Investing.com -- Bernstein analysts identified the data center and neocloud companies most exposed to a potential slowdown in frontier AI model training, following a weekend essay from Anthropic CEO Dario Amodei calling for the industry to deliberately pace capability gains. Other AI leaders, including Sam Altman and Elon Musk, endorsed Amodei’s essay.

They clarified that the proposal centers on minimizing security risks through independent reviewers, safety standard coordination, and international cooperation, and "at this point, it’s not a call for a lowering of capex or stopping model training," Bernstein analyst Madison Rezaei said in a note. Still, she said many investors have begun questioning what happens if training slows.

From an infrastructure standpoint, Rezaei said a decline or slowdown in training activity would pull demand away from rural data center locations, many of which were purpose-built for latency-insensitive training workloads. Bernstein tracks a U.S. development pipeline of 488 gigawatts of nameplate capacity, of which it considers 170 gigawatts credible. Within that pipeline, 36% sits in rural areas and another 34% in Tier 3 markets such as West Texas, meaning roughly 70% was built with training or latency-insensitive inference in mind.

There’s little debate that inference will take a growing share of AI data center capacity over time, but the pace of that shift and its latency sensitivity remain unclear, since "we don’t yet know what the (profitable?), widely adopted use cases will be," Rezaei wrote. Some inference workloads, like bulk document processing and overnight batch scoring, can tolerate higher latency, while agentic workflows, real-time voice applications, and robotics are likely to require lower-latency infrastructure.

Regardless of timing, the analyst said Tier 1 metros remain the safest and most valuable locations, while Tier 4 rural markets carry the highest risk. She said impacts would disproportionately hit developers’ portfolios of new, not-yet-leased construction, along with neoclouds’ contracted-but-not-sold or shorter-term contracts.

CoreWeave was flagged as the most exposed name in Bernstein’s coverage. Rezaei estimates 25% of CoreWeave’s existing active U.S. power sits in Tier 3 and Tier 4 markets, along with roughly 74% of its contracted power. While the backlog is largely composed of take-or-pay contracts that limit near-term threat, she said it could see a pullback in demand for contracted-but-not-yet-sold rural power if training development slows.

On the flip side, Rezaei said Equinix, Digital Realty and CyrusOne are better insulated, with 95%, 92% and 94% of their existing U.S. capacity, respectively, located in Tier 1 and Tier 2 metros. Most future development from these companies is also expected to fall within those tiers, with interconnection density, particularly at Equinix, supporting demand in a more inference-driven market.