CoreWeave CRWV: $12 Billion OpenAI Compute Deal and the Leverage Question, September 2026

OpenAI reportedly signed a $12 billion compute capacity agreement with CoreWeave—ratifying the specialist neocloud model even as a $4.2 billion convertible note offering and a Rothschild Redburn sell rating at $54 underscore the leverage risk embedded in a business running 201% capex intensity.

When OpenAI reportedly signed a $12 billion compute capacity agreement with CoreWeave, it did more than add a contract to an already formidable order book. The deal ratified a division of labour reshaping the AI-infrastructure stack: even the most capitalised frontier labs are routing substantial inference workloads through specialist GPU cloud operators rather than owning every rack themselves. For CoreWeave, which went public earlier in 2026 and has since assembled a contract backlog exceeding $104 billion—reportedly expanded to over $129 billion in the months since—the OpenAI agreement signals that the neocloud model's position at the centre of the inference build-out may be more durable than its critics allow.

The financial foundation supporting that positioning has been growing at speed. Second-quarter 2026 results showed revenue of $2.58 billion, up 112% year-over-year, with customer concentration deepening rather than diversifying: Meta's contract has reportedly expanded to $21 billion, and all four major AI laboratories are now under contract, according to coverage this month. That breadth hedges against the risk of any single customer pivoting to in-house compute—though it does not eliminate it. Slicast's XBRL-derived compilation puts FY2025 capital expenditure at $10.31 billion against revenue of $5.13 billion, a capex intensity of 201%, and the highest absolute investment among neoclouds tracked. That ratio, which exceeded revenue by two to one, describes a business that must continuously outrun its own balance sheet to maintain the hardware-generation lead that customers pay for.

September has brought a concentrated burst of capital-markets activity to finance the next phase. CoreWeave's September 17, 2026 filing announced a planned $3 billion convertible senior notes offering; a September 22, 2026 8-K confirmed the deal was completed at $4.2 billion—upsized on strong demand—at a coupon of 2.875% and maturing in 2033. The company simultaneously registered an equity distribution facility covering up to 35 million Class A shares. Together, these instruments extend the runway but deepen the leverage profile. Coverage this month places the company's backlog at approximately $104 billion against debt of roughly $72 billion—a ratio that holds so long as contract revenue drawdown stays on schedule, but that leaves limited buffer if lease economics shift or customer throughput falls short of commitment.

On the infrastructure side, CoreWeave is already mid-transition to the next hardware generation. It brought its first multi-rack NVIDIA Vera Rubin NVL72 cluster online at its Livingston facility this month, following Dell's delivery of what the companies described as the world's first Vera Rubin NVL72 server rack. Reported separately, HRT selected CoreWeave to provision Vera Rubin accelerators, adding a quantitative-trading client alongside the AI-lab anchor tenants. Power supply is expanding in parallel: Galaxy Digital's Helios campus completed Phase I delivery of 133 megawatts of critical IT load to CoreWeave this month, with the full campus carrying a planning target of 3.6 gigawatts; Blockfusion reportedly signed a 15-year lease with CoreWeave in Niagara Falls, New York. Coverage also indicates that Q3 new-signed compute pricing has risen to approximately $40 million per megawatt—a level that, if sustained, would support margin improvement as capacity scales.

The bear case is stated most sharply by Rothschild Redburn, which downgraded CoreWeave to sell this month with a price target of $54—against a current price of $90.13—citing leverage and rental rates as the core concerns; the firm argued that the credit market is 'sending the more important signal.' That framing is not without basis in the numbers: a 201% capex intensity means the company perpetually finances tomorrow's infrastructure with today's contracts, and any sustained softening in AI training demand—flagged in recent coverage as a source of uncertainty—could widen the gap between backlog commitments and actual revenue realisation. Competitive dynamics are also shifting: Lambda's B200 pricing reportedly undercuts CoreWeave on spot rates; Meta has disclosed plans to scale its own compute capacity; and SpaceX has entered the infrastructure market, according to earlier coverage. CoreWeave's structural advantage—long-duration contracts locked in at scale—provides insulation, but does not render the company immune to a repricing cycle.

Three concrete signals will determine which of these narratives prevails. First, the trajectory of new-contract pricing: the reported Q3 rise to approximately $40 million per megawatt is encouraging, but whether that level holds as Vera Rubin supply broadens will reveal how durable the pricing premium is against an increasingly competitive spot market. Second, free cash flow conversion against the capital programme—the single quarter ending March 2026 saw $7.70 billion in capital expenditure, the largest period in Slicast's eleven-quarter XBRL dataset—which will show how quickly the company can shift from a growth-capital mode toward self-financing. Third, the behaviour of its largest customers: OpenAI's reported $12 billion agreement and Meta's reported $21 billion commitment provide multi-year revenue visibility, but that backlog is only as durable as those counterparties' own infrastructure plans—a meaningful shift toward in-house compute at the frontier labs would challenge the demand foundation on which CoreWeave's entire capital structure rests.