AI Toll Collectors
ghoul@jeffrey
−12.4%
since Oct 7
The idea
I don't think you need to pick which AI startup wins to make money off the boom, you just need to own what every one of them has to pay to play: chips, cloud compute, data centers and networking, and the power to run it all. That's what this basket holds, concentrated across those four layers rather than spread thin, with 10% kept in cash to add on pullbacks. It's built around roughly two years of accelerating AI spend, and I'd revisit it if that spending trend actually breaks instead of just wobbling.
What it owns
- Every AI model in training runs on silicon, and whoever wins the software wars still needs to buy the hardware to get there. · NVIDIAIt manufactures the processors that train most large AI models, so it collects revenue whenever a startup or cloud rents compute to build one., Taiwan SemiconductorIt manufactures advanced chips for NVIDIA, AMD, and the custom silicon that hyperscalers design, so it captures a cut across multiple parts of the stack., BroadcomIt manufactures the custom AI chips that hyperscalers design to reduce their reliance on NVIDIA, and those orders are beginning to scale., Advanced Micro DevicesIt sells processors for AI compute as a credible second supplier to NVIDIA, and hyperscalers are ordering them for training workloads., Micron TechnologyIt manufactures high-bandwidth memory that ships on every AI chip, so its revenue scales with the volume of AI processors being built., Marvell TechnologyIt manufactures custom connectivity chips and processors that data centers deploy as they scale infrastructure for AI workloads.32%
- AI startups and enterprises rent compute rather than build their own data centers, and these companies are the landlords collecting that spending. · MicrosoftAzure rents compute to AI startups and enterprises, collecting their spending regardless of which model or startup ultimately succeeds., OracleIt signed major compute contracts with model builders and collects as those backlog commitments convert into actual revenue., AmazonAWS captures a large share of AI startup spending on compute, collecting burn rates whether or not any individual startup survives., CoreWeaveIt rents GPUs to AI companies with no other business, so accelerating AI compute spending flows directly to its revenue.22%
- AI compute lives in physical buildings wired together with specialized networking, and those buildings and connections must expand to handle the load. · Arista NetworksIt sells the networking switches that wire thousands of GPUs together into training clusters inside each data center., VertivIt manufactures cooling systems and power distribution equipment for data centers, and every new facility built requires its gear., Digital RealtyIt owns data center buildings and collects rent from clouds and AI companies regardless of which vendor's hardware or model wins., CoherentIt manufactures optical transceiver modules that move data between GPUs inside data centers at speeds copper cannot achieve., Astera LabsIt manufactures connectivity chips that allow GPUs, memory, and processors to communicate inside AI servers at scale.20%
- Data centers running AI chips consume enormous amounts of electricity and generate intense heat, and grid capacity to supply and cool that compute is the constraint. · Constellation EnergyIts nuclear plants are signing long-term contracts to supply baseload power to data centers, capturing premium pricing for always-on electricity., VistraIt sells power into markets where AI data centers are driving electricity demand higher, and firm power prices benefit its revenue., GE VernovaIt manufactures turbines and grid equipment to deliver new power capacity from generation sources to data centers across the country., EatonIt sells electrical distribution and backup power systems inside data center facilities, collecting a slice of every new site constructed.16%
- Cash10%