Key Takeaways

  • Andrew Thomas projects that AI inference compute demand will expand to several multiples the physical footprint and capacity of training clusters.
  • Latency requirements are redirecting data center capital away from remote rural tracts and back into tier-one metro hubs located near end users.
  • Network interconnection density is replacing raw acreage as the primary asset differentiator, boosting valuations for existing urban data platforms.
  • Incumbent platforms benefit from high switching costs, structural barriers to entry, and firmer contract structures that insulate them from inexperienced developers.

The Compute Pivot From Remote Power to Local Interconnection

Capital spent on artificial intelligence over the past two years focused almost entirely on building centralized clusters for model training. These massive facilities prioritized access to bulk power over geographic placement. Because training runs take weeks or months without requiring live user interaction, operators parked gigawatts of compute in remote jurisdictions with cheap electricity and vast land reserves.

That capital allocation model is hitting a structural turning point. Thomas observes that production workloads are shifting rapidly toward real-time query responses and application delivery. As Thomas explains, “We are going to see a material step change in the number of deployments that are catering towards AI inference. Most of what we're seeing and most of the management teams that we spend a lot of time with think that the inference opportunity is going to be several multiples the size of the AI training capacity.”

Inference workloads cannot tolerate the transit delays inherent in cross-country data transmission. When enterprise software, autonomous systems, and consumer interfaces query a trained model, round-trip latency dictates performance. That dynamic changes where compute hardware must live.

As Thomas notes, “That is going to be more latency specific. That is going to be back to tier one markets that are closest to end consumers. We're going to see a meaningful amount of investment and shift for people who have existing data center platforms to try to address that demand need.”

Defensibility in Urban Interconnection Assets

The return to urban centers highlights a sharp division between raw real estate developers and mature digital infrastructure platforms. Building in tier-one metropolitan areas introduces severe power distribution constraints, complex municipal permitting, and high land costs. Capital alone cannot manufacture new carrier hotels or duplicate dense optical cross-connect networks overnight.

Thomas points out that market participants will place higher value on strategic network connectivity: “When you have more of a move to inference and a heightened focus again on what is the interconnection position that a lot of these assets have to capitalize upon the shift into inference, I think you'll see people recognize, all right, that's a strategically important asset.”

This structural moat protects established platforms against the flood of new capital attempting to build speculative mega-campuses in secondary regions. Thomas emphasizes the durable economics of these urban nodes: “You've got mission criticality of assets. You've got very high barriers to entry, very high switching costs. You have improved contract quality as well.”

Why It Matters

This deployment shift signals a valuation reset across digital infrastructure portfolios. Capital allocated to remote, training-only facilities faces long-term asset stranding risk as compute budgets migrate toward inference at the edge. Investors holding tier-one facilities with dense interconnection networks can capture higher pricing power, longer contract commitments, and lower tenant churn as enterprise workloads move into production.