Local AI Hardware Wave
Kevin Lee@kevinleeme
−11.2%
since Sep 30
The idea
I believe AI is going personal: prosumers will start running free local models on their own hardware instead of renting the cloud, and over the next 2-3 years that pulls spending toward the silicon inside those machines, especially memory bandwidth, which is the real bottleneck local models keep hitting. The basket leans into memory makers and edge chips, backed by the foundry and packaging layer behind them, plus a smaller slice of public plumbing companies in vector search, edge inference, and dev tooling.
What it owns
- Running AI models on personal devices instead of in data centers requires chips designed for low power and on-device inference, a shift that favors multiple silicon makers competing for sockets in phones, PCs, and embedded systems. · NVIDIAIts GPUs are the default hardware prosumers buy to run local models, since the free-model ecosystem is built around its CUDA platform., Advanced Micro DevicesIts APUs with unified memory let developers run large models on a desktop without a separate graphics card, offering the primary alternative to NVIDIA for prosumer AI hardware., QualcommIts mobile chips enable AI models to run directly on phones without cloud connectivity, positioning it as a beneficiary if smartphones become platforms for on-device AI agents., Arm HoldingsNearly every edge and mobile chip in this take is built on its instruction set architecture, so it collects royalties across the entire wave without backing a single chip winner., Lattice SemiconductorIt supplies small, power-efficient programmable chips for on-device AI in constrained devices like cameras and appliances that cannot accommodate full GPUs., AmbarellaIt provides vision AI chips that process camera and video data on the device itself, capturing the extension of local inference into edge vision applications.32%
- Local models demand fast memory bandwidth and storage capacity, so the companies that supply DRAM, storage, and the controllers connecting them benefit regardless of which chip design wins. · Micron TechnologyIt supplies the high-bandwidth DRAM and memory modules that become the constraint when prosumers build machines for local model inference., SanDiskIt supplies the flash storage where multi-gigabyte model files are stored on personal machines, growing as local AI adoption increases storage demands in prosumer boxes., RambusIt licenses the interface standards that move data between processors and memory at speed, capturing royalties whenever bandwidth becomes the limiting factor in local AI hardware., Western DigitalIt supplies the large hard drives where prosumers store model checkpoints, datasets, and training data locally, benefiting from growth in personal AI hardware ownership., Silicon MotionIt supplies the controllers that manage high-speed SSDs in local AI machines, benefiting as prosumers invest in faster storage for model inference.27%
- Software tools that enable local models to run efficiently, including edge inference, search capabilities, and developer frameworks for building AI agents. · CloudflareIt runs inference engines at the network edge close to users, capturing the shift of AI workloads away from centralized cloud data centers toward distributed compute., ElasticIt provides search and vector database technology that developers run on their own hardware for AI applications, capturing the self-hosted alternative to cloud-based retrieval., MongoDBIt provides a database platform developers can deploy anywhere, including on personal hardware, to store state and memory for AI agents running locally., GitLabIt supplies development tools with self-hosted options for the developer-heavy prosumer audience building and deploying local AI applications.17%
- Every chip in this take must be manufactured and packaged by a small number of specialists, so foundries and advanced packaging houses win volume growth regardless of which chip designer the market picks. · Taiwan SemiconductorIt manufactures nearly every advanced chip in this take, so it benefits from growth in local AI hardware regardless of which chip designer wins adoption., Amkor TechnologyIt packages advanced AI chips with memory stacked close to processors, benefiting as high-bandwidth memory packaging becomes essential for local model performance., ASE TechnologyAs the world's largest chip packager, it processes all advanced chip designs at scale, so it benefits from rising volumes in edge and local AI silicon regardless of which designers succeed.14%
- Cash10%