Agentic Stack, No GPUs
Stephen@stephen
−12.0%
since Oct 9
The idea
I believe agentic AI is early innings, and the money flows to everything around the GPU rather than the chip itself. So this owns the plumbing agents actually run on: CPUs, memory, storage and networking, the clouds and sandboxes that host them, identity and payments so agents can act and transact, databases for agent memory, and the power and cooling behind it all. I kept a cash buffer and spread bets across layers rather than betting on single names, holding for years unless the thesis that spend shifts from training to always-on inference breaks down.
What it owns
- Always-on agents burn CPUs, memory, storage and networking far harder than one-shot chatbots, creating demand beyond the GPU everyone already owns. · Arm HoldingsIt collects royalties on most CPUs wherever agents run, providing clean exposure to compute growth without direct GPU dependence., Micron TechnologyAgents that run constantly and hold state consume memory at scale, and it supplies the memory that inference workloads require., Arista NetworksIt builds the network backbone inside datacenters where agents run, earning revenue on every buildout regardless of which chips fill the racks., Marvell TechnologyIt designs custom inference silicon for major cloud providers, capturing the shift from training GPUs toward chips built for running agents., Seagate TechnologyAgents that hold state generate data requiring bulk storage, and it supplies the drives where that data lives long-term., Astera LabsIt makes the connectivity inside AI racks, the component that becomes the bottleneck when compute and memory are packed densely for inference., Credo TechnologyIt sells the cabling that links everything inside AI datacenters, a small cost per rack multiplied across many facilities.30%
- Identity, payments, memory and monitoring for agents are barely priced today but become essential once agents act autonomously at scale. · DatadogOnce companies run fleets of autonomous agents, someone must watch what those agents do, and it is positioned as the default monitoring platform for agent behavior., OktaMachine identities already outnumber human ones, and it brokers credentials for nonhuman entities at scale., VisaIf agents buy things on their owner's behalf, they need payment rails, and it operates live agent-payment programs today., MongoDBAgents hold state and need memory between requests, and it is positioned as the database serving that agent memory layer., TwilioAgents still reach humans through voice and messaging, and it owns the communication rails they use to do so., SailPointIt governs nonhuman identities at the enterprise level, addressing agent identity governance as a distinct product category.28%
- Agents need isolated sandboxes and serverless compute to run in, and these companies provide exactly that infrastructure. · CloudflareIt sells the sandboxes and edge runtime where agents actually execute, forming the first layer of managed infrastructure for agent code., OracleIt rents cloud capacity for always-on agents and inference workloads without requiring a bet on any single chip manufacturer., Akamai TechnologiesIt is repositioning its edge network as compute for inference close to users, where agents that need real-time response prefer to run., DigitalOceanIt is the cloud where small teams and solo builders deploy agents, positioned to benefit as agentic development spreads beyond enterprises.20%
- Agents that never sleep need steady electricity and dense cooling, which forms the physical foundation under the entire theme. · Vertiv HoldingsDatacenters running always-on agents never idle, requiring continuous power and cooling infrastructure that it supplies and installs., GE VernovaIt supplies the turbines and grid equipment that deliver electricity to datacenters, with order books driven by rising AI power demand., Constellation EnergyAgents running around the clock need electricity around the clock, and it provides always-on nuclear power under long datacenter contracts.14%
- Cash8%