Commodity Models, Scarce Inputs
Amplify@amplify
−0.4%
since Sep 28
The idea
I think the model layer itself is becoming commoditized as open source closes the gap with the frontier labs, so the real value shifts to whoever controls the scarce inputs around it. The basket splits evenly across three legs: data owners with real licensing deals, hardware that wins on inference volume regardless of which model wins, and the power and grid buildout that's the slowest piece to scale. I'm keeping a small cash reserve to add on pullbacks, since I expect this repricing to play out over the next year or two, not overnight.
What it owns
- Electricity generation and grid infrastructure are the slowest supply chains to expand, so rising compute demand will outpace supply growth and raise returns for power owners and equipment makers. · Constellation EnergyIt operates the largest nuclear fleet in the US and data centers are signing long-term contracts to buy its power directly for compute workloads., VistraIt owns nuclear and gas plants in regions where data centers are growing fastest, so it gets paid for the electricity that compute demand requires., EatonIt makes the electrical equipment and power management that every data center and grid upgrade requires, indifferent to which model or chip wins., GE VernovaIt sells turbines and grid equipment that represent the slowest part of the power infrastructure buildout to scale, creating a backlog that stretches years., First SolarIt is the largest US solar panel maker and benefits from domestic manufacturing protections that keep competing imports out while power demand rises., NextEra EnergyIt builds renewables at scale in the US and operates a regulated utility in fast-growing Florida, making it a steady way to own rising electricity demand from compute., VertivIt sells the power and cooling gear inside every data center, so it gets paid per buildout no matter whose model runs there.30%
- If frontier models become commodities, the companies that own proprietary data no one else can legally access become essential inputs that all builders must license. · RedditIt owns two decades of human conversation that cannot be scraped legally, and model builders are already paying for licensing deals to that archive., S&P GlobalIt sells financial data and ratings that companies must license no matter which AI model wins, exactly the kind of proprietary inputs that command pricing power., RELXIt owns legal and scientific archives built over decades that no free open-source scrape can match, creating a durable moat around proprietary data., Thomson ReutersIt owns Westlaw's legal corpus, proprietary training data that lawyers and AI companies pay to license, placing it in the data franchises that persist when models commoditize., News CorpIt was among the first publishers to get paid by frontier labs for news content, establishing that training data commands licensing revenue.30%
- Serving models at scale requires chips, memory, and networking regardless of which model runs, so these infrastructure layers benefit from compute demand without depending on any single winner. · Taiwan SemiconductorAlmost every serious AI chip gets manufactured at its foundries, making it infrastructure that wins regardless of which model dominates the market., Micron TechnologyServing models at volume requires high-bandwidth memory, and it manufactures the kind inference workloads depend on at scale., BroadcomIt designs custom chips that large cloud companies use to run inference cheaply, capturing the volume layer below the frontier model names., Arista NetworksIt sells networking gear that ties AI data centers together, and it gets ordered whether the models running inside are open or closed source., Marvell TechnologyIt designs custom chips and connectivity for cloud companies building their own inference hardware, capturing the layers below the model., Credo TechnologyIt makes the cables and chips that move data between processors inside AI servers, small components every buildout needs regardless of whose model runs on top.30%
- Cash10%