To The Moon

CoreWeave ( $CoreWeave, Inc.(CRWV)$ ) just delivered a quarter that perfectly captures the paradox of the AI infrastructure boom: the faster it grows, the more cash it burns.

Q3 revenue hit $3.661 billion (+168% YoY, +$1.086 billion QoQ), explicitly beating the high end of guidance. Adjusted operating income came in at $379 million, a 46% beat over the $200–$260 million guide. But GAAP net loss ballooned to $709 million, compared to a $110 million loss a year ago. Adjusted net loss widened to $529 million, from just $41 million last year.

The revenue and production engines are firing on all cylinders. But interest expenses and depreciation are obliterating shareholder profits. The market’s core debate is no longer about revenue growth. It is about whether AI compute pricing is sustainable, and whether CapEx can be covered by contracted cash flows. Demand and pricing are robust; financing remains tight. Operating leverage has arrived, but free cash flow and net income have not.

1. The Revenue Engine: Heavy-Asset Leasing and "Take-or-Pay" Contracts

FY2025 revenue was $5.131 billion, with a 71.7% gross margin. But technology infrastructure costs of $2.929 billion and interest expenses of $1.229 billion pushed the company to a $46 million operating loss and a $1.167 billion net loss. The gross margin is fat, but depreciation and interest are eating it alive.

Customer Concentration: In FY2025, Microsoft accounted for roughly 67% of revenue. The 10-K disclosed that the top two customers made up about 77% in 2024.

The OpenAI contract carries a maximum commitment of roughly $6.5 billion through May 2031 (or $11.9 billion through October 2030 under an earlier MSA).

The Meta commitment is up to $14.2 billion through December 2031. Revenue is highly dependent on a handful of AI hyperscalers. Diversification is the goal, but the concentration risk remains an undisputed fact.

The Revenue Model: About 98% of FY2025 revenue came from committed contracts, typically 1–6 years, mostly take-or-pay.

Customer prepayments are booked as deferred revenue, then recognized over the service period; on-demand business is billed monthly in arrears. FY2025 ended with $8.2 billion in deferred revenue and $60.7 billion in Remaining Performance Obligations (RPO). Prepayments are real cash, but they represent future service obligations.

The backlog is massive, but execution and continued customer compliance are the key variables.

2. Unit Economics: Volume, Pricing, and Margins

• Volume: Active power capacity grew from 850 MW at the end of 2025 to over 1.7 GW. However, the sequential addition was only 200 MW, down from 500 MW last quarter. Contracted power stands at 4.6 GW, with a revenue backlog of $13.25 billion. The Q3 revenue beat was driven more by capacity coming online at the end of last quarter, rather than new capacity added during Q3.

• Pricing: Long-term contracts are priced at $20–$25 million per MW per year; short-term contracts exceed $40 million. The company raised prices by roughly 25% in July, and another 10% late in the quarter. Legacy Ampere and Hopper GPUs saw a 30%–50% QoQ price hike. Analyst calculations show annualized revenue per active MW reached roughly $9.15 million, up from $8.24 million last quarter, confirming the rental pricing uptrend on the P&L. Current supply-demand favors the seller; whether short-term pricing holds depends on industry capacity additions.

• Margins: Q3 gross margin was 67.7%, down from 73.0% a year ago. Cost of revenue hit $1.183 billion (+220.6% YoY), growing 52.4 percentage points faster than revenue (+168.2%). After deducting cost and infrastructure expenses, the "residual gross margin" is roughly 12.4%, meaning depreciation and infrastructure are consuming the bulk of the gross profit. Adjusted operating margin was 10.4%, up 5.4pts QoQ (beating the target a quarter early) but down 5.5pts YoY. The QoQ improvement came from operating leverage and high-priced short-term contracts; the YoY decline came from depreciation, power, and data center costs. The expanding net loss is an interest problem, not a gross margin problem.

Product Metrics: The booked ARR for managed inference exceeded $250 million, constrained by capacity allocation. 70% of new cloud customers are already Forge customers, and about 10% of the US Fortune 500 are CoreWeave clients. This proves demand for inference products and Forge's role in enterprise acquisition. However, without product-level revenue and profit breakdowns, there is no proof that the software layer is actually profitable.

3. The Two Core Contradictions

Contradiction 1: Can legacy GPU renewals become low-capital-intensity cash flow?

Management stated that roughly 130,000 GPUs will come off contract in 2027, and another 170,000 in 2028. At current pricing, 2027 renewals could generate nearly $4 billion in incremental annualized revenue, with 2028 adding up to nearly $10 billion combined—without requiring new CapEx. The initial contract financing for these GPUs has already been paid off, keeping marginal renewal costs low.

If 2027 renewal pricing holds at $20–$25 million per MW per year, legacy assets become low-capital-intensity cash flow, improving FCF. Management's guidance is "nearly $4 billion," not a "no less than $4 billion" commitment. The falsification threshold: if renewal revenue falls below $3 billion, or long-term pricing drops below $20 million/MW/year, the economic viability of legacy assets is disproven, and the timeline for adjusted net income turning positive gets pushed back.

Contradiction 2: Can the financing structure cover CapEx and interest?

Management stated that asset-level delayed draw term loans (DDTL) can cover 70%–100% of corresponding CapEx. Since the IPO, pricing has been SOFR + 225–550 bps; the next tranche is expected at a 7%–7.5% yield, covering over 90% of CapEx. Over the next three months, the company expects to receive roughly $9 billion in customer prepayments and deposits, with over $3 billion tied to renewals of paid-off existing infrastructure. 2027 EBITDA minus interest is expected to more than double; adjusted net income is expected to turn positive in 1H 2028.

If the $9 billion in prepayments arrives on schedule and the next loan lands at 7%–7.5%, the funding gap is manageable. The risk: if prepayments fall short of $9 billion, or if spreads widen beyond 550 bps, equity financing and dilution risks rise sharply.

4. Key Metrics to Track

Short-term view: Q3 changed the expectation of "whether operating leverage can materialize"—the answer is yes. But it did not change the expectation of "negative free cash flow and expanding net losses."

Long-term view: Watch whether the CapEx model based on successfully signed contracts, asset-level DDTLs, and customer prepayments can push adjusted net income and FCF into positive territory by 2028. Until then, this remains a "strong operations, weak equity cash flow" setup.

Three Falsification Metrics:

1) Q4 revenue of at least $4.69 billion, and adjusted operating income of at least $747 million. The FY2026 revenue guide is $13.0–$13.3 billion; with $8.314 billion YTD, Q4 needs $4.686–$4.986 billion. The adjusted operating income guide is $1.275–$1.375 billion; with $528 million YTD, Q4 needs $747–$847 million—double Q3's $379 million. A miss below the low end means the full-year guide is busted.

2) 2027 legacy GPU renewal incremental annualized revenue near $4 billion, with long-term pricing no lower than $20 million/MW/year. The threshold: less than $3 billion, or pricing below $20 million/MW/year.

3) Q4 Adjusted EBITDA to interest expense ratio of at least 2.0x (analyst estimate). Q3 was $2.256 billion / $846 million = 2.67x. Q4 interest guidance is $1.05–$1.15 billion. If coverage drops below 2.0x, interest coverage deteriorates, and the financing structure contradiction escalates.

4) Actual revenue realization from managed inference and Forge. Right now, we only have booked ARR >$250 million, 70% of new cloud customers from Forge, and ~10% of Fortune 500 clients. There is no product-level revenue/profit breakdown. Next quarter, if they disclose Forge or inference revenue and backlog attachment rates, it will prove the software layer is not just bringing customers but also profits. Another soft metric: can top customer revenue concentration drop from 67% in FY2025 to below 50%? If not, customer concentration remains a core risk.