The Inference Stack
hrishikesh kamath@recursiveagi
−10.7%
since Sep 30
The idea
I believe running AI models, not training them, becomes the real bottleneck over the next decade, so I built exposure across the whole stack rather than betting on one layer. This basket owns inference chips, memory bandwidth, networking and optics, foundries and equipment, and the power and data centers underneath it all, leaning on larger established names as ballast. I keep a modest cash reserve to add on weak days, and I'd rethink this if inference demand growth actually stalls out.
What it owns
- The processors that run AI models are the most direct way to own rising inference demand. · NVIDIAIt runs most AI models in production today, so rising inference demand lands there first., BroadcomIt builds custom inference chips for the biggest AI companies looking to run models faster and cheaper., Advanced Micro DevicesIt is the main alternative source for inference GPUs, so owning the stack means owning both routes to competition., Marvell TechnologyIt builds custom inference chips for cloud giants, representing the second major route to proprietary silicon., QualcommIt makes chips for on-device inference on phones and laptops, capturing the models running outside data centers.27%
- A decade of inference load needs electricity and buildings, the steadiest layer and the foundation of this basket. · Vertiv HoldingsIt sells the cooling and power gear that lets data centers run inference chips densely without overheating., EatonIt sells switchgear and power distribution for every new data center, growing with the buildout even if any one vendor stumbles., Constellation EnergyIts nuclear fleet provides the clean, steady electricity that running models billions of times a day requires., Digital RealtyIt owns the buildings inference load runs in and collects rent as capacity fills with new demand., EquinixIt runs the interconnection hubs where clouds and enterprises plug into each other, mattering more as inference traffic moves closer to users.20%
- Fast inference at scale means moving data between thousands of chips, which the networking and optics layer controls. · Arista NetworksIt sells the switches that connect clusters of inference chips talking to each other constantly at scale., CoherentIt makes optical transceivers that move inference data as light at the speeds data centers increasingly demand., Astera LabsIt makes connectivity chips inside AI servers that move data between processors and memory as inference speeds climb., Credo TechnologyIt makes high-speed cables and chips that carry data between racks in AI data centers as capacity expands.18%
- Every chip in the stack has to be manufactured somewhere, and the makers and their toolmakers collect from all layers. · Taiwan SemiconductorIt manufactures nearly every inference chip in this take, collecting from every layer as the stack grows., ASML HoldingIt is the sole maker of machines that print the most advanced chips, so a decade of inference demand runs through its order book., Applied MaterialsIt sells equipment that fabs buy when they expand capacity to meet inference demand., Amkor TechnologyIt assembles modern inference chips from multiple stacked pieces, a step that has been a real bottleneck.17%
- Inference speed is usually limited by how fast memory can feed the chip, making memory suppliers a key constraint. · Micron TechnologyIt makes high-bandwidth memory that solves the bottleneck of feeding data to inference chips fast enough., RambusIt sells interface chips and designs that connect fast memory to processors, getting paid whenever memory bandwidth constrains speed.10%
- Cash8%