Over the last week’s earnings/news cycle, the message sharpened: the constraint isn’t just GPUs—it’s the full bill of materials (especially high-bandwidth memory) plus the capex and depreciation burden of scaling inference. Coverage this week tied rising memory component costs directly to higher AI infrastructure spend and a higher profitability hurdle for Big Tech’s AI push.

AI economics are shifting from “software-like margins” toward “infrastructure-like margins” unless monetization catches up. If HBM stays tight, the cost of each incremental unit of useful inference won’t fall as quickly as the market narrative assumes—and depreciation becomes the silent margin killer on the income statement (not just a capex headline).

Most investors model AI as a demand story (more tokens, more subscriptions). The under-modeled variable is input cost inflation: memory pricing, networking, cooling, and power buildouts that don’t scale down nicely. When unit costs are sticky, the winners are the ones who can (1) price discriminate, (2) push inference to cheaper footprints, or (3) vertically optimize the stack.

This week’s enterprise platform announcements increasingly framed the problem as managing autonomous work, not just adding copilots. ServiceNow’s Knowledge 2026 theme (AI Control Tower / “Autonomous Workforce”) is essentially a pitch for a governed control plane that sits above proliferating agents and workflows—integrating across major clouds and enterprise apps.

Once agents touch real systems (tickets, identity, security policies, spend, customer actions), the bottleneck becomes control: permissions, audit, monitoring, and rollback. That shifts spending toward platforms that can provide observability + governance for AI actions, not just model access. This is where “enterprise readiness” hardens into a defensible moat.

People are treating “AI agents” as an application layer story. It’s quietly becoming an operating model story: who owns policy, how exceptions are handled, how autonomous work is measured, and who is liable when it breaks. That favors incumbents that already sit in system-of-record workflow paths (ITSM, security, GRC, IAM), because governance is easiest where the data/permissions already live.

A notable signal (not because it’s huge—because it’s structured) is the warehouse pilot where humanoid robots were dispatched through SAP Extended Warehouse Management to perform inspections autonomously inside a real facility. That’s less about a single robot and more about integration into existing operational systems.

Robotics adoption doesn’t scale on videos; it scales on workflow fit. When robots are triggered by the same systems that assign work to humans (WMS/EWM), you get the beginnings of “robot labor as a callable service.” That’s the bridge from pilots to procurement.

The near-term value isn’t replacing whole job families—it’s removing the annoying, repetitive inspection/handling tasks that create throughput losses and safety issues. That means the first ROI shows up as fewer exceptions, fewer stoppages, and better inventory integrity—not headline-grabbing headcount elimination.

Score: Level 3 — Structural
Explanation: The market still anchors on “AI gets cheaper over time,” but the cost stack is being pulled upward by constrained components (HBM) and by the accounting reality of depreciation on massive datacenter builds. Even if demand is real, margin capture becomes the question. This changes how quickly AI can become profit-positive at scale.
Market pricing: Underpricing the durability of cost pressure (and the advantage it gives vertically-optimized stacks).

Score: Level 3 — Structural
Explanation: This is the “unsexy layer” that enterprises always end up buying once experimentation hits production. Governance becomes mandatory when agents can act across systems. The control plane vendors can turn chaos into a recurring budget line, which is where long-duration software returns are rebuilt.
Market pricing: Correct-to-underpricing (recognized as a theme, but not fully modeled as a category expansion).

Score: Level 2 — Meaningful
Explanation: This is still early and mostly pilot-stage, but the integration pattern (robots dispatched via enterprise warehouse software) is the real tell. If that interface standardizes, adoption can compound quietly. The near-term is operational ROI, not sweeping labor replacement.
Market pricing: Underpricing the importance of integration moats (overpricing splashy hardware narratives).

Closing: The next 12 months won’t be decided by “who has the best model,” but by who can turn expensive AI and early robotics into governed, reliable production work.