Equinix, Digital Realty Positioned to Withstand AI Training Slowdown, Bernstein Says

A potential slowdown in the development of frontier artificial intelligence models could reshape the data center landscape, favoring established operators in major metropolitan areas over facilities in rural locations, according to an analysis from Bernstein.

The Catalyst for Caution

Investor inquiries have risen following a recent essay by Anthropic CEO Dario Amodei, endorsed by other industry leaders, which called for the industry to deliberately pace AI capability gains to manage security risks. While Bernstein analyst Madison Rezaei noted this is not an explicit call to cut capital expenditures, it has prompted questions about the potential impact on infrastructure demand.

If AI model training activity were to slow, demand would likely shift away from rural data centers, many of which were specifically developed for such latency-insensitive workloads. Bernstein estimates that of the 170 gigawatts of credible U.S. data center capacity in the development pipeline, roughly 70% is located in rural (Tier 4) and Tier 3 markets.

Geographic Exposure is Key

Bernstein's analysis identifies a clear divergence based on geography. The firm views Tier 1 metropolitan markets as the "safest and most valuable locations," while rural markets carry the highest risk in a slowdown scenario. The long-term shift from AI training to inference workloads is widely expected, but the pace and latency requirements of widely adopted applications remain uncertain.

Ad

• High-Latency Inference: Workloads like bulk document processing can be handled in remote locations.
• Low-Latency Inference: Applications such as real-time voice assistants and robotics will require infrastructure closer to end-users in major urban centers.

Potential Winners and Losers

According to the note, this dynamic creates distinct winners and losers among data center operators. CoreWeave was identified as the most exposed, with an estimated 25% of its active U.S. power and 74% of its contracted power located in riskier Tier 3 and Tier 4 markets. While its take-or-pay contracts limit near-term threats, a slowdown could dampen future demand for its rural capacity.

Conversely, established players are seen as better insulated. The analysis highlighted that Equinix (EQIX), Digital Realty (DLR), and CyrusOne have the vast majority of their existing U.S. capacity in more resilient Tier 1 and Tier 2 metros — 95%, 92%, and 94%, respectively. Bernstein added that Equinix's high degree of interconnection density provides an additional advantage in a market driven more by inference workloads.