Tokens Over Barrels
−7.3%
since Oct 6
The idea
I believe AI inference is on track to become the largest commodity market in the world, eventually rivaling and passing crude oil. Commodities reward the lowest-cost producer and the owners of scarce inputs, so I weight this most heavily toward the hyperscalers and neoclouds selling inference, then spread the rest across chips and fabs, power and grid, and data centers and land. I keep a slice in cash and avoid betting too heavily on any single name, since this is a multi-year thesis I expect to play out slowly.
What it owns
- The companies actually selling tokens at scale; if inference becomes the biggest commodity, they are its producers. · AlphabetIt makes its own TPU chips to produce tokens cheaper than rivals paying NVIDIA prices, rewarding the lowest-cost producer in a commodity market., MicrosoftAzure sells inference at a scale almost nobody else can match, making it one of the biggest sellers if tokens pass barrels in market size., AmazonAWS moves more compute than anyone, making it a top seller in any world where inference becomes the biggest commodity., OracleIt has signed huge multi-year contracts to supply AI compute, locking in inference revenue before the market gets there., CoreWeaveIt sells nothing but AI compute, making it the most direct way to own the inference-seller side of this take., Nebius GroupIt is the smaller neocloud trying to win on price and speed of buildout, exactly how new low-cost producers emerge in commodities.37%
- Compute is the crude that gets refined into inference, and a handful of companies control its supply. · NVIDIANearly every token sold today runs through its chips, so it collects a toll on the whole inference market as it grows toward oil-sized scale., Taiwan SemiconductorAlmost every serious AI chip gets made at its fabs, making it the scarcest input in the inference supply chain., BroadcomIt builds custom inference chips that let hyperscalers make tokens cheaper than off-the-shelf GPUs allow., Advanced Micro DevicesIt is the credible second source for inference chips in a commodity market that won't tolerate a single supplier forever., Micron TechnologyServing tokens requires enormous amounts of high-bandwidth memory, and it is one of only three companies on earth that make it.25%
- Every token costs electricity, and generation and grid equipment are the hardest inputs to add quickly. · Constellation EnergyIt owns nuclear plants that can run data centers around the clock, supplying the clean always-on electricity tokens require., VistraIt sells power directly into data center buildout, and inference at oil scale means electricity demand the grid wasn't planning for., GE VernovaIt builds the turbines and grid equipment that handle the new generation and connections all this inference runs on., EatonEvery new data center needs its switchgear and power distribution gear before a single token gets served., VertivIt sells the power and cooling systems that keep inference chips alive, including the liquid cooling newer chips require.18%
- Someone owns the buildings the commodity is produced in, and they collect rent whoever wins. · EquinixIf inference becomes the biggest commodity market, the buildings it lives in become valuable real estate, and it is the biggest landlord., Digital RealtyIt owns the big-footprint data centers hyperscalers rent when they can't build fast enough themselves., Iron MountainIt is converting land and buildings it already owns into data centers, the cheapest way into the scarce-input side of this take.10%
- Cash10%