A slowdown in frontier-model training would favor metro data centers over rural, training-oriented capacity.

Overview

Bernstein examines how deliberately slower frontier-model capability development could reshape U.S. data-center demand. Its central conclusion is that demand would shift away from rural, training-oriented capacity toward lower-latency metro infrastructure, benefiting EQIX, DLR and CSQR relative to CoreWeave.

Core views

The report responds to investor concerns that frontier-model training may slow after calls for more deliberate capability development and stronger security safeguards. Bernstein stresses that the underlying framework concerns independent review, safety standards and international cooperation rather than an immediate call to cut capital expenditure or stop training. Nevertheless, a reduction or deceleration in training would shift demand away from rural locations purpose-built for latency-insensitive workloads. Bernstein estimates a U.S. data-center development pipeline of 488GW of nameplate capacity, of which 170GW is considered credible. Rural markets account for 36% of that pipeline and Tier 3 markets, including West Texas, account for another 34%. Thus, 70% of the pipeline is positioned for training or latency-insensitive inference. The report expects inference to take a rising share of AI data-center capacity, but says the timing and degree of latency sensitivity remain uncertain because profitable, widely adopted use cases are not yet known. The report separates inference workloads by latency needs. Bulk document work, synthetic-data generation and distillation, and overnight batch scoring can operate in training-type facilities. In contrast, agentic workflows involving serialized tool calls, real-time voice, live translation, payment-fraud authorization and robotics are likely to require lower-latency infrastructure. This makes Tier 1 major metros the safest and most valuable locations under either timing outcome, while Tier 4 rural markets carry the greatest risk. Bernstein expects the pressure to fall disproportionately on developers' new, unleased construction and on neocloud capacity that is contracted but not sold or secured on shorter-term contracts. CoreWeave is identified as the most exposed company in Bernstein's coverage. The firm estimates that 25% of CoreWeave's active U.S. power and about 74% of its contracted power are in Tier 3 and Tier 4 markets. Bernstein does not see its primarily take-or-pay backlog as immediately threatened, but it could see demand weaken for rural contracted-but-not-yet-sold power if training development slows. Conversely, 95% of Equinix's, 92% of Digital Realty's, and 94% of Csquare's existing U.S. capacity is in Tier 1 and Tier 2 metros, where most of their future development is also expected to occur. Bernstein further identifies interconnection density—especially at Equinix—as a demand support for these metro locations in a more inference-driven market. It maintains Outperform ratings on EQIX, DLR and CSQR, with targets of $1,270, $226 and $27, respectively, while maintaining an Underperform rating and $74 target on CRWV.

Analysis framework

Bernstein links a potential slowdown in AI training to the geographic and latency profile of U.S. data-center capacity. It compares rural and Tier 3 exposure with Tier 1 and Tier 2 metro exposure, distinguishes latency-tolerant from real-time inference workloads, and applies forward valuation multiples to its covered equities.

Methodology notes

• Industry AnalysisSupply-demand framework

  Data-center demand is assessed through the expected mix of AI training and inference workloads and their location requirements.

  The report argues that slower training would reduce the relative demand for rural capacity, while latency-sensitive inference would support metro data centers.

• Valuation methods

  Forward earnings and cash-flow multiple valuation.

  Bernstein values CRWV using forward EV/Adjusted EBIT, CSQR using EV/Adjusted EBITDA, and DLR and EQIX using forward price-to-AFFO per-share multiples.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• CoreWeave (CRWV)

  Most exposed to slower training because of substantial Tier 3 and Tier 4 power exposure.

  Strengths

  Its backlog is primarily composed of take-or-pay contracts.

  Weaknesses

  An estimated 25% of active U.S. power and ~74% of contracted power are in Tier 3 and Tier 4 markets.

  Comparison

  Less insulated than EQIX, DLR and CSQR, whose existing capacity is concentrated in Tier 1 and Tier 2 metros.

  Risks

  Hyperscaler build delays, specialized-software development, or continued power shortages could sustain its value proposition and pricing power.

• Equinix (EQIX)

  Beneficiary of an inference-led shift toward low-latency metro capacity.

  Strengths

  95% of existing U.S. capacity is in Tier 1 and Tier 2 metros; interconnection density is a particular advantage.

  Comparison

  Higher Tier 1 and Tier 2 concentration than CoreWeave's Tier 3 and Tier 4 exposure.

  Risks

  Enterprise-demand slowdown, greater interconnection competition, and declining data-center supply market share.

• Digital Realty Trust (DLR)

  Beneficiary of a shift toward metro, lower-latency infrastructure.

  Strengths

  92% of existing U.S. capacity is in Tier 1 and Tier 2 metros.

  Comparison

  More insulated than CoreWeave from a slowdown in training-oriented rural demand.

  Risks

  Lower-than-expected growth in the sub-1MW segment, slower enterprise demand, and market-share declines that could increase pricing pressure.

• Csquare (CSQR)

  Beneficiary of a shift toward metro, lower-latency infrastructure.

  Strengths

  94% of existing U.S. capacity is in Tier 1 and Tier 2 metros.

  Comparison

  More insulated than CoreWeave from a slowdown in training-oriented rural demand.

  Risks

  Slower-than-expected deleveraging, execution risk in unlocking incremental capacity, and Brookfield's retained board control and majority vote.

Key data

• U.S. data-center development pipeline488GW nameplate; 170GW considered credibleBernstein's estimate of the U.S. pipeline.
• Pipeline in rural and Tier 3 markets36% rural; 34% Tier 3; 70% combinedCapacity oriented toward training or latency-insensitive inference.
• CoreWeave exposure25% of active U.S. power; ~74% of contracted powerLocated in Tier 3 and Tier 4 markets.
• Metro concentrationEQIX 95%; DLR 92%; CSQR 94%Share of existing U.S. capacity in Tier 1 and Tier 2 metros.
• CRWV valuation basis25.5x NTM+1 adjusted operating income of $5.2BSupports Bernstein's $74 price target.
• DLR valuation basis27x NTM+1 AFFO per share of $8.36Supports Bernstein's $226 price target.
• EQIX valuation basis~25x NTM+1 AFFO per share of $51.19Supports Bernstein's $1,270 price target.
• CSQR valuation basis15x NTM+1 adjusted EBITDA of $568MBernstein's stated valuation basis.

Impact & implications

Bernstein's location-based framework favors operators with dense, interconnected Tier 1 and Tier 2 footprints if inference becomes more important, while leaving rural, unleased or not-yet-sold training-oriented capacity more exposed to a slowdown in frontier-model development.

Risks

• For CoreWeave, delays in hyperscaler capacity, development of its specialized software platform, or continued power shortages could support demand and pricing power.
• For Csquare, deleveraging could take longer than expected, incremental-capacity execution may disappoint, and Brookfield retains board control and a majority vote.
• For Digital Realty, sub-1MW growth or enterprise demand could be weaker than expected, while market-share loss could raise pricing pressure.
• For Equinix, enterprise demand could slow, interconnection competition could increase, and market-share loss could raise pricing pressure.