Wall Street Dissects Meta's Move to Sell Compute Power: Not an Industry Inflection Point, CoreWeave Faces the Most Pressure
Claire Weston
Meta's plan to sell surplus compute sent its stock up over 9%, adding roughly $98 billion in market cap in a single day, while CoreWeave fell 14% and Nebius dropped 17%. UBS, Morgan Stanley, and Bernstein all reached the same conclusion: this is not a signal of industry-wide overcapacity or an AI capex inflection.
"Selling compute = overcapacity" — why do all three banks disagree?
The market's shortest trade logic runs: selling compute = overcapacity = capex cuts. All three banks pushed back.
Morgan Stanley's model shows Meta expects to add roughly 2 GW and 3.5 GW of owned IT capacity in 2026 and 2027, respectively. Amazon and Google are adding on the order of 5 GW and 9 GW over the same period. This means → Meta leasing out a slice cannot move the overall hyperscaler buildout needle.
Bernstein cited reports that Google is limiting Meta's compute access due to its own capacity constraints. In plain terms = Meta is simultaneously chasing external compute and preparing to sell some — this looks more like reallocation across "different chip generations, use cases, and time windows" than a simple surplus.
Zuckerberg said publicly in October 2025 and May 2026 that selling compute "is an option we have." UBS therefore classified the news as not new.
What does selling compute mean for Meta's financials?
UBS noted that selling cloud compute can provide near-term revenue and margin cushion before Meta AI chatbots and Business Agents reach meaningful scale, easing concerns about flat EPS in 2027. UBS models Meta's diluted EPS at roughly $32.6 for 2026 and $33.0 for 2027 — essentially flat.
Morgan Stanley's sensitivity analysis is more granular: leasing 250 MW for one year at $40/watt could add roughly $2.97 to 2028 EPS — about 8% upside. Scaling to 500 MW–1,000 MW would amplify the effect further.
The market also drew a parallel to xAI leasing compute to Anthropic: 500 MW implies roughly $1.25 billion per month, or about $30 billion/GW/year. This reflects that high-quality compute remains tight in certain use cases, with extremely high implied returns.
Morgan Stanley stressed that Meta's valuation core remains first-party product innovation. In plain terms = selling compute can bridge EPS, but it does not automatically lift the valuation multiple.
Will capex actually come down?
Morgan Stanley's current model assumes Meta capex rises from $145 billion in 2026 to $175 billion in 2027 and $205 billion in 2028, premised on Meta building mainly for its own products.
This means → if Meta scales up external cloud services meaningfully, capex could face upward pressure — a full cloud business requires longer-duration data-center capacity, a more complex software platform, and enterprise delivery capabilities.
Bernstein drew a distinction between "temporary leasing" and "permanent cloud expansion," arguing the capex implications are fundamentally different and hard to judge before 2027.
Morgan Stanley also noted that Meta's Muse model family has not performed strongly on TerminalBench or SWE Bench Verified. In plain terms = the narrative that "Meta sells compute = Meta exits the model race" does not hold — to sell compute credibly, the models cannot fall behind either.
Why is CoreWeave hit the hardest?
Bernstein rates CoreWeave Underperform with a $67 price target; Meta is rated Outperform at $850. CoreWeave fell roughly 14% on the day.
The core tension: Meta is itself a major CoreWeave customer. Per Bernstein, Meta currently holds $35.2 billion in CoreWeave contracts — over one-third of CoreWeave's backlog. Add Microsoft's roughly $14 billion, and nearly half of CoreWeave's orders come from clients that could become competitors at renewal.
Near-term risk is not immediate — existing contracts are binding, and revenue or debt pressure may not worsen right away. But the long-term problem is harder: if customers build their own clouds and sell their own compute, CoreWeave's bargaining power at renewal drops sharply — it would face not just a demand-side party, but a potential supply-side competitor with capital, technology, and data-center experience.
Hardware sell-off — fundamentals or positioning?
Nvidia, Broadcom, and AMD all fell. Micron dropped 10.5%, Western Digital 10.6%, and Nebius 17%.
JPMorgan's trading desk split the debate into two sides: whether Meta's news represents a narrative shift, and whether crowded positioning and deleveraging amplified the decline. The desk leaned toward the latter carrying more weight.
Semiconductor and memory positioning was near the 100th percentile; long and short additions over the prior four weeks both ran at +2 standard deviations. Goldman Sachs' high-beta momentum basket fell 9% in a single day, after gaining 57% in the first half. This reflects a classic crowded-trade deleveraging pattern.
Software, crowded shorts, and China ADRs rose more than 1.4 standard deviations on the day — consistent with short-covering behavior.
What comes next?
The market is watching several potential reversal signals: whether Meta clarifies the scale and pace of compute sales, whether overseas AI application ARR — annual recurring revenue — accelerates, whether hyperscaler capex guidance continues to rise, and whether Q2 earnings beat expectations.
The timeline centers on July through August. In plain terms = this looks more like a watch-and-wait period than a settled inflection point.
Bloomberg's "Magnificent Seven" index posted a single-day return gap of +8% versus the Philadelphia Semiconductor Index — the widest since 2015. This means → the market is aggressively repricing the split between "platform giants" and "hardware supply chains," but whether the direction is confirmed still requires data.
Content is for reference only, not financial advice.