CoreWeave vs Nebius: Who Pays for the AI Data Center Boom?

A data center can be fully booked and still run into financial trouble before its first invoice. That is the central tension behind two of the AI industry's most closely watched public cloud operators: CoreWeave and Nebius.

Both have huge customer commitments and expensive infrastructure to deliver. Both increasingly use customer prepayments and loans backed by contracted compute to support construction. But their scale, accounting and financing structures differ enough that a simple ranking by backlog or capital expenditure is misleading.

CoreWeave and Nebius: comparison at November 2026

Read carefully: CoreWeave's backlog and Nebius's remaining performance obligations (RPO) are not automatically comparable measures. CoreWeave's backlog definition includes amounts beyond strict RPO. Nor can you compare CoreWeave's 2026 spending with Nebius's 2027 plan as if they were from the same financial year. These numbers are different views, not a league table of who is winning.

Four ways to pay for the same bank of GPUs

1. Equity: the owners pay

A company issues shares to investors, which brings capital without a fixed interest-payment schedule. Existing shareholders may be diluted. The money belongs to the company after closing, but investors take ownership risk.

An example: Nebius and NVIDIA announced a $2B strategic investment commitment in March 2026. This is a strategic capital arrangement, not the same as a multiyear capacity-purchase agreement.

2. Convertible notes: investors lend, with possible equity later

Convertible notes are debt securities that may convert into company shares under defined terms. Companies can raise very large sums while paying a coupon that reflects the conversion option. Debt still has maturity terms, and conversion may dilute owners.

Nebius closed $5.75B of convertible notes in August 2026. CoreWeave closed $4.2B of 2.875% notes due 2033 in September. Neither transaction should be described as a new ordinary venture round, and principal is not the same as future compute revenue.

3. Secured loans: the assets and contracts support the bank

A lender can take security over GPUs or contract-linked cash flows, subject to detailed terms and covenants. That can lower lending risk compared with completely unsecured financing, but the borrower still owes principal and interest. If equipment underperforms, lender protection is tested rather than guaranteed.

Nebius announced a $775M GPU-backed facility in July at SOFR + 2.50%. CoreWeave announced a $2.6B delayed-draw loan in August, designed to support shorter customer commitments within a longer financing term. These loans do not have identical collateral, tenor or pricing, so comparing headline spreads alone is not a fair measure of company quality.

A signed $3.5B or $2.6B loan commitment is also not proof that exactly that much cash was borrowed on the announcement date.

4. Customer prepayments: the buyer helps fund capacity

A customer can pay an advance to secure GPUs and priority service capacity. This effectively helps finance deployment before the supplier recognizes the associated future service revenue.

The economic attraction is obvious: less initial external financing may be required. The constraint is equally important: a prepaid contract remains an obligation to deliver. Paying for the infrastructure early does not mean it is profitable, that penalties can never apply or that the customer cannot have contractual remedies.

Not every customer advance pays for new construction. CoreWeave says more than $3 billion of its expected $9 billion in customer deposits is tied to infrastructure already installed and paid off. That can improve near-term cash generation without an additional equipment purchase, although the company still owes customers compute service in return.

In Q3, Nebius says prepayments on newly signed deals exceeded 100% of the capital expenditure associated with those deals. CoreWeave expects roughly $9B in near-term prepayments and deposits. One figure describes new-deal contract terms; the other is a forecast of receipts. They are not competing answers to the same metric.

A simple example: prepayments reduce funding needs, not construction risk

Imagine an AI cloud sells a three-year contract for $120 million of capacity. It needs $70 million of GPUs, network equipment and power-related investment. If a customer pays $50 million upfront, the supplier may need to raise only the remaining construction and working capital through its own cash or a loan. If a secured lender makes part of that available, the owner can preserve equity.

Now imagine delivery is delayed six months. The supplier still faces equipment, site and interest costs, but the contract may permit billing only when capacity is usable. Depending on the contract, the customer may receive credits, exercise remedies or renegotiate timing. A deposit does not remove these obligations.

The $120M, $70M and $50M amounts are illustrative, not figures from any actual Nebius or CoreWeave deal. The example shows why comparing prepayment receipts with recognized revenue is an accounting mistake.

Why capacity delivery is now a financial variable

Nebius's Vineland, New Jersey project illustrates the connection. A third-party site transitioning from temporary generation to Bloom Energy fuel cells is running behind plan, moving some Microsoft-linked capacity into Q2 2027. The New Jersey regulator independently fined the site's operator for unpermitted generators in September; that enforcement is documented, but it does not prove every delay had the same cause.

The contract and the GPUs do not automatically disappear when delivery slips. But a later start can defer revenue, create covenant pressure and increase the number of months before a deployment earns its cost of capital.

For CoreWeave the question is broader: its $132.5B backlog is economically useful only as capacity is built, becomes available and earns cash. The $47.5B debt load and rapidly rising interest expense mean that debt duration, refinancing and deployment schedules matter alongside sales growth.

Are higher GPU prices enough?

Both operators describe strong demand and favorable pricing. Nebius reported Q3 deals priced more than 20% higher per GPU-hour than Q2. CoreWeave reported rising revenue and better adjusted operating margins. Those are important signals, but neither proves future capacity will always sell at today's prices.

An investor needs to separate:

• Initial GPU acquisition cost: how much the systems cost.
• All-in project capex: power, networking, facilities, deployment and integration.
• Contract rate and duration: the committed revenue per unit of deployed capacity.
• Utilization and uptime: actual billable hours.
• Cash funding cost: interest, facility fees and repayments.
• Replacement needs: when hardware must be upgraded or replaced.

A short 'payback' statistic calculated from revenue, as Nebius publishes for certain deal cohorts, is not necessarily free-cash-flow payback to shareholders. The calculation can exclude the practical timing of cash, reinvestment and financing.

What Reddit readers are really asking

In recent discussions about the November earnings releases, five recurring questions emerged:

1. Why spend $40–45B when the next year's revenue guidance is $12–13B? Because capex purchases assets expected to generate revenue over multiple years. The return and timing remain to be proven.
2. Is 75–85% funding coverage already in the bank? No. It is a management expectation combining prepayments and financeability of signed contracts.
3. Why can CoreWeave lose $709M while reporting 62% adjusted EBITDA margin? Because EBITDA and the company-defined adjustments exclude important expenses, notably interest and depreciation.
4. Can short contracts safely finance long-lived GPU assets? Only if renewals, residual value and lender protections work; shorter demand visibility can increase risk.
5. Is a customer prepayment risk-free capital? No. It arrives against future services, and the supplier must still meet performance and contract terms.

These are common discussion themes, not statistically representative surveys or verification of individual commenters' financial claims.

What investors should monitor in 2027

What is the bigger signal?

The AI infrastructure boom has not stopped using capital markets; in 2026 it is combining equity, convertible bonds, equipment-backed loans, contract-backed loans and customer advances at very large scale. Customer credit can be a competitive advantage, particularly when buyers demand scarce capacity. It is not a substitute for engineering, permitting, power delivery or earning returns on expensive hardware.

For a broader list of public and private competitors, read Neocloud companies and stocks (2026). The GPU financing explainer compares actual credit ratings, loan spreads and collateral structures across CoreWeave, IREN, Lambda and Nebius.

Read the underlying company reports in Nebius's Q3 breakdown and CoreWeave's Q3 breakdown. For historical context see Lambda's GPU-backed Microsoft financing, USD.AI's $128.9M equipment finance and our 2026 AI infrastructure market map.