Bernstein Names CoreWeave as Top Casualty in AI Slowdown Warning

Wall Street has long treated AI infrastructure stocks as a single interchangeable trade, rising and falling together on demand headlines. That pattern broke on Monday, Sept. 14, when Bernstein not only flagged sector-wide risk from a potential AI slowdown but also named the single stock poised to lose the most: CoreWeave.

The warning traces back to an essay by Anthropic CEO Dario Amodei titled “We Must Pace the Frontier,” in which he argued that unchecked AI progress, particularly as systems begin to self-improve, is outpacing the industry’s ability to safely test and control the technology. OpenAI CEO Sam Altman and Elon Musk endorsed the call within hours.

The timing complicated Amodei’s message. Anthropic confidentially filed for an IPO in June and is widely expected to list within weeks, according to CNBC, meaning the same executive urging restraint is also asking investors to fund faster growth.

The contradiction unsettled a market that had priced years of uninterrupted spending into valuations. CoreWeave Inc. (CRWV), the cloud provider renting AI computing capacity to labs and enterprises, saw shares fall roughly 8% on Sept. 14. They failed to recover from an early morning slide, closing down 6.75% at $82.98.

The drop reflects how exposed investors now see CoreWeave’s business model if the industry slows down. Bernstein analyst Madison Rezaei said the exposure stems from geography: about 25% of CoreWeave’s active U.S. power sits in Tier 3 and Tier 4 rural markets built for training, not low-latency inference. Another 74% of contracted but inactive power is concentrated there as well. Training tolerates distance from users, while inference does not, so a slowdown would hit demand for power CoreWeave has not yet locked into contracts. Its existing backlog remains largely protected by take-or-pay terms.

Bernstein rates CoreWeave Underperform with a $74 price target, implying significant downside from current levels. Wall Street overall remains more constructive, with an average target near $139. The company raised its 2026 capex forecast to $35 billion to $39 billion, and its revenue backlog swelled past $104 billion, but it carried roughly $35 billion in debt as of the end of June.

Metro-based data centers look safer to Bernstein

Equinix, Digital Realty, and Csquare sit on the opposite side of Bernstein’s split. Each keeps more than 90% of its U.S. capacity in major or minor metro markets, at 95%, 92%, and 94%, respectively. Bernstein rates all three Outperform, with price targets of $1,270 for Equinix, $226 for Digital Realty, and $27 for Csquare.

Metro locations matter because inference, the work that runs AI products after training, needs to sit close to users and customers. Bernstein called those markets the safest and most valuable part of the sector as it shifts toward inference.

The sell-off reached far beyond data centers

The Sept. 14 decline was not confined to CoreWeave or its direct rivals. It spread across companies that supply and connect AI infrastructure. Neocloud peers dropped alongside CoreWeave, with Nebius down 8%, Applied Digital down 6%, DigitalOcean down 5%, and IREN down 4%. Networking equipment makers fell sharply, with Lumentum down 8%, Ciena and Coherent each down 7%, and Arista Networks down 5%. Chipmakers tied to the AI buildout slid, too, as Marvell and Arm each fell 8%, AMD and Intel dropped 6%, and Broadcom fell 4%.

The AI trade just stopped being one basket

For most of this cycle, infrastructure stocks moved together, rising and falling on demand headlines rather than balance sheets or site maps. Bernstein’s note marks a shift toward pricing where a company’s capacity physically sits, not just how much of it exists.

That distinction will matter more if Amodei’s call for deliberate pacing gains traction. “We must slow the pace at which we improve the capabilities of AI models,” he wrote. A slower training cycle would not erase AI demand. It would reallocate that demand toward inference, and inference rewards proximity to customers over raw scale.

The pattern echoes the telecom fiber buildout of the late 1990s. Demand for bandwidth was real, but timing and location errors sank operators years before traffic caught up to the capacity.

Investors who spent the past two years buying AI infrastructure as a single trade may now need a map, not just a thesis.

Source link