AI Power Value Chain
Masheed Ahadi@masheed
−6.7%
since Oct 6
The idea
I believe AI is fundamentally an energy trade, so the money flows to whoever powers and houses the compute, not just whoever builds the models. This basket spans the whole chain: nuclear-heavy power producers and utilities, the grid and cooling equipment behind them, chips, and data center landlords, with a small 7% bucket for the riskier neoclouds and early-stage nuclear names so they can't dominate the mix. I'm keeping cash in reserve and holding for years, and I'd rethink the thesis if those long-term power deals between utilities and hyperscalers started falling apart.
What it owns
- Data centers operate around the clock and require firm baseload power, creating decade-long supply contracts with the nuclear and thermal plants that can reliably fuel them. · Constellation EnergyIt operates the largest US nuclear fleet and is signing long-term power deals with hyperscalers, earning from both the electron shortage and the firm pricing that comes with it., VistraIt owns both gas and nuclear plants in the regions where data centers are landing, capturing margin from the same electron shortage that drives power prices up., Talen EnergyIt operates nuclear plants selling power directly to AI campuses through long-term contracts, capturing margin from the reliability premium buyers pay for firm baseload., CamecoIt supplies uranium fuel to the nuclear plants that power data centers, gaining from utilities extending reactor runtime and contracting forward for fuel., NextEra EnergyIt is the largest builder of new renewable generation that hyperscalers can actually add to their power mix, growing its project backlog as data center demand accelerates.27%
- Transformers, switchgear, turbines and cooling systems form the physical bottleneck in every data center build, with multi-year backlogs and pricing power tilted toward suppliers. · GE VernovaIt builds the gas turbines and grid equipment that are sold out for years, exactly the physical bottleneck that gets paid first as data centers race to add capacity., VertivIt sells the cooling and power systems that AI compute racks depend on, with demand rising alongside denser chip generations and higher power densities., EatonIts switchgear and power distribution equipment is installed before any chip turns on inside a data center, keeping it in tight supply as construction accelerates., Quanta ServicesIt builds the transmission lines and substations that physically connect data centers to their power sources, winning from every dollar of grid expansion spending., nVent ElectricIt produces the enclosures and liquid cooling hardware that go into every data center rack, a direct exposure to the same cooling bottleneck driving faster growth.25%
- The silicon layer captures the first and largest share of AI capital spending, making it the most direct beneficiary of scaling data center buildouts. · NVIDIAIt supplies the chips that capture the first and largest share of every hyperscaler's AI capacity spending, making it the default silicon choice as long as capital spending keeps growing., BroadcomIt designs the custom chips and networking silicon that hyperscalers use to reduce their reliance on any single supplier, winning either way inside the chip layer as custom orders grow., Taiwan SemiconductorIt manufactures nearly every advanced AI chip, making it the single narrowest point in the supply chain, with advanced capacity that stays sold out and pricing power that holds., Micron TechnologyIt makes high-bandwidth memory, the fast memory stacked next to every AI processor, and remains one of only three suppliers for the most demanding applications.23%
- Real estate owners of data center capacity collect rent regardless of which AI model or chip architecture wins, capturing value from scarcity and utilization. · EquinixIt owns the physical data center buildings and collects rent from whoever wins the AI race, capturing value from scarcity as long as space stays tight and rents keep rising., Digital RealtyIt leases the wholesale data center shells that hyperscalers are racing to fill, profiting from the gap between leasing demand and new supply coming online.11%
- Leveraged neoclouds and pre-revenue nuclear companies carry outsized upside and downside, so the group starts at a small allocation to manage risk and leave room for drift. · CoreWeaveIt rents GPU capacity faster than most suppliers can build it, leveraging debt to fund growth, and depends on keeping its contracts full and its financing serviceable., IRENIt pairs cheap renewable power with AI compute capacity, executing the speculative version of the idea that electrons get paid first as data center demand soars., Cipher MiningIt holds grid connections and power sites that a bitcoin mining operation secured and that AI tenants now want to lease, betting on converting those sites into paying AI hosting., TeraWulfIt operates compute on zero-carbon power and is signing AI hosting deals on the back of nuclear and hydroelectric capacity, with the bet that those deals remain full and self-funding., OkloIt designs small reactors meant to sit next to data centers and depends on years of regulatory approval and first commercial deployments, making it the longest-duration bet in the group.6%
- Cash8%