February 21, 2026
Security, Storage, and the Edge: The Infrastructure Beneath the AI Cycle
Most of the conversation around AI investing focuses on the obvious: models, chips, and applications. The more I look at how this cycle is playing out, the more the durable opportunity looks like it sits a layer beneath — in the infrastructure that lets everything else run securely, reliably, and at scale.
Security Is Becoming a Platform Problem
As compute scales, so do attack surfaces — digital and physical at once.
Data centers now house model weights, proprietary datasets, and inference systems representing billions of dollars in intellectual property. Yet physical security at many facilities still runs on fragmented tooling: standalone badge readers, siloed camera systems, and outsourced monitoring that never talks to the network layer.
The opportunity is in unification — platforms that connect physical access events with network authentication and anomaly detection in one view. The pattern is already visible at both ends of the market. Motorola Solutions has spent a decade rolling up video security and access control (Avigilon, Openpath, and a string of smaller deals) into a command-center software layer sold on recurring revenue. Axon did the same in public safety: hardware as the entry point, but the durable economics living in the evidence and workflow platform on top. On the digital side, CrowdStrike and Palo Alto are running the identical consolidation play — replacing dozens of point tools with a platform — because buyers have learned that fragmentation is itself the vulnerability. The companies that own the unifying layer turn security from a cost center into infrastructure, and as hyperscale buildouts accelerate they become structurally important to anyone operating at scale.
Storage Is AI's Long-Term Memory
Every AI system runs on a storage hierarchy. High-bandwidth memory and high-performance SSDs serve the computation happening right now — fast, expensive, and scarce; Micron and SK Hynix have had HBM capacity effectively sold out quarters in advance. But training runs also produce an enormous and growing tail of artifacts worth keeping: checkpoints, training archives, logs, and datasets that aren't needed in the current inference cycle but remain structurally critical. That persistence layer falls to high-capacity nearline drives, which is why Seagate and Western Digital — businesses the market spent a decade treating as terminal-decline commodities — have been reporting record nearline exabyte shipments into hyperscaler demand, with Seagate's HAMR transition extending the capacity roadmap.
As models grow and every run produces more worth preserving, the unglamorous end of the hierarchy quietly becomes a bottleneck. Unglamorous is often where the repricing happens last.
Lower Barriers to Building Mean More Demand for Infrastructure
This is the theme that ties the other two together.
AI is collapsing the cost and time of building software — small teams now ship full-stack applications in days on platforms like Vercel and Netlify. The interesting dynamic is that as it gets easier to build software, the infrastructure required to run it doesn't get simpler. Every one of those new applications needs deployment, DNS, DDoS protection, and global distribution, which routes demand to the edge layer that Cloudflare, Akamai, and Fastly operate. The structural tailwind is that these companies are indifferent to which applications win: more builders means more customers, and the number of builders is going up by an order of magnitude.
What This Means
These three themes — unified security, the storage hierarchy, and edge infrastructure for a growing application layer — share one characteristic: none of them is a bet on which AI model wins. They're bets on what every AI system, and increasingly every software application, requires to function.
The picks-and-shovels metaphor is overused, but the logic holds. When an entire industry is scaling rapidly, the companies supplying structural necessities tend to compound quietly while the spotlight stays on whatever is newest. The test I apply: does demand for this layer grow regardless of which specific technology or application captures the market's imagination? Where the answer is yes, that's where I'm spending my research time.