finIO

Markets Brief

← All briefs

CoreWeave's $104 Billion Backlog Lands With a $49 Million Operating Loss

The AI compute trade got its cleanest demand print yet this quarter — and a reminder that scale has not yet become profit.

August 12, 2026 AI & Tech

Demand is not the question anymore

CoreWeave's second quarter is the strongest single data point available on AI infrastructure demand: revenue of $2.6 billion, up 112% from $1.2 billion a year earlier, third-quarter guidance of $3.4–$3.6 billion, and a revenue backlog of $104 billion as of June 30 that excludes a further $25 billion of new customer commitments booked in the third quarter. Meta alone committed an additional $21 billion during the quarter, and Jane Street deepened its partnership alongside $1 billion in strategic investments. Citi called it "one of the cleaner quarters" since the IPO, citing stronger pricing power and better-than-expected margins; the stock was up 18% in premarket trade after closing Tuesday 26% higher year to date. Super Micro rose 7% on a sales forecast beat, and Foxconn reported a 35% jump in second-quarter profit — three independent reads pointing the same direction.

The asterisk is the income statement

CoreWeave's operating expenses also more than doubled, to $2.6 billion from $1.2 billion, marginally exceeding revenue and flipping $19 million of prior-year operating income into a $49 million operating loss. Full-year guidance of $12.4–$13.2 billion in revenue against $960 million to $1.15 billion in adjusted operating income sets the bar for whether the promised operating leverage actually arrives. The company has taken on significant debt to build, which means the backlog is both the bull case and the obligation.

Capital is rotating to the picks and shovels

The money is spreading to the layers around the compute. Blacksmith, which validates code before it ships, raised a $45 million Series B led by Peak XV at a $550 million valuation — roughly 9x the $60 million mark it carried less than a year ago — with customers up from more than 700 to more than 5,000; its pitch is that AI code generation creates a testing bottleneck, not eliminates one. IonQ posted $80 million of quarterly revenue, up 287%, with remaining performance obligations up 297% to $485 million, while still burning $120 million in adjusted EBITDA losses at a 62x price-to-sales multiple. Wintermute Trading is reportedly planning a $1 billion investment in AI data centers. Read together, these are three different bets that the buildout keeps compounding.

The valuation math has inverted

AMD trades at 63x forward earnings versus Nvidia's 24x, despite Nvidia carrying the greater substitution risk from hyperscaler ASICs — Google's TPUs, Amazon's Trainium and Inferentia, Microsoft's Maia. AMD's second quarter showed revenue of $11.5 billion, up 50%, with data center up 107% to $6.7 billion and free cash flow of $1.6 billion. Our reading: the spread prices catch-up potential into AMD while assuming the incumbent's multiple already reflects the ASIC threat — an unusually asymmetric setup for the two names most exposed to the same capex cycle.

Adoption and trust are the friction points

Marc Benioff, whose Salesforce carries a $162 billion market cap and whose stock is down 25.4% year to date despite a 15% bounce over the past month, says AI models still "cannot operate autonomously" and that his 15,000 engineers remain necessary — while backing June AI, an agentic startup deploying AI agents across enterprise systems, and having already cut Salesforce support headcount from 9,000 to about 5,000. Spotify will begin applying an AI Persona badge to profiles from mid-September, excluding flagged artists from personalised recommendations, after listeners made clear they object to discovering an apparent human is synthetic. Meta and 29 states head to court in the biggest test yet of youth social media litigation. The consumer-trust and legal layers are becoming a distinct cost line for platforms, separate from compute.

Sources

Not investment advice.