella's entire business is low-power chips that run vision AI on device in cameras, cars and robots, with no data-center business to offset edge demand., Lattice SemiconductorLattice makes tiny, low-power programmable chips that handle AI tasks at the edge where larger processors would drain batteries too fast., SynapticsSynaptics builds chips that connect sensors, touch and audio to on-device AI in phones, PCs and smart gadgets, winning if higher-spec devices need smarter interfaces.11%
- The foundry that manufactures most advanced chips benefits from higher volumes and more complex nodes across the entire on-device AI ecosystem. · Taiwan SemiconductorTSMC manufactures the advanced chips for nearly every other company in this take, winning on higher volumes and more advanced nodes across the ecosystem.7%
- Cash11%