美国通信基础设施：Neoclouds（CoreWeave）：市场传闻——关于首次覆盖的顶级投资者反馈

Neoclouds (CoreWeave): Word on the Street...top investorfeedback on our initiation We initiated coverage on US Communications Infrastructure just over a month ago, includingan Underperform on CRWV (PT $56), our noisiest call. Since then, we’ve spoken with 70+investors, the company, and our industry contacts. Here are the top responses we've heard(with our opinions and justification): Madison Rezaei+1 917 344 8622madison.rezaei@bernsteinsg.com Nancy Wu+1 917 344 8545nancy.wu@bernsteinsg.com •The Bulls:“You’re nuts. How can you possibly imagine a slowing of demand forneoclouds? CRWV has sold out all of their capacity and says they will continue to do sowell into the future. This is fundamental to AI growth and development.” •The Bears:“We agree with your thesis and think it’s wild that there are so many peoplebacking this. But where is the bottom? How do we time it?” •The Cloud Optimists:“Traditional cloud is a great business. Do Neocloud margins havethe potential to be similar at scale? If you ignore the questions around who should win, isthis a valid concept?” •The Balance Sheet (BS) Detectors:“CRWV talks a lot about their borrowingsophistication and ability to bring down their cost of capital. Is that valid? What ishappening behind the scenes with the debt?” •Team “In Jensen We Trust”:“There’s no way NVDA lets this fail, right? Why would theynot backstop, especially given CRWV’s scale relative to their own - it’s nothing.” •The Neocloud Negotiators:“We hear you on some specifics around CRWV, but is thisunique to them or pertinent to all neoclouds?” While we very much respect the opinions of the shrewd individuals we disagreewith, our take on CRWV remains unchanged.Even if you believe our forecast on AIdemand is meaningfully over-conservative, we do not believe CRWV is likely to be thelong-term beneficiary of that upside: hyperscalers are too big and too savvy. In termsof timing, we do not believe the bottom falls out immediately - we model CRWV signingan incremental $35B in contracts over the next two years. We begin to diverge fromconsensus in our revenue forecasts in 2027 and accelerate into 2028. Market sentimentmay well shift before then—this is a notoriously momentum-driven stock. Our Underperform is more a structural revenue call than anything else. We believe CRWVshould be able to execute against their guides and may reach 25-30% operating marginsover the long-term and bring down their overall cost of capital into the ~7% range. Will NVDA prop up the company? They obviously could—the CRWV investment is penniesfor a company with a market cap in the trillions. Our view is that NVDA invested in CRWV(and NBIS) to seed the ecosystem and potentially push the hyperscalers along; we don'tlove this as the sole insurance policy over the long term. On other neoclouds, check out ourcomparison piece. We only cover CRWV, but our colleagues are more positive on IREN. BERNSTEIN TICKER TABLE INVESTMENT IMPLICATIONS We are Underperform on CRWV with a price target of $56. We value the company using an EV/Adj. EBIT multiple (28.4x)on 2027 E Adj. EBIT per share of $5.31.Our case hinges on our longer-term revenue expectations for CRWV, whichare significantly lower than street expectations.In general, we question CRWV’s ability to continue to sign large-scalehyperscale deals as data center supply becomes more readily available (2028+) and are concerned that hyperscalers are likelyto attempt head-on competition with neoclouds for enterprise business. We disagree with CRWV’s premise that its existing software moat will hold up to hyperscaler attacks, especially givenhyperscalers’ scale and ability to spend. We do not believe the structural concerns abate with increased AI demand, though thetiming for our call may be pushed further out. Importantly, we do not believe the tide will shift for CRWV in the very near term and anticipate them signing an incremental$35B of deals between now and the end of 2027. We are roughly in line with consensus for 2026 expectations. By 2027, webegin to diverge, and by 2028, we are more than 30% below consensus revenue estimates. DETAILS RESPONSE TO THE BULLS There is indeed significant AI demand - we’re modeling it at ~60 GW by 2030, but it could easily be twice that, particularly if theLLMs find their way towards profitability (or the market tolerates unprofitability for a longer time). Nobody knows. Here’s whatwe’re banking on: •Power efficiency.The eye-popping high-case GW estimates literally cannot be satisfied by today’s building constraints.Instead of remaining in what is extreme supply constraint forever (and stomaching very high per-token costs), we have strongconviction that players up and down the value chain will find ways to dramatically increase efficiency and do more with less. •Hyperscaler self-builds and new leases.Hyperscalers have announced ~30GW of self-builds through 2030, ~23 GW ofwhich should be online by 2028. Additionally, there is ~70 GW of capaci