The Shifting Sands of the AI Trade

Investors continue to debate whether the first phase of the AI trade — hardware (semiconductors and servers) and infrastructure investment (data centers and associated technologies) is entering bubble territory. But prudent investors should recall Wayne Gretzky’s maxim: “I skate to where the puck is going to be, not where it has been.”

So where is the puck heading next? Toward the early adopters of agentic technology. Agentic systems represent the evolution beyond single-pass inference — the familiar input → model → output → stop pattern used in applications such as spam classification or summarizing support calls. In contrast, agentic AI follows a plan → act → evaluate → adapt cycle, enabling automation of entire processes and reducing or eliminating human intervention. Both inference and agentic approaches will define the next phase of AI adoption, with inference retaining a roughly 4:1 cost advantage and serving as the initial workhorse.

This transition is underscored by a recent Microsoft announcement. The company has begun routing certain simple AI requests from Excel and Outlook to its in-house models rather than to OpenAI or Anthropic. This shift represents a meaningful cost-saving move and will have a negative marginal impact on those external model providers. It also broadens and accelerates the inference market by lowering operating costs. Still, we expect inference-only approaches to be short-lived; the industry will move quickly toward full agentic applications that deliver on AI’s broader promise.

The winners and losers are becoming clearer. Microsoft is a winner. Corporate customers benefit. Infrastructure players — especially communications providers — stand to gain as the market expands. Hyperscalers also win. On the other side, software companies unprepared for agentic technology and high-cost component suppliers, including some semiconductor firms, may face near-term pressure.

Another critical factor is corporate data protection. Today, enterprise data is siloed, fragmented, and inconsistent. Before AI agents can operate effectively, that data must be normalized, unified, governed, and contextualized — and importantly, kept in the possession of the end user rather than transferred externally. Here again, firms that enable this approach will benefit.

Skeptical investors who argue that the AI trade is over may be missing the forest for the trees. Yes, the first phase of the opportunity may be maturing, but the shifting sands suggest the next phase is only now beginning. The question is not whether the hardware cycle is in bubble territory — it’s what comes next.

Bruce M. Lupatkin
Managing Member

Michael P. DeSantis
Managing Member