Inference Everywhere
Rishwajeet@rjrazzle
−11.1%
since Sep 30
The idea
I believe AI's real growth is shifting from training models to running them everywhere, all day, on two different clocks. So this basket leans hardest into data center inference: the chips serving queries, the memory and networking that inference actually strains, and the clouds selling it by the token, with a smaller, patient sleeve for the on-device chipmakers powering the slower local wave.
What it owns
- When applications fire inference queries all day, the chips serving those requests in data centers capture the first wave of demand. · NVIDIA CorporationIt supplies the GPUs running most inference workloads in data centers as long as demand for serving queries keeps growing., Broadcom Inc.It designs custom inference chips for the biggest cloud providers, winning as long as hyperscalers prefer their own silicon over off-the-shelf options., Advanced Micro DevicesIt serves as the credible second source for data center inference chips, capturing share when buyers want alternatives to the dominant supplier., Taiwan Semiconductor ManufacturingIt manufactures every inference chip in this take across both data centers and devices, so it gets paid from the entire wave regardless of which processor wins., Marvell TechnologyIt designs custom inference chips and their surrounding connectivity for major cloud providers, growing as bespoke silicon programs ramp.30%
- Inference queries travel between chips within servers and across data centers, creating constant demand for the switches and optical links that carry that traffic. · Arista NetworksIt makes the switches connecting clusters of inference chips, capturing value from the constant traffic between processors serving queries., Astera LabsIt makes the connectivity inside AI servers that moves data between chips, addressing the exact bottleneck inference workloads hit., Coherent Corp.It makes the optical links that carry inference data between distant clusters in data centers as AI infrastructure scales., Credo TechnologyIt supplies high-speed cables and links inside inference data centers, capturing value from expanding connections between server racks.15%
- Companies that charge per inference call or per token collect revenue on every AI query an application makes through their cloud platforms. · Microsoft CorporationIt sells inference by the token through Azure and embeds it across its product suite, collecting revenue on rising query volume., Alphabet Inc.It runs inference on its own TPU chips in its own cloud, keeping more margin as inference demand grows and maintaining competitive serving costs., Oracle CorporationIt rents data center capacity for inference workloads, acting as a landlord for the infrastructure surge driven by query volume., CoreWeave Inc.It rents GPU capacity focused on inference, capturing pure upside from the query surge while remaining vulnerable if demand softens.15%
- Models running on phones and laptops represent a slower adoption wave, dependent on on-device AI becoming a meaningful reason for users to upgrade their hardware. · Qualcomm IncorporatedIts chips power models running on most smartphones and an expanding share of laptops, positioned for the patient bet on on-device AI adoption., Apple Inc.It owns the largest installed base of devices capable of running models locally, positioned for eventual upgrades driven by on-device AI., Arm HoldingsIt collects royalties on nearly every device that will run local inference, profiting from the entire on-device wave regardless of timeline., Monolithic Power SystemsIt provides precise power delivery for AI chips in both data centers and devices, profiting from growth on both sides of the inference split.15%
- Inference workloads move massive amounts of data in and out of memory, so suppliers of high-bandwidth memory get paid regardless of which processor wins. · Micron TechnologyIt makes high-bandwidth memory where inference chokes on data movement, capturing value from the tightest constraint in the chain., Rambus Inc.It collects licensing fees on the memory interface technology that faster inference memory depends on, profiting from bandwidth improvements.12%
- Data centers running inference around the clock generate intense heat, and the equipment that powers and cools these facilities becomes essential infrastructure. · Vertiv HoldingsIt supplies the power and cooling systems that keep inference data centers running around the clock as capacity demands expand.5%
- Cash8%