CoreWeave's quarterly report hits record highs, but Bernstein and HSBC still maintain "sell" ratings

Cloud computing company CoreWeave delivered an almost flawless quarterly report but failed to convince the two most steadfast bears on Wall Street to change their stance.

Wallstreet Insights wrote that CoreWeave surged more than 27% intraday on Wednesday, marking the best single-day performance in over a year. The day before, the company released its Q3 financial report: revenue reached $3.66 billion, exceeding expectations, margins surpassed forecasts, annual revenue and operating profit guidance were both raised, and contract backlog orders reached $132.5 billion. CEO Michael Intrator called it a “breakthrough quarter.” This report temporarily eased market concerns over slowing AI infrastructure demand. Brian Mulberry, chief market strategist at Zack Investment Management, noted that if CoreWeave can maintain this level of execution for several consecutive quarters, market concerns will gradually subside.

However, Bernstein analyst Madison Rezaei and HSBC analyst Abhishek Shukla, having both reviewed the same financial report, did not change their ratings. Bernstein maintained an “Underperform” rating and a $54 price target; HSBC maintained a “Reduce” rating and a $39 price target — the latter is about 30% below Tuesday’s closing price. The core disagreement between these two analysts is not about the strength of the report itself but rather the three things this report fails to prove: pricing that cannot be justified by financial data, a financing gap independently estimated by both institutions at about $33 billion, and the influx of new market compute supply by 2028 (55GW), which is equivalent to creating 20 more CoreWeave (2-3GW each)—once this flood of supply arrives, can current high prices be sustained?

Bernstein: Bulls Need to Believe Three Things, None of Which Stand Up

Rezaei, in a rare move, acknowledged the strength of the report even while maintaining a Sell rating. She wrote at the beginning of her report: “CoreWeave delivered the best quarterly report we could hope for… Just like after last quarter, we are struggling with our ‘Underperform’ rating.” She even raised her earnings forecasts, narrowing the expected adjusted per share loss for 2026 from $5.14 to $3.54 and for 2027 from $6.60 to $3.20.

But the rating remains unchanged, with a long-term price target of $54 (equivalent to 22.5x forward adjusted EBIT), and she noted that if AI sentiment rebounds, the stock price could touch the $64–73 range.

Rezaei believes that the bull case relies on three assumptions being true at the same time:

First, data center supply remains tight, maintaining current high pricing. CoreWeave has chosen not to sell 2028 compute power for now, waiting for higher pricing opportunities. However, Bernstein’s own data show that the major cloud providers, emerging cloud companies, and SpaceX together have announced new capacity totaling over 55 GW by 2028 (75 GW before discounts), while CoreWeave’s share is only 2 to 3 GW. In order to sustain current prices, total market demand would have to multiply.

Second, GPU renewal prices remain high. Management indicated that contracts for more than 130,000 GPUs will expire in 2027 and another 170,000 in 2028. Renewals could respectively generate about $4 billion and $10 billion in annualized revenue, requiring no new capex. Bernstein estimates this implies a GPU rental price of around $3.81/hour, or $24 million per megawatt per year. However, the market price for the same GPUs 13 months ago was less than half this assumption, and at that time, the 55 GW of new supply had not yet entered the market.

Third, CoreWeave achieves genuine hyperscale and a self-reinforcing economic model. In Q3, interest expenses were $846 million, more than twice its adjusted operating profit; debt rose by $11.8 billion in a single quarter to $46.9 billion, with equity at only $4 billion; adjusted net profit is not expected to turn positive until at least the first half of 2028. Bernstein forecasts that adjusted operating profit in Q3 2027 will be about $1.02 billion, still below the $1.56 billion in interest expense for the same period. Net debt to EBITDA ratio has reached 7.3x.

HSBC: $20 Million per Megawatt Pricing Cannot Be Found in the Financials

HSBC’s analysis focused on the most explosive language in CoreWeave’s financial report: The company can now sign four- to five-year contracts at $20–25 million per megawatt per year, with short-term contracts even reaching $40 million; it raised prices twice in Q3 (25% in July, then another 10%), with further increases planned for Q4.

But HSBC estimates that the actual fee for CoreWeave’s historical contracts is about $12–13 million per megawatt per year; Bernstein divides reported revenue by average utilized power and finds that the actual value in Q3 was only $9.2 million, below $10.7 million two years ago.

Notably, Goldman Sachs’s own example contract—$1.1 billion per year for 1 GW—breaks down to exactly $11 million per megawatt, closely matching the bears’ estimates of the historical average, and about half of management’s latest quote.

Abhishek Shukla’s verification method is straightforward: If pricing really doubled, the asset turnover (revenue/value of installed equipment) should reflect this. In Q3, this metric was 30.5%, 28.6% in Q2, about 30.0% for all of 2025, and 34.8% for 2024. The revised guidance implies a Q4 end-run rate of about 32.6%, almost exactly matching HSBC’s previous model and on par with Oracle Cloud’s level.

HSBC’s conclusion is:

The so-called $20–25 million pricing is either sales messaging or indicates that per-megawatt construction costs have increased significantly; the higher price is simply covering higher costs, making little difference to shareholders.

In addition, HSBC questioned the better-than-expected margin. Q3 adjusted operating profit margin was 10.4%, well above the guidance of 6.5%, giving the appearance of a turning point. However, the company’s average has historically been about 16%, and its long-term target is 25%. HSBC argues that the expectation base was already suppressed by conservative guidance.

Financing Gap: Two Paths, One Number—$33 Billion

Management stated in the financial report that funding issues are basically resolved: $6.5 billion in cash and securities (only $4.2 billion unrestricted), nearly $5 billion in undrawn DDTL facilities, about $9 billion in customer prepayments and deposits expected in the next three months, and less than $5 billion in new parent-level financing needs before 2027.

Bernstein compared these sources of funds with a required $60 billion in capex to deliver the backlog orders and calculated an approximate $32.5 billion shortfall. HSBC reached a similar estimate from a different angle, concluding CoreWeave needs to raise about $33 billion in new financing by 2027, with undrawn credit lines halved in a quarter from $10 billion to $5 billion.

The calculations of the two institutions can be reconciled: management’s statement of “less than $5 billion” refers only to parent-level financing, with the remaining ~$30 billion expected to come from additional asset-level DDTL and customer prepayments. This means that the entire plan depends on both doors staying open:

• Lenders: The next DDTL facility is expected to be priced at 7.0%–7.5%, covering more than 90% of relevant capex. Getting this rate would be good news, but it also means that the equity buffer for each new data center is thinning, while the bears are precisely questioning the residual value of the collateral.

• Customers: The prepayment ratio for new contracts has risen from 15%–25% at the end of 2025 to 15%–50%. Bernstein points out that the reason enterprises sign multiyear cloud contracts in the first place is to convert upfront capex into operating expenses, but being forced into large prepayments shows compute power is scarce. HSBC adds that these prepayments may resemble Oracle’s practice and carry interest, which in essence is still a loan—just with the customer as creditor.

In addition, this quarter CoreWeave issued $4.2 billion in convertible notes and applied for an “At The Market” (ATM) plan for up to 35 million shares. Bernstein estimates that if the ATM is used, there will be a 2%–6% dilution impact on the share price.

Broader Skepticism: From D.A. Davidson to Chamath Palihapitiya

The bear camp is not limited to Bernstein and HSBC. Gil Luria at D.A. Davidson stated bluntly:

“Their interest expense greatly exceeds adjusted operating profit — this shows that the business income generated by these data centers is not sufficient to service the accumulated debt… It is a structurally challenging business model.”

Nick Del Deo at MoffettNathanson pointed out the policy risk in Texas (the incoming governor hopes to halt related projects) and that parent company debt yields 16%. UBS’s Karl Keirstead believes it will take several quarters to rebuild market confidence. Stifel cut its price target citing financing issues.

Goldman Sachs’ near-simultaneous “Top of Mind” report contained the most insightful skepticism from the technology business podcast “All-In Podcast” and venture capitalist Chamath Palihapitiya. He likened emerging cloud companies to airlines: GPUs are engines, data centers are the fuselage, power-supplied land is the airport, and one hour of compute is a perishable seat. He put breakeven utilization at about 70%–75% and warned that companies with long-term commitments on a certain generation of GPU cannot upgrade at will whenever Nvidia releases a new product. His bearish scenario requires no accidents:

“The bear case is largely a continuation of the status quo: enterprises continue to spend on AI, but the majority of spending remains at the experimental stage, rather than forming large-scale, repeatable, enterprise-level commitments. In this scenario, compute consumption never explodes upward. Emerging cloud companies will linger long term in a gray area of 75%–85% utilization — profitable, but not enough to pay off debt before hardware becomes obsolete.”

Palihapitiya also differentiates between cloud companies that own their land, electricity, and buildings versus those that lease. He calls those who lease “middlemen,” whose business model requires near-perfect operations. CoreWeave leases data centers; as of September 30, it had lease liabilities of $18.7 billion, up significantly from $8.2 billion at the start of the year.

Notably, Borges himself, in the same report, also pointed out the same risk as Bernstein’s core logic: “A lot of capacity will come online over the next two years—the key is whether demand is strong enough for emerging cloud companies to earn attractive returns on the new supply. If supply overshoots, compute prices may fall.”

Two Bears, One Rating, Starkly Different Numbers

It’s worth noting that Bernstein and HSBC have little consensus other than their ratings. HSBC is the relatively optimistic bear: it forecasts CoreWeave will be profitable by 2028, earning $2.58 per share in 2029. Even so, based on 2030 numbers, CoreWeave’s EV/EBIT (including stock compensation) would be 11.7x, higher than Oracle’s 6.6x, and slightly higher than Amazon, Meta, and Microsoft—each of which funds capex with free cash flow and does not issue CDS “on points.”

Bernstein is the more pessimistic bear: it predicts losses widening each year, reaching a loss per share of $8.09 by 2029.

As HSBC raised its 2028–2030 revenue, EBITDA, and operating profit forecasts, it cut its 2028 earnings per share projection by 15% to $0.82 (down from $0.96) due to higher capex from a larger backlog (2028 capex forecast up 27%), resulting in more debt, with interest costs 14%–20% higher than previously estimated. HSBC’s conclusion remains concise: “The price of growth exceeds the value it creates.”

The Core Issue: A Strong Report That Can’t Address the Long-Term Questions

No one denies this was a strong quarterly report. Bernstein’s Rezaei wrote at the end of her report: “Execution is improving steadily; the fundamentals remain an overhanging cloud.”

This quarterly report demonstrates that when compute is scarce, CoreWeave can raise prices, collect prepayments, and borrow at 7% interest. This has never been in doubt. What the bears question is what cannot be verified by a single quarterly report: Before a 55 GW supply surge enters the market, will scarcity persist long enough for a company with $47 billion in debt and only $4 billion in equity—and no free cash flow this decade—to repay its loans before the value of its chips is halved?

Goldman’s financing structure diagram consists of six links, of which the first five are functioning well. The entire equity story revolves around the sixth link—the residual value of GPUs four years from now.

The next critical test will come in a few weeks: the investment-grade DDTL facility, expected to be priced at 7.0%–7.5%, will enter the market soon. The final execution rate will be the most direct test of the entire bull thesis.