By Grace Kim

CoreWeave ended the June quarter with a $104.2 billion revenue backlog and has added roughly $25 billion more since, the GPU rental firm disclosed this week. The figure dwarfs current billings and recasts neoclouds as capital vehicles funded by long-dated AI demand rather than short-cycle utility providers.

Second-quarter revenue grew 2.1 times year over year to $2.58 billion. CoreWeave cut its operating loss to $49 million, expanded its net loss to $626 million, and spent $9.4 billion on capital expenditure. Cash and equivalents stood at $6.41 billion at quarter close.

What separates neoclouds from hyperscalers?

Neoclouds are GPU-native compute providers formed largely after 2022, some out of bitcoin mining farms, to undercut the major clouds on accelerated compute pricing. Many of those same major clouds now rent capacity back from the neoclouds they once threatened to displace. The irony reflects a market short of allocations rather than a structural reversal. With GenAI demand still outrunning supply, pricing power now sits with whoever holds the chips.

How concentrated is CoreWeave's book?

CoreWeave has set a target ceiling of 35 percent of revenue from any single customer in the near term. Long-term contracts cover OpenAI at $22.4 billion, Meta Platforms at $32 billion, plus Google and Anthropic. Hedge fund Jane Street committed $1 billion strategically, a structure that presumes Jane Street will rent CoreWeave iron alongside its financial stake. CoreWeave's 2022 through 2025 history was dominated by Microsoft as the anchor customer, and the multi-year OpenAI and Meta agreements reset that concentration.

How large is the physical fleet?

CoreWeave exited Q2 with 51 datacenters worldwide and 1,500 MW of active installed power. Cross-industry analysis puts the GPU count at roughly 1.1 million chips across the deployed base. Contracted power for buildouts through Microsoft and the model builders stands at 3,700 MW. During the quarter, the operator became the first to bring up an Nvidia "Vera-Rubin" NVL72 rack-scale system, drawing on the tight partnership that made CoreWeave Nvidia's preferred neocloud counterweight to the hyperscalers building their own accelerators.

What is Nebius doing differently?

Nebius, the Netherlands-domiciled spinout of Russian internet conglomerate Yandex, runs at one-tenth CoreWeave's revenue scale but spends roughly half the capex. Deferred revenue sits just shy of $6 billion. Customer capacity commitments exceed $40 billion, with Microsoft and Meta Platforms widely believed to anchor the book. Nebius has framed 2027 as the GenAI build-out window it must reach before its peers lock down allocations.

The company's contracted power target tells the steepest story of the year. Nebius guided above 1 GW in August 2025, lifted the figure to above 2.5 GW in November 2025, then to more than 3 GW in February, more than 4 GW in May, and 5 GW for delivery by year-end 2026 at quarter close. That is a fivefold expansion in twelve months.

How is pricing shifting at the short end?

Nebius has begun signing contracts of three to six months duration, opening a price band the company described to Wall Street as a "price opportunity in the $40 million to $50 million per MW range." The core book still runs one to three years, where contracted power commands lower headline rates than the emerging spot market. Short-cycle pricing will test whether capacity stays scarce once Vera-Rubin systems hit general availability at scale.

What comes next?

Both firms are spending faster than they bill, accepting widening losses today to convert contracted power into billable GPUs before allocations tighten further. The first Vera-Rubin NVL72 deployment at CoreWeave this quarter and the fivefold Nebius power ramp set up the 2027 capacity race that will determine which neocloud captures the next leg of model builder demand.