A backlog that looks enormous and a loss that looks worse
CoreWeave ended its June quarter with a revenue backlog of 104.2 billion dollars, and the company says another 25 billion dollars has been added since that count was taken. Quarterly revenue came in at 2.58 billion dollars, more than double the prior year's figure for the same period, continuing a growth trajectory that has made CoreWeave one of the most closely watched infrastructure companies in the current AI buildout. On paper, those two numbers together read as a company riding a genuinely durable wave of enterprise demand for raw GPU capacity at scale.
The net loss figure tells a different and considerably less comfortable part of the same story. CoreWeave reported a 626 million dollar net loss for the quarter, which the company's own disclosures describe as having grown larger year over year, even as the operating loss actually narrowed to just 49 million dollars over the same period. The gap between a shrinking operating loss and a widening net loss usually points squarely at financing costs, meaning interest expense on the debt used to build out the 51 data centers, 1,500 megawatts of active power, and roughly 1.1 million GPUs CoreWeave now operates across its footprint.
Customer concentration is a named risk, not a hidden one
CoreWeave has signed long-term capacity deals worth 22.4 billion dollars with OpenAI and 32 billion dollars with Meta Platforms, alongside additional commitments from Google and Anthropic and a 1 billion dollar strategic investment from the trading firm Jane Street. Those are genuinely credible, well-capitalized counterparties, and their presence on CoreWeave's customer list lends real weight to the backlog figure. But that same roster also concentrates CoreWeave's future revenue in a fairly small handful of relationships, which is precisely why the company has publicly set an internal target of keeping no single customer above 35 percent of total revenue going forward.
That stated target is worth watching closely over the coming quarters as it gets tested against reality. If OpenAI or Meta scales back spending for any reason, shifts a portion of its workloads to in-house custom silicon, or simply renegotiates existing terms downward, CoreWeave's revenue outlook moves with those decisions in a way that a more diversified hyperscaler's revenue base simply would not. Any enterprise signing its own capacity deal with CoreWeave is, by extension, implicitly exposed to the ongoing health of those other large contracts it has no direct visibility into.
Nebius is playing a different game: speed over scale
Nebius runs at roughly a tenth of CoreWeave's revenue and a little over half its capital expenditure, but its contracted power has grown just as aggressively on a relative basis: from just over 1 gigawatt in August 2025 to over 2.5 gigawatts by November, past 3 gigawatts by February 2026, and over 4 gigawatts by May, with an explicit company target of 5 gigawatts contracted by the end of 2026. Deferred revenue sits just under 6 billion dollars against total customer capacity commitments of more than 40 billion dollars already on the books.
What genuinely sets Nebius apart from its larger rival is its willingness to sign short-term, 3 to 6 month contracts specifically to capture what the company itself describes as a pricing opportunity in the 40 to 50 million dollar per megawatt range, even while its core underlying business still rests on more conventional 1 to 3 year agreements. For enterprise buyers, that flexibility can translate into genuinely better near-term pricing on raw compute, but it also means Nebius's own forward revenue visibility is considerably less locked in than a hyperscaler's would be, a factor that should weigh directly on how much multi-year operational reliance you build on top of it.
Neoclouds did not emerge from nowhere, and that history matters
Industry analysis of this sector has been unusually blunt about its origins: several prominent neocloud providers trace their physical infrastructure roots back to cryptocurrency mining operations that were explicitly repurposed to undercut hyperscaler GPU pricing once the AI compute boom took hold. That origin story is not, by itself, any kind of disqualifier for a serious enterprise buyer. It genuinely explains why neoclouds are able to offer such aggressive pricing on raw compute in the first place, since their underlying cost structures and institutional risk tolerance differ meaningfully from those of a large, diversified cloud provider.
It does mean, however, that the sourcing conversation around a neocloud contract needs to be structured differently from one with a hyperscaler. A hyperscaler's financial stability ultimately rests on a broad, highly profitable software and services business that happens to include cloud infrastructure as one part of a much larger whole. A neocloud's stability rests almost entirely on continued access to capital markets, ongoing chip allocation from suppliers, and the specific customers already sitting in its backlog, which together represent a fundamentally more concentrated set of dependencies underpinning any multi-year contract signed against it.
A practical due diligence checklist before signing
Before committing to a multi-year neocloud GPU contract, enterprise buyers should ask three specific questions that a typical hyperscaler negotiation might never even raise. First, what share of the provider's total backlog traces back to your largest direct competitor's workloads, and what contractually happens to your own capacity allocation if that particular customer relationship changes materially. Second, what is the actual gap between contracted power and currently delivered, fully operational power, since a wide gap between those two figures is itself a strong signal of execution risk hiding inside the provider's build-out timeline.
Third, ask directly and specifically about net loss trends and financing costs relative to revenue growth, essentially the same set of questions a lender would ask before extending meaningful credit to any company. None of this due diligence means neoclouds are an inherently bad choice for sourcing GPU capacity; their pricing and raw availability can be genuinely superior to hyperscaler alternatives in the right circumstances. It means the financial due diligence applied to the deal needs to be every bit as rigorous as the technical evaluation, because the real risk here sits in a fundamentally different place than it does when contracting with AWS, Azure, or Google Cloud.
