−10.7%
since Oct 3
The idea
I believe AI compute demand keeps compounding over the next couple years, and the real bottlenecks aren't the famous chip names but the choke points around them. So this owns the whole chain: power generation and grid gear, wafer equipment, advanced packaging, data center plumbing, and the construction crews who actually build these sites, rather than betting on one winner. Power gets the heaviest weight since those backlogs run years out. I'm holding cash to add on rough days and built this concentrated for a two-year window, ready to rethink it if the buildout story breaks.
What it owns
- Every new data center needs power the grid can't yet deliver, and turbines, transformers and switchgear have multi-year backlogs. · GE VernovaGas turbine order books stretch years out because data centers need generation faster than anyone can build it., EatonEaton sells transformers, switchgear and electrical gear every data center and grid upgrade needs, with lead times stretched and orders growing., Quanta ServicesQuanta is the crews that actually string lines and build substations, capturing value from grid buildout spending and scarce skilled labor., VistraVistra owns plants that sell power into the shortage, benefiting when data center demand keeps power prices and contract terms favorable., Constellation EnergyConstellation runs the biggest nuclear fleet in the country, the steadiest clean power available to hyperscalers signing long deals., Talen EnergyTalen sells nuclear power directly to AI data centers, the purest version of a power-scarcity bet dependent on direct deals staying intact., Powell IndustriesIt manufactures switchgear that utilities and data centers need to upgrade power distribution systems, with order backlogs extending visibility into future revenue.25%
- Chips can't exist without fab capacity and the equipment that builds it, and packaging is the single tightest step in the chain. · Taiwan SemiconductorNearly every AI chip gets made at TSMC, so if compute demand compounds, orders flow through here regardless of which designer wins., ASMLASML is the only company that makes lithography machines for advanced chips, with no substitute as foundries expand leading-edge capacity., Lam ResearchLam's etch and deposition tools are required for every new fab and especially for stacking memory, riding the capacity buildout., Applied MaterialsApplied Materials sells the widest range of chipmaking tools, capturing the whole fab buildout rather than one step., KLAKLA's inspection tools catch defects and advanced AI chips need more inspection than anything before them, keeping KLA's share of fab spending growing., Amkor TechnologyAdvanced packaging is the tightest spot in the semiconductor chain and Amkor is the biggest outsourced way to own it, with capacity sold out.23%
- The chip designers selling the compute everyone is racing to buy. · NVIDIANVIDIA remains the default chip for AI, and in a market that isn't winner-take-all, it just needs to keep growing with overall demand., BroadcomBroadcom builds custom AI chips hyperscalers design to lessen NVIDIA dependence, which fits a market where multiple suppliers gain share., Advanced Micro DevicesAMD is the one real alternative to NVIDIA with chips buyers actually deploy, and in a non-winner-take-all market, even modest share gains work.15%
- Networking, cooling and interconnects that every data center needs no matter whose chips sit inside. · Arista NetworksArista sells the networking that ties thousands of AI chips into one cluster, growing as clusters get bigger., VertivVertiv makes cooling and power distribution gear inside data centers, demand that scales directly with every megawatt of new power coming online., Astera LabsAstera's connectivity chips move data between AI processors inside a server, a necessary piece of every big AI system., Credo TechnologyCredo's high-speed cables connect gear inside AI data centers, a plumbing piece that grows with every new cluster., CoherentCoherent makes the optical parts that carry data between racks, and bigger AI clusters need exponentially more of them.12%
- Skilled electrical and mechanical labor is the binding constraint on actually getting data centers built. · Comfort Systems USAComfort Systems installs the mechanical and cooling systems data centers can't open without, the other half of the construction labor squeeze., EMCOR GroupEMCOR is the largest electrical contractor doing data center work, capturing value from the binding constraint of skilled electrical labor., CumminsEvery data center needs backup generators and Cummins is the standard supplier, riding the construction wave of data center buildouts.8%
- AI servers need several times the memory of ordinary ones, and supply is tight. · Micron TechnologyAI chips need high-bandwidth memory and Micron is one of three companies that make it, with supply staying sold out ahead of demand., Western DigitalTraining data and AI output have to live somewhere, and Western Digital's storage rides that wave as storage demand stays firm.7%
- The hyperscalers whose capital spending is the demand this whole take rides on. · MicrosoftMicrosoft's AI infrastructure spending underwrites orders across power, chips and construction, making it proof that compute demand keeps compounding., AlphabetAlphabet's capital spending plans are the clearest evidence that AI compute demand keeps compounding, serving as proof for the take's premise.4%
- Cash8%