CoreWeave Sells AI Capacity at $40M Per MW Amid Rising Power Costs
Neocloud pivots toward customer prepayments to fund infrastructure
CoreWeave announced major changes to how it finances its AI data center expansion during its third-quarter earnings report, seeking customer prepayments and allocating capacity to short-term deals commanding over $40 million per megawatt annually. The cloud computing provider reported $3.66 billion in revenue alongside a $709 million net loss and $846 million in quarterly interest expenses. Facing rising yields in public debt markets, the company is adjusting its commercial terms to manage the escalating capital costs of powering artificial intelligence workloads.
Key details
According to third-quarter financial results disclosed by CoreWeave, the company is turning to secured debt facilities and upfront customer capital rather than relying on unsecured bond markets, where yields recently topped 15%. Chief Executive Officer Michael Intrator stated that the company expects roughly $9 billion in customer deposits and prepayments over the next three months. This includes more than $3 billion tied to contracts for computing capacity that has already been built, requiring no incremental capital expenditure for servers or facility hardware.
The company is capitalizing on acute compute scarcity by shifting substantial power allocation into short-term contracts. CoreWeave raised customer prices by approximately 25% in July and added another 10% increase later in the third quarter. In its earnings call, management reported selling short-term computing capacity at rates exceeding $40 million per megawatt per year, compared to between $20 million and $25 million per megawatt annually for standard four- and five-year infrastructure commitments. To finance this strategy, CoreWeave closed a $2.6 billion delayed-draw term loan facility specifically structured around short-term customer agreements.
Why this matters
The pricing dynamics at CoreWeave illustrate the immense financial pressure and capital intensity underpinning modern AI infrastructure. Delivering high-density clusters requires massive upfront investments in electrical substations, grid interconnections, backup generation, and advanced cooling systems. When interest expenses alone reach $846 million in a single quarter, compute providers must extract higher margins per megawatt to service their debt, directly increasing API and hosting costs for downstream software developers and enterprises.
Context
As utilities across the United States and Europe struggle to deliver gigawatt-scale interconnections, energized data center real estate has become one of the most constrained commodities in technology. Other specialized infrastructure operators, including Nebius and IREN, have similarly negotiated substantial upfront payments from major technology companies to underwrite construction costs. CoreWeave holds a total contract backlog of $132.5 billion, but executing on those commitments requires billions of dollars in physical plant before revenue can be recognized.
Risks and open questions
Selling compute capacity on shorter-term agreements at premium pricing exposes providers to significant market risk if AI demand moderates or if token economics compress margins for foundation model developers. While lenders value four- and five-year contracts for predictable cash flows, short-term contracts introduce periodic renewal risks. Furthermore, relying on customer prepayments locks in long-term service commitments that could constrain profitability if electrical tariffs, cooling water rates, and equipment replacement costs continue to rise.
What happens next
CoreWeave plans to deploy its $2.6 billion installment loan to build out additional high-density data centers while collecting anticipated customer prepayments over the winter quarter. Industry observers and financial markets will monitor whether enterprise buyers accept annualized rates above $40 million per megawatt, or if rising compute overhead accelerates a broader optimization wave in model training and inference efficiency across hyperscale cloud environments.
Source: TechCrunch Published on AI Usage Global, author: AUG Bot
