Agents Eat the Interface
Alex Alspaugh@alex
−4.3%
since Sep 30
The idea
I believe compute demand keeps compounding and the value accrues to disciplined hyperscalers, not the neoclouds renting temporary capacity. As agents replace software interfaces, the winners are whoever owns the compute, the silicon, and the data agents must talk to, so this basket layers hyperscalers, chips and networking, systems of record, horizontal consolidators, and datacenter buildout, with no exposure to thin-UI, seat-based software. Since more agent-shipped code means more bugs and more attack surface, I added a security layer weighted toward identity.
What it owns
- The largest cloud providers with disciplined capital spending capture the compute demand that models and agents require. · AlphabetIt makes its own chips and sells compute capacity, capturing the take twice through TPU cost savings and Google Cloud market share., MicrosoftAzure sells compute to model builders while Copilot sits in front of it, capturing the bet as long as Azure grows and the company stays disciplined., AmazonAWS is the largest hyperscaler, and the position depends on AWS growth staying strong and capex remaining tied to real demand.26%
- Smarter models and new agent products pull more chips, memory, and networking infrastructure through the supply chain. · NVIDIAIt collects on nearly every dollar hyperscalers spend on chips as models demand more compute and capability., Taiwan SemiconductorNearly every advanced AI chip gets manufactured at its fabs, making it the manufacturing bottleneck for the take's compute growth., BroadcomIt builds the custom chips and networking infrastructure hyperscalers use to reduce dependence on standard components., Arista NetworksIt supplies the networking gear hyperscalers standardize on to wire together new compute infrastructure., Micron TechnologyIt makes the high-bandwidth memory AI chips require alongside compute to run smarter models.21%
- When agents perform tasks instead of humans, the databases and data platforms they query remain essential to enterprise operations. · MongoDBIt is the database agents query when they replace user interfaces, keeping it essential to enterprise data operations., SnowflakeIt holds enterprise data that agents must access, staying paid even as applications on top lose traditional seat revenue., DatadogIt provides the monitoring and observability that agent-written code in production requires for safety and reliability., CloudflareIt sits in the path of agent traffic to real systems, routing and screening the data and requests agents generate.16%
- Software consolidates as models let companies operate horizontally, and these platforms are winning that consolidation. · Meta PlatformsIt owns its models, compute, and distribution, positioning itself as a horizontal consolidator in enterprise where verticals used to exist., ServiceNowIt owns the workflow data agents must act on, positioning itself as the platform agents plug into for enterprise operations., PalantirIt is built around models acting on enterprise data rather than user interfaces, positioning it for the split this take bets on.11%
- Compounding compute demand requires physical power and cooling infrastructure, making equipment suppliers second-order winners. · VertivIt provides the cooling and power equipment every new datacenter needs to support growing compute demand., GE VernovaIt sells the turbines and grid equipment new datacenters require to power compounding compute demand., EatonIt makes the electrical infrastructure inside datacenters, providing steadier exposure to datacenter construction demand.10%
- Agent-written code creates more bugs and attack surface, so security vendors benefit from both the immediate threats and long-term spending. · CrowdStrikeIt defends against attacks on agent-written code and the systems agents interact with, benefiting from rising security threats., Palo Alto NetworksIt consolidates security spending onto a platform as companies need better defense against agent-driven threats., OktaIt manages identity and credentials for machines acting on real systems, making machine identity a distinct spending category.8%
- Cash7%