Eben@CANSLIM ResearchBernstein has flagged CoreWeave (CRWV) as the most exposed name in its US communications infrastructure coverage if frontier-model training slows, noting that 25% of CoreWeave’s active US power and roughly 74% of its contracted power sits in Tier 3 and Tier 4 markets. The broker maintains an Underperform rating and $74 price target on CRWV, while keeping Outperform ratings on Equinix (EQIX, $1,270), Digital Realty (DLR, $226) and Csquare (CSQR, $27). The call follows an essay by Dario Amodei arguing frontier models should deliberately slow capability gains — endorsed by Sam Altman and Elon Musk — which has prompted investors to ask what happens to data centre demand if training decelerates.
Key Takeaways
- Bernstein estimates 70% of the 488GW US data centre development pipeline — of which 170GW is considered credible — is built for latency-insensitive training demand, with 36% in rural areas and 34% in Tier 3 markets such as West Texas.
- CoreWeave (CRWV) carries an Underperform rating and $74 price target from Bernstein, with 25% of active US power and about 74% of contracted power located in Tier 3 and Tier 4 markets.
- Equinix (EQIX), Digital Realty (DLR) and Csquare (CSQR) are insulated, with 95%, 92% and 94% respectively of existing US capacity in Tier 1 and Tier 2 metros.
- Bernstein maintains Outperform ratings on EQIX ($1,270), DLR ($226) and CSQR ($27), citing interconnection density as a structural advantage in a more inference-driven market.
- CoreWeave’s backlog is largely take-or-pay, so Bernstein does not see a threat there, but warns of a possible demand pullback for contracted-not-yet-sold rural power.
What Bernstein’s Data Centre Note Says About Slowing AI Training
In a report titled “Quick Take: Data centers and neoclouds – winners and losers if frontier models do slow development,” Bernstein analysts Madison Rezaei and Nancy Wu argue that a deliberate slowdown in frontier-model capability gains would redirect demand away from rural data centre locations purpose-built for latency-insensitive training workloads. The note, first published 14 September 2026, follows Dario Amodei’s weekend essay calling for frontier models to slow capability gains, which Sam Altman and Elon Musk both endorsed. Bernstein stresses the framework centres on minimising security risks — independent reviewers, coordinated safety standards and international cooperation — and is not a call to cut capex or halt model training. Even so, the analysts note many investors have begun questioning the consequences if training slows.
Subscribe to continue reading
Become a paid subscriber to get access to the rest of this post and other exclusive content.