Nebius and CoreWeave AI Infrastructure Earnings Surge

About 70% of Nebius's Q2 2026 deals included customer prepayments, with the company expecting more than $9 billion in prepayments across 2026 ^[1]. That mechanic matters for reading the headline 454% year-over-year revenue growth figure ^[2]: prepayments accelerate cash and revenue recognition ahead of when compute actually gets consumed, so part of the reported jump reflects contracting velocity rather than workloads running today. Nebius also disclosed four AI-cloud contracts signed in the quarter averaging more than $1 billion each ^[2], a reminder that a handful of anchor deals can move the topline sharply in either direction. None of this makes the growth fictitious - money changed hands - but it explains why skeptics discount the multiple the market is attaching to it.

Nebius stock fell as much as 35% in mid-July 2026 as an 'AI infrastructure oversupply' thesis swept through neocloud names ^[3]. Weeks later, Morgan Stanley published a bottom-up model projecting hyperscaler compute capacity nearly quadrupling, from about 30.5 GW in 2025 to 116.6 GW by 2028, and framed the July selloff as a reaction to a bear case its own numbers didn't support ^[4]. CoreWeave reported first, on August 11, with revenue up 112% and a backlog near $104 billion ^[5]; Nebius followed a day later with its 34% single-day rally, and CEO Arkady Volozh said the company could sell its entire 2027 capacity under current commercial terms but is deliberately holding some back for higher-value deals ^[6]. The same underlying demand signal produced opposite stock reactions six weeks apart - which says as much about how narrative-driven neocloud valuations have become as it does about the demand itself.

Michael Burry disclosed a short position against Nebius on August 6 ^[7], and rather than covering after the 34% earnings pop, doubled it on August 12 ^[7]. His case isn't about demand - it's about accounting: he argues neoclouds extend the assumed useful life of GPUs well beyond their real two-to-three-year turnover and carry undisclosed off-balance-sheet liabilities, including uncommenced leases and purchase commitments ^[7]. That puts him at odds with Goldman Sachs, which initiated Nebius at buy with a $68 price target ^[8], though the skepticism isn't uniform: former Deloitte cloud strategist David Linthicum separately warned neoclouds may be 'delusional' about timing, expecting an inflection point in one to two years when it could realistically take five to ten ^[8], and industry consultants cited alongside the earnings coverage flagged that neoclouds' bare-metal-as-a-service economics remain structurally fragile despite the boom ^[8]. The disagreement isn't over whether demand is real - everyone concedes it is - it's over whether reported profitability and depreciation schedules honestly capture the cost of chasing it.

Meta anchors the exposure on both balance sheets: an up-to-$27 billion, five-year supply agreement with Nebius signed in March ^[9], plus an additional $21 billion committed to CoreWeave during Q2 ^[1]. Nvidia has a direct stake too, having bought pre-funded warrants for roughly 21 million Nebius shares for $2 billion in gross proceeds as part of a broader infrastructure collaboration ^[10]. The more notable shift is in who's showing up that isn't a frontier AI lab: quantitative trading firm Jane Street made a $6 billion compute commitment to CoreWeave during the quarter, and CoreWeave signed a new multi-year deal with Anthropic ^[11]. That customer diversification cuts against the concentration risk bears have pointed to, but it doesn't reduce the capital intensity - CoreWeave raised 2026 capex guidance to $35-39 billion ^[5]and Nebius spent $5.7 billion in a single quarter ^[2]- meaning both companies stay dependent on the prepayment and commitment pipeline staying open.