Own the Model Layer
nik bonaddio@nik
−5.2%
since Sep 29
The idea
I believe the durable value in AI sits below the wrapper apps, in the model owners and the infrastructure they run on: chips and foundries, power, networking and data centers, and compute clouds. So this basket leans into the four large model owners, then builds down through silicon, energy, and data center layers, keeping no single name too large. Cash stays modest, mainly to add on dips, and the thesis holds as long as value keeps flowing down to this layer rather than the apps on top.
What it owns
- The buyable companies that own frontier models, either outright or through large stakes in the private labs. · Alphabet Inc.It owns a frontier model and the TPU chips it trains on, collecting revenue at multiple layers as long as Gemini stays competitive., Meta Platforms Inc.It builds Llama and deploys it across products with billions of users, capturing the model layer's value inside its own business., Microsoft CorporationIt holds a large stake in OpenAI and runs Azure where those models operate, collecting rent from both the model layer and the cloud beneath it., Amazon.com Inc.It owns a major stake in Anthropic and operates AWS, the largest cloud platform for AI workloads, combining model ownership with infrastructure rental.28%
- Every model, no matter who wins, is trained and run on this silicon, the purest shovel layer. · NVIDIA CorporationNearly every frontier model trains on its chips, making it the most direct supplier to the entire AI ecosystem., Taiwan Semiconductor ManufacturingIt manufactures advanced chips for every major AI chip designer, getting paid regardless of which designer wins market share., Broadcom Inc.It designs custom AI chips for major model owners seeking to reduce accelerator costs, plus the networking silicon connecting them together., Advanced Micro DevicesIt is the credible second source for AI accelerators, and cloud operators want it to keep pricing competitive against the leading supplier., ASML HoldingIt is the only company selling the EUV machines required to make advanced AI chips at leading edge, the deepest supplier in the stack.26%
- A decade of model training runs on electricity, and the grid is the slowest layer to build. · GE VernovaIt sells gas turbines and grid equipment being ordered to power new data centers, with a backlog stretching multiple years into the future., Constellation EnergyIt operates the largest nuclear fleet in the country, providing the always-on clean power that model owners are signing long-term contracts to lock up., Eaton CorporationIt makes the electrical distribution equipment and switchgear sitting between the grid and every rack of AI chips in data centers., Vistra CorpIt owns nuclear and gas plants in regions where data centers are being built fastest, so rising electricity demand flows directly to its prices.14%
- The buildings, switches, and cooling that AI compute physically lives in, a bottleneck the headline story skips. · Arista NetworksTraining frontier models requires wiring tens of thousands of chips together, and it supplies the networking gear that does it., Vertiv HoldingsAI chips run hot enough that data centers need its cooling and power equipment to function, and density keeps growing with each chip generation., Digital Realty TrustIt owns the physical buildings where AI clouds rent space, collecting literal rent as data center capacity stays scarce and pricing rises.13%
- Models are rented, not installed, so the companies selling raw AI compute and the memory it demands collect rent from every wrapper. · Micron TechnologyEvery AI accelerator ships wrapped in its high-bandwidth memory, and it is one of only three companies making this critical component., Oracle CorporationIt has signed massive contracts to rent compute infrastructure to model labs, converting its data center network into an AI cloud platform., CoreWeave Inc.It rents GPU compute exclusively to the model layer, making it the purest play on sustained AI training and inference demand.12%
- Cash7%