The Whole AI Stack
Tom White@twhite
−12.8%
since Oct 1
The idea
I believe cheaper AI means people use far more of it, not less, so demand for compute and memory keeps expanding for a decade rather than a single cycle. This basket spreads across the whole stack: chips and cloud compute, the equipment that makes those chips, memory and bandwidth, networking, power and cooling, and the data centers housing it all. I favor companies that stay essential regardless of which model or chip wins, and I keep a small cash reserve to add on weakness.
What it owns
- The chips doing the AI work and the clouds renting them out are where the demand explosion lands first. · NVIDIA CorporationIt sells the compute everyone reaches for first when cheaper AI means vastly more AI use, holding as long as data center buildout accelerates., Taiwan Semiconductor ManufacturingWhoever wins the chip race, it makes their chips, staying intact as long as advanced AI chips need its leading-edge factories., BroadcomIt builds custom AI chips for big clouds cutting costs, and custom silicon orders keep growing alongside standard chips if usage explodes., Advanced Micro DevicesIt is the only credible second source for AI compute in a decade-long buildout too big for one supplier., Nebius GroupIt rents out AI compute directly, betting that demand for rented GPU capacity outruns what big clouds can supply.28%
- Every new chip factory buys from the same handful of tool makers, whoever wins the chip race. · ASML HoldingEvery advanced chip gets made on its machines and nobody else can build them, selling more tools every year of a decade-long buildout., Applied MaterialsIt sells the widest range of chipmaking tools, so it wins whenever anyone, anywhere builds more chip capacity., KLA CorporationAs chips get harder to make, more of the factory budget goes to its inspection machines, growing with chipmaking complexity., Lam ResearchIt makes the tools memory factories need most, so it wins when rivals expand to feed AI demand.18%
- Data centers are hitting the limits of the electric grid, which makes power equipment a bottleneck. · Vertiv HoldingsIt sells the liquid cooling and power gear that replaces air cooling when AI chips run too hot, exploding with compute demand., Eaton CorporationEvery new data center needs its electrical gear to get power from the grid to the racks, growing faster than it can ship., GE VernovaIt sells the turbines and grid equipment for new generation and capacity that a decade of AI buildout requires., Constellation EnergyIt operates the biggest US nuclear fleet for reliable, carbon-free data center power, growing as clouds sign long-term deals for it.16%
- Thousands of chips only act like one big computer if the connections between them keep up. · Arista NetworksIt sells the switches that connect thousands of chips together fast, holding as long as AI clusters keep getting bigger., Marvell TechnologyIt sells the chips that move data inside and between data centers plus custom AI silicon, riding more data in motion from AI use., Astera LabsIt makes the connectivity chips that let AI processors and memory talk at full speed inside a server, riding purely on AI server numbers., Credo TechnologyIt sells the high-speed cables and chips that wire AI racks together, holding as long as cluster sizes keep growing.15%
- AI chips are starved for fast memory, and only a few companies make it. · Micron TechnologyIt is the US maker of the high-bandwidth memory AI chips need, growing as long as AI servers demand more memory per chip., RambusIt makes the small chips that let memory keep up with faster processors, getting paid on nearly every memory stick as AI demand climbs., SandiskMore AI use means more data stored and moved, and it sells the storage growing with AI data center volumes.12%
- Someone owns the buildings all this hardware sits in and collects rent on the boom. · EquinixIt owns the buildings where clouds and companies connect, collecting rent as long as data center space demand keeps outrunning supply., Digital RealtyIt leases big blocks of data center space to clouds building AI capacity, with rents and occupancy climbing through a decade-long buildout.6%
- Cash5%