AI Picks and Shovels
Jordan Kobert@jordankobert
The idea
I think the real winners of the AI boom are the suppliers making the buildout physically possible, not just the headline chip maker. So this take splits across two sides: the physical layer of grid gear, power generation, and cooling, and the rack-level plumbing of memory, networking, optics, and custom chips. Nvidia stays in but as a smaller piece, with more weight on bottleneck suppliers I think are underpriced. It's built for a multi-year buildout, spread across many names rather than one bet, with a small cash reserve to add if prices drop.
What will be bought
- The processors, custom chips and high-speed memory are the core of every AI server. · Micron TechnologyIt makes high-bandwidth memory that stacks next to every AI chip and has been in tight supply for data center buildout., Broadcom Inc.It designs custom AI chips for major cloud companies and makes the switching chips that tie GPU clusters together at scale., Taiwan Semiconductor ManufacturingIt manufactures nearly all advanced AI chips for every designer, so it gets revenue no matter which architecture wins out., NVIDIA CorporationIt makes the GPUs that form the foundation of every AI data center, anchoring the infrastructure ecosystem., Marvell TechnologyIt builds custom AI chips and optical connection chips for data centers, competing for designs from major cloud companies.28%
- New data centers can't switch on without turbines, transformers and grid connections, which are in short supply. · GE VernovaIt sells turbines and grid equipment that new data centers need to connect and power on as AI infrastructure expands., Eaton CorporationIt makes switchgear and power distribution equipment that connects the grid to servers inside every new data center., Quanta ServicesIt builds the transmission lines and substations that utilities need to deliver power from generation sources to new data center sites., Hubbell IncorporatedIt sells transformers, connectors and other equipment that utilities purchase when expanding the grid for new data center loads.21%
- Thousands of chips have to talk to each other fast, which makes switches, optical links and connectors a growing bottleneck. · Arista NetworksIt sells the high-speed switches that connect thousands of GPUs into single AI clusters for major cloud operators., Coherent CorpIt makes the optical transceivers that convert electrical signals into light, enabling fast data movement between server racks., Credo TechnologyIt makes active electrical cables and connectivity chips that link servers and switches inside AI racks more cost-effectively than optical links for short distances., Astera LabsIt makes connectivity chips that let GPUs, CPUs and memory communicate with each other inside AI servers.19%
- AI chips run hot, so every new rack needs serious cooling and electrical work to get built. · Vertiv HoldingsIt supplies the power and liquid cooling systems that AI racks require because the chips generate more heat than air cooling can handle., EMCOR GroupIt performs the electrical and mechanical contracting inside data centers, installing wiring and cooling systems before servers power on., Modine ManufacturingIt supplies cooling equipment for data centers, shifting its business mix toward this faster-growing segment.13%
- Data centers need huge amounts of steady electricity, and the companies that produce it can charge more as demand climbs. · Constellation EnergyIt operates the largest US nuclear fleet, providing the round-the-clock carbon-free power that major tech companies need for data centers., Vistra CorpIt owns gas and nuclear plants and sells power at market rates, benefiting from rising data center demand in Texas and the mid-Atlantic., Bloom EnergyIt makes fuel cells that let data centers generate power on site, avoiding years-long waits for new grid connections.12%
- Cash7%