Physical AI Supply Squeeze
Alex Mathews@ajmat
−12.1%
since Oct 6
The idea
I believe humanoid robots and physical AI are about to become massive new customers for compute, arriving into a supply chain that simply cannot scale fast enough. So instead of betting on one robot maker, I own the bottlenecks everyone will have to pay: the foundries and chipmaking tools, memory and networking, the power and grid buildout, and the companies building robotic bodies and sensors. It is spread across these layers on purpose, with a small cash buffer, because this is a decade-long buildout, not a quick trade.
What it owns
- The chips every robot and AI model runs on face demand that already outruns supply. · NVIDIAIt sells the compute brains that humanoid robots and physical AI systems run on, from training to robot platforms, as the default choice for both., Taiwan SemiconductorIt makes nearly every advanced AI and robot chip at the only fabs that can do leading-edge production at scale, so fab capacity limits the entire supply chain., BroadcomIt builds the custom AI chips that largest buyers order when standard compute supply runs short, which is exactly what a squeeze produces., Advanced Micro DevicesIt is the clearest alternative to the default compute choice, so buyers take its AI accelerators when supply for the standard option runs short.26%
- The machines that make chips are the hardest layer to expand, so scarcity shows up here first. · ASML HoldingIt is the only company that makes the machines needed to print the most advanced chips, so a decade-long compute buildout runs through its order book., Applied MaterialsIt sells the widest range of chipmaking equipment, so it gets paid whenever anyone builds or upgrades a fab to ease the squeeze., Lam ResearchIts etch and deposition tools are required steps in making both logic and memory chips, two layers that cannot grow fast enough., KLA CorporationIts inspection tools help fabs squeeze more good chips from finite capacity, which matters most when every wafer is already spoken for.20%
- Every new data center and robot factory needs electricity the grid does not have yet. · EatonIt makes the electrical equipment that connects data centers and factories to the grid, a finite layer where lead times are already long., GE VernovaIt sells the turbines and grid equipment that take years to build, which is the hardest layer to add when compute demand is not going away., Vertiv HoldingsIt builds the power and cooling gear inside AI data centers, so every new building in the buildout becomes its customer., Constellation EnergyIt owns the largest US nuclear fleet, the cleanest way to supply the round-the-clock power that AI compute needs at scale.19%
- The builders, bodies and senses of humanoid machines represent the physical side of the bet. · TeslaIts Optimus humanoid robot is the most visible entrant with its own factories to build and deploy them, making it a direct bet on the race itself., TeradyneIt gets paid twice in this take: its robot arms are sold into automation today, and its chip-testing machines are needed for every AI chip made., Rockwell AutomationIt runs the factory automation systems where humanoid robots will get deployed first, so it gets paid as factories automate regardless of which robot wins., ON SemiconductorIt makes the image sensors and power chips that give robots eyes and motion, the sensing layer needed for physical AI to become real.15%
- AI compute is useless without fast memory and networks, and both are in short supply. · Micron TechnologyIt is one of only three companies making the high-bandwidth memory that every robot brain needs next to compute, a finite layer paid regardless of the winner., Arista NetworksIt builds the networks that wire thousands of chips together for training robot brains, the infrastructure for scaling AI compute., Marvell TechnologyIt makes the chips that move data between AI processors, a layer that gets paid whenever compute scales, whoever builds the robots.12%
- Cash8%