On-Prem AI Buildout
Austin@nfteague
−9.6%
since Oct 14
Practiced to +1.4% over 14 days before going real.
Practice chapter: simulated results; no money was invested.
The idea
I believe enterprise AI ends up hybrid: custom models fine-tuned on internal data running on premise, local AI on employee devices, and cloud reserved for heavy compute, while the market still prices inference as a pure cloud story. That leaves the cloud leg overbought and the on-prem and on-device pieces underpriced. This basket owns the on-prem hardware sellers, the memory and networking bottlenecks that get paid regardless of who wins, on-device chipmakers, and the software enablers for fine-tuning on internal IP, deliberately skipping hyperscalers.
What it owns
- Server and storage vendors selling systems enterprises buy to run custom models inside their own walls rather than renting cloud compute. · Dell TechnologiesThe largest seller of servers enterprises buy to run custom fine-tuned models on-premise instead of renting cloud capacity., Hewlett Packard EnterpriseIt sells on-prem AI systems and private-cloud setups aimed at enterprises that want models trained on internal IP inside their own walls., IBMIts watsonx platform is built for fine-tuning on a company's own data and running models wherever the company chooses, including on-premise., NetAppFine-tuning on internal data requires fast local storage, and it sells that storage systems enterprises need for on-prem AI work., Super Micro ComputerIt assembles AI servers quickly and in custom configurations that suit enterprises building their own on-prem clusters.26%
- Suppliers of memory, networking, and connectivity that on-prem AI systems need regardless of which hardware vendor wins, capturing value from the underlying scarcity. · Micron TechnologyEvery AI box needs substantial memory, and it supplies that to all on-prem and on-device systems regardless of which hardware vendor wins., Arista NetworksIt sells networking gear that wires AI servers into clusters, which enterprises building on-prem setups need just as much as cloud providers do., BroadcomIt makes the switching chips inside nearly every AI network cluster, capturing value from on-prem buildouts regardless of whose logo is on the server., Marvell TechnologyIt designs custom AI chips and connectivity around them, capturing value on the component bottleneck regardless of who assembles the server.23%
- Chipmakers that put AI directly onto employee laptops and phones, enabling local inference without cloud dependency. · QualcommIts processors run AI locally on phones and laptops, delivering the on-device inference leg of this take without touching the cloud., Advanced Micro DevicesIts processors go into AI PCs that put local AI on employee desks, and it is deepening the story through AI model acquisitions., Arm HoldingsIts chip designs sit inside most processors that would run AI on devices, collecting royalties on the on-device leg almost regardless of which chipmaker wins., IntelIt offers an unloved way to own the AI PC upgrade cycle as corporate laptop refreshes tilt toward AI-capable machines.22%
- Software platforms that make fine-tuning models on internal data and running them on-premise actually practical for enterprises. · Palantir TechnologiesIts software lets companies deploy AI on top of their own internal data, often inside their own walls, which is the enabler layer of this take., NutanixIt provides the software layer that makes managing private infrastructure as straightforward as cloud platforms, critical for enterprises moving AI workloads on premise., MongoDBIts vector search capabilities let companies use their internal data in AI systems and it can run self-hosted, not just in the cloud., ElasticIt provides search and retrieval over internal documents that companies can self-host, which is the plumbing custom fine-tuned models need to be useful.19%
- Cash10%