Off the Run

The Adolescence of AI

Last week, when Meta announced they would be repositioning much of their AI data center infrastructure for resale to third parties as it sat unused by their own AI models, the market reaction was swift: Supply glut! AI bubble! Sell the neoclouds! Buy Meta? Putting aside the market’s confusion, the real story here is far more nuanced and, we believe, a credit positive signal. If you have been following the technology industry for more than two decades, as we have, this moment is quite familiar. The Meta announcement does not signal the popping of the AI bubble, but rather that the market is beginning to mature out of its feverish, wide-eyed infancy into a more grounded, disciplined adolescence.

The Meta Compute announcement is a major signal that the AI ecosystem is starting to mature. Meta may never win the frontier model race, and that’s OK. Its 5 GW of contracted capacity in just the first half of 2026 does not represent a company retreating from AI. It represents a company that has figured out where it fits in the stack today and prioritizing ROI on its capital investment. This is similar to SpaceX’s recent leasing of its Colossus data center space to Anthropic just ahead of its IPO. Neither of these events signal an existential end of the AI ecosystem, rather they positively reflect the market rationalizing and responding to true supply and demand signals – something mature markets do.

We are simultaneously getting signals for what the end state of AI compute might look like as the market matures. Palantir CEO Alex Karp recently went viral with his commentary on CNBC when articulating (perhaps more emphatically than we would) our core thesis on how AI infrastructure and enterprise AI adoption evolve. Enterprises will eventually route AI workloads through an orchestration layer for all major queries. That layer will disburse work based on cost, latency, and data sensitivity. Complex tasks will be routed to high-performance frontier models at premium prices, commodity inference to low-cost public cloud and importantly, anything touching proprietary data will likely eventually be routed to a private infrastructure stack running secure, and low-cost open models on the enterprise’s own hardware (many of which will likely reside in third-party infrastructure). Despite the market’s reaction to Karp’s comments, the fact is that every major technology stack since the cloud has converged on this kind of hybrid architecture. More providers, more competition, and more variety signals market maturation.

The implication for infrastructure investors is straightforward. More paths to AI adoption means more demand for the physical layer (land, power, AI-oriented data centers, fiber connectivity) if it is in the right location with the right power and technical specifications. Enterprise demand is real and the orchestration layer does not care which model wins the frontier race. The private cloud stack needs the same fiber, the same cooling, and the same power as the hyperscaler next door. Physical infrastructure is the common thread running through every version of this future, and it does not become optional as the market matures. If anything, broader adoption makes it more essential, even as the names of the leading players change over time (and they will).

The caveat here is timing. Markets and technologies mature on their own schedule, and the transition from early infant exuberance to responsible maturity has never been orderly. Some of today’s most celebrated players will not survive the journey to the end state. Not only is that OK, it’s to be expected.