CoreWeave reported third quarter 2026 revenue of $3.66 billion on November 10, up 168% year over year, and a revenue backlog of $132.5 billion, a record high for the specialized AI cloud provider.
The company's earnings call, transcribed by Aol, laid out a system that now behaves less like a traditional infrastructure vendor and more like a contract-execution engine: committed demand in, financed capacity out, cash flow returned as the hardware amortizes. For anyone tracking the emerging architecture of computational sovereignty, the numbers describe a governance model in which capital allocation is bound to signed customer contracts rather than speculative build plans.
The Q3 2026 numbers in full
Revenue of $3.66 billion exceeded the high end of management's own guidance. Adjusted operating income reached $379 million against a guided range of $200 million to $260 million, with adjusted operating margin crossing 10% one quarter ahead of the plan management set at the start of the year. The margin trajectory is steep and legible: roughly 1% in Q1, 5% in Q2, above 10% in Q3.
Adjusted EBITDA came in at $2.3 billion, a 62% margin and the highest in five quarters. Operating cash flow reached $2 billion. The company still posted a net loss of $709 million, compared with a net loss of $110 million a year earlier, driven largely by interest expense of $846 million as debt levels rise to fund the build-out.
Active power reached 1.7 gigawatts, doubling the operational IT load footprint since the end of 2025. Capital expenditures totaled $14.5 billion, above the guided range, as the company pulled forward some deliveries originally scheduled for Q4. Construction in progress fell to $9.3 billion from $11.9 billion in the prior quarter, a signal that inventory is converting to revenue-generating capacity rather than accumulating.
Financing as a distributed protocol
Chief Executive Officer Michael Intrator described a financing architecture with a single operating principle: match every dollar of capital to a specific customer contract, and let the contract's cash flow service and retire the debt. The weighted average cost of debt fell to 7.9%, down 250 basis points year over year, a figure also reported by Benzinga.
CoreWeave has already paid down $4 billion in principal, including nearly half of its first two delayed draw term loans, which are its most expensive facilities. Management stated it will not issue parent-level unsecured debt at current yields through 2027. Instead, asset-level secured facilities will carry the load: more than 75% of remaining 2027 committed capital expenditures are tied to investment-grade and enterprise customers, and the company expects to launch a new investment-grade DDTL in the coming weeks, pricing at a spread implying a 7% to 7.5% yield over SOFR while financing more than 90% of associated capex.
Customer prepayments are the second pillar. Approximately $9 billion in prepayments and deposits will be collected over the next three months, including more than $3 billion tied to recontracting existing, fully amortized infrastructure at a premium. That $3 billion carries no debt service and no new capital expenditure. It is pure protocol output: hardware that completed one contract cycle now re-enters the market as yield-bearing capacity.
Pricing power and the renewal flywheel
Short-term capacity is now selling at up to $40 million per megawatt annually, against $20 million to $25 million per megawatt for new four- and five-year contracts. Intrator said pricing for prior-generation NVIDIA Amperes and Hoppers rose 30% to 50% quarter over quarter, above the levels of their original contracts.
The renewal schedule quantifies the flywheel. More than 130,000 GPUs come off contract in 2027, with roughly 170,000 more in 2028. At current prices, 2027 renewals would generate approaching $4 billion in incremental annualized revenue with no new capex, rising to nearly $10 billion when 2028 renewals are included. Management characterized qualified demand as ten times available capacity through the end of 2027.
Payback periods compress accordingly: six to 18 months for short-term contracts, one to two years for long-term hardware commitments at current pricing.
Platform layer: Forge and the enterprise migration
CoreWeave introduced Forge in September, a unified environment for AI training, inference, evaluation, and agent development. Adoption metrics suggest it functions as the on-ramp the company intended: 70% of new cloud customers added in Q3 were already Forge users. Managed inference has already booked $250 million in annual recurring revenue, with growth constrained mainly by the capacity allocated to it.
The customer base is diversifying beyond AI labs. Ten percent of Fortune 500 companies are now CoreWeave customers, financial services accounts for 10% of backlog, and the number of organizations committing more than $100 million grew 2.5-fold over the past year. More than half of new customers this year came from industries outside AI and technology, including the Chan Zuckerberg Initiative in computational biology, Hippocratic AI in patient care, and Nissan and Aston Martin Aramco Formula 1 in engineering simulation.
CoreWeave also claims a first-mover node on the next hardware generation. Intrator said the company was