Bernstein: If AI model development slows down, CoreWeave (CRWV.US) may be the most affected

Investment bank Bernstein said that the slowdown in AI model training called for by industry giants could have a negative impact on data center developers and neocloud service providers, with CoreWeave (CRWV.US) potentially being one of the most affected companies. Anthropic CEO Dario Amodei called for slowing the pace of frontier AI model development in a warning article published on September 12. In the article, he stated bluntly that the risks posed by AI are "severe" and that time must be taken to address these risks. This call was quickly echoed by two key figures. Musk reposted Amodei's post on social media platform X, adding: "Dario is right." OpenAI's leader Altman wrote on X: "I agree with Dario, we need to control the pace of frontier AI advancement." Bernstein analyst Maidson Rezaei said in an investor report on Monday that a slowdown in model training speed or a decline in training volume would mainly affect data centers in rural areas. The analyst said: "From the perspective of infrastructure and neocloud service providers, a decline or slowdown in training demand would cause demand to flow away from rural data center locations, many of which were specifically built for this kind of latency-insensitive training demand." The analyst added: "We have a data center development pipeline in the US with total nameplate capacity of 488GW, of which 170GW is what we consider credible. Of this 488GW pipeline, 36% is located in rural areas, and another 34% is in Tier 3 markets (such as West Texas). This means 70% of the development pipeline is built around training demand (or latency-insensitive inference demand)." According to this analysis, CoreWeave has the most prominent exposure. The analyst said: "We estimate that 25% of CoreWeave's currently operational power capacity in the US is located in Tier 3 and Tier 4 markets. In addition, about 74% of its contracted power capacity is also located in these markets." "Its backlog is mainly composed of 'take-or-pay' contracts, so we do not expect this portion to be threatened, but if the pace of training development slows, we may see a pullback in demand for rural power capacity that has already been contracted but not yet sold." By contrast, data centers located in metropolitan and small metropolitan areas appear best able to withstand the impact of a slowdown in training demand. This includes Equinix (EQIX.US), Digital Realty Trust (DLR.US), and Csquare (CSQR.US). Respectively, 95%, 92%, and 94% of these companies' data center capacity is located in large or small metropolitan areas. Bernstein gave CoreWeave an "underperform" rating with a target price of $74. At the same time, the investment firm gave Csquare, Digital Realty, and Equinix "outperform" ratings, with target prices of $27, $226, and $1,270, respectively. Bernstein noted that even if data centers shift from training-oriented to inference-orienteda shift necessary to enable more agentic workflowsmetropolitan-based data centers "remain the safest and also the most valuable."