Earnings Call Highlights: CoreWeave (CRWV) | Q3 FY2026 — $3.66B Revenue, $132.5B Backlog, Old GPUs Reprice Higher, $9B in Prepayments Replace Unsecured Debt
1. Executive Takeaway: This Quarter CoreWeave Answered One Question — "Where Does the Money Come From?"
For the past year the market's doubt about CoreWeave was simple: revenue is growing, but every new block of capacity is built with borrowed money, and interest has been growing faster than revenue. On this call, management spent nearly half its time answering exactly that.
Three core conclusions:
The hardest stretch of the operating story is behind it. Adjusted operating income was $379 million, well above the top of the $200–260 million guidance; adjusted EBITDA was $2.26 billion at a 62% margin, the highest in five quarters. As new capacity comes online, revenue starts "growing on capital that has already been spent" — that is operating leverage.
AI compute is a seller's market, and even old GPUs are getting more expensive. CoreWeave raised prices across SKUs by about 25% in July, added another 10% later in the quarter, and pricing kept moving higher in Q4. Ampere and Hopper pricing rose 30–50% quarter over quarter, above their original contract prices.
The funding structure is switching tracks. The company said plainly it will not use unsecured debt at current yields, "nor do we need to." The three main funding sources from here: delayed draw term loans (DDTL) tied to customer contracts, customer prepayments, and operating cash flow. About $9 billion of prepayments and deposits arrives in the next three months alone.
The headline this quarter is not 168% revenue growth. It is CoreWeave showing it can keep building without high-yield unsecured debt.
2. Revenue and Financials: Operating Income Turns Positive, Net Loss Widens
Metric | Q3 FY2026 | Q3 FY2025 | Change |
Revenue | $3.661B | $1.365B | +168% YoY, +42% QoQ |
GAAP operating income (margin) | $207M (5.7%) | $52M (3.8%) | Q2 was a $49M loss |
Adjusted operating income (margin, non-GAAP) | $379M (10%) | $217M (16%) | Q2 was $128M (5%) |
Adjusted EBITDA (margin, non-GAAP) | $2.256B (62%) | $838M (61%) | +169% YoY |
Interest expense, net | $846M | $310M | +173% YoY |
GAAP net loss | $709M | $110M | Loss widened |
Diluted net loss per share (GAAP) | $1.27 | $0.22 | — |
Operating cash flow | $1.975B | $1.689B | +17% YoY |
Purchases of PP&E (cash) | $9.441B | $3.279B | +188% YoY |
Revenue backlog | $132.5B | $55.6B | +138% YoY |
▲ CoreWeave Q3 FY2026 key financials: revenue $3.66B (+168% YoY), revenue backlog $132.5B, GAAP operating margin 5.7%, net loss $709M, full-year revenue guidance raised to $13.0–13.3B (Image source: Simple Tech Trend original; data from CoreWeave Q3 2026 earnings release, 8-K Ex.99.1)
Three things to read from this table:
The operating leverage is real. Adjusted operating income nearly tripled from $128 million in Q2 to $379 million in Q3, while interest expense rose only about a third, from $640 million to $846 million. In the CFO's words, "that is the path to earnings" — operating income is now rising faster than interest.
But the absolute gap is still wide. Interest of $846 million is still 2.2x adjusted operating income, leaving an adjusted net loss of $529 million. The company also booked an $83 million tax provision despite the loss (a valuation allowance on deferred tax assets), an extra accounting drag.
Keep the CapEx definitions straight. The $14.5 billion Q3 CapEx cited on the call is accrual-based and above the top of guidance because some Q4 deliveries were pulled forward; cash purchases of PP&E on the cash flow statement were $9.44 billion. Construction in progress fell from $11.9 billion to $9.3 billion, meaning a large amount of equipment was actually energized in Q3.
STT estimate: how much repricing room is left per MW? Annualized Q3 revenue is about $14.6 billion. Divided by 1.7 GW of active power at quarter-end, that is roughly $8.6 million per MW per year — less than half the $20–25 million per MW on new 4–5 year contracts. Not all of the gap is "cheap old contracts"; capacity energized late in the quarter and not yet billing a full quarter also drags the average down. But the direction is clear: the unit price of CoreWeave's existing revenue is far below what it can sign today, and every expiring contract is a repricing opportunity.
3. Business Highlights: Old GPUs Still Getting Pricier, First to Bring Vera Rubin Online
3.1 Pricing: $40M per MW on Short Contracts, $20–25M on Long Ones
CoreWeave now sells capacity three ways:
Long-term contracts (4–5 years): new infrastructure at $20–25 million per MW per year, with hardware CapEx paid back in about 1–2 years.
Short-term contracts (3–6 months): more than $40 million per MW per year, with payback in 6–18 months.
Premium services: managed inference, launched only this year, already has more than $250 million of booked annual recurring revenue (ARR), and growth is constrained mainly by the capacity allocated to it.
Management was blunt: it could sell all of its capacity to buyers already in the queue but is "choosing not to," because bringing capacity to market later earns better pricing. 2028 capacity has not formally gone on sale, yet demand is already several times the available capacity; through the end of 2027, qualified demand is 10x available capacity.
This matches the logic we covered in NVIDIA at OCP 2026: every megawatt counts — the pricing unit of an AI cloud has shifted from "per GPU per hour" to "per MW per year," and power is the scarce resource.
3.2 Fleet Renewals: 130,000 GPUs Roll Off in 2027, 170,000 in 2028
This is the most interesting argument of the quarter. More than 130,000 GPUs come off contract in 2027, and about 170,000 more in 2028. Re-contracted at current pricing:
The 2027 expirations would bring nearly $4 billion of annualized revenue;
Including the 2028 expirations, nearly $10 billion of annualized revenue;
None of this is in the $132.5 billion backlog, and it needs no new CapEx — DDTLs are structured so the debt is repaid by the end of the contract and CoreWeave owns the equipment free and clear.
STT estimate: $4 billion divided by 130,000 GPUs is about $31,000 per GPU per year, or roughly $3.50 per GPU-hour. For the Hopper generation that is a strong price, consistent with management's 30–50% QoQ increase for Ampere and Hopper.
This directly challenges the market's worry about GPU depreciation lives: if four- or five-year-old cards can be re-leased above their original contract prices, the "six-year depreciation is too optimistic" argument loses its footing. But note this is a calculation at today's prices; 2027–2028 pricing is the real test.
CoreWeave's hedge is clever: if the industry builds fast, it grows with it; if the industry builds slowly, its paid-off fleet becomes more valuable.
3.3 Vera Rubin, Forge, Omni
Vera Rubin: CoreWeave was the first cloud to bring multi-rack NVIDIA Vera Rubin NVL72 deployments online. Its first production customer is Cognition, which according to the press release achieved 4.8x the token throughput of GB200 NVL72. Management also plans to deploy the Vera CPU on its own.
Forge: an AI development environment launched in September. 70% of the new cloud customers added in Q3 were already Forge customers — Forge is the on-ramp for enterprises.
Omni: extends the CoreWeave platform to third-party-owned infrastructure, so customers keep ownership of the hardware while CoreWeave earns software and operations revenue. The first customer is signed. It is a much less capital-intensive growth lever.
3.4 Power and Customer Mix
Active power exceeds 1.7 GW (more than 200 MW added in Q3), with a year-end target above 1.85 GW; contracted power is 4.6 GW, up more than 900 MW in the quarter.
More than 1.3 GW of contracted power is outside the U.S., spanning Canada, Europe and Asia; this quarter CoreWeave entered Indonesia and India and expanded in Denmark and Spain.
More than half of this year's new customers come from outside AI labs and tech; 10% of the Fortune 500 are customers; customers committing more than $100 million grew more than 2.5x year over year; and 10% of backlog comes from financial services.
4. Management Outlook: Full-Year Guidance Raised, a Path to Profit Through 2028
Item | Latest guidance | Prior |
FY2026 revenue | $13.0–13.3B | $12.4–13.2B |
FY2026 adjusted operating income | $1.275–1.375B | $960M–1.15B |
FY2026 CapEx | $37–39B | $35–39B |
Q4 interest expense, net | $1.05–1.15B | New |
Exit-2026 annualized run-rate revenue (ARR) | $20–21B | Raised |
Active power at end of 2026 | More than 1.85 GW | Reaffirmed |
STT breakdown of the implied Q4:
Full-year revenue minus the first three quarters' $8.31 billion implies Q4 revenue of $4.69–4.99 billion, up 28–36% QoQ.
Full-year adjusted operating income minus the first three quarters' $528 million implies Q4 adjusted operating income of $750–850 million, a margin of roughly 16–17% — a notch above the "low teens" Q4 margin flagged on last quarter's call.
But Q4 interest of $1.05–1.15 billion is still larger than adjusted operating income, so the net loss continues in Q4.
Reading the language:
"We expect adjusted net loss to reduce meaningfully in 2027 and to become adjusted net income positive in the first half of 2028 and for the full year 2028." — the first time a profitability timeline has been given, but it says "expect," and it is on an adjusted basis.
"We expect EBITDA less interest, which excludes the benefit of prepayments, to more than double in 2027." — the baseline is roughly $1.4 billion in Q3 alone ($2.26 billion minus $0.85 billion).
"Should we choose to focus only on our committed backlog, we would expect CapEx for 2027 to remain in line with our 2026 CapEx investment, while revenue and EBITDA grow significantly." — a message to the market that CapEx has brakes, and applying them does not jeopardize the 2027 ARR target.
5. Supply Chain and Customer Signals: Cost of Capital Is the Variable Optical Suppliers Should Watch
5.1 The Details of the Funding Switch
No more parent-level unsecured debt: none through 2027 at current yields. Market reports put the yield on CoreWeave's 2030 bonds at around 15%, which explains why.
The next DDTL is investment grade: launching within weeks, with multiple rating agencies indicating a high investment-grade rating, an expected yield of about 7–7.5%, and financing for more than 90% of the associated CapEx. Historical DDTL spreads ranged from SOFR plus 225 to 550 basis points, with investment-grade customers at the tight end and AI labs at the wide end.
Less than $5 billion of incremental parent-level capital through 2027; more than 75% of remaining committed 2027 CapEx is tied to investment-grade and enterprise customers.
The weighted average cost of debt is 7.9%, down 250 basis points year over year; more than $4 billion of debt has already been repaid through the DDTL mechanism.
During the quarter the company also raised about $4.2 billion in convertible senior notes, set up an at-the-market program for up to 35 million Class A shares, and closed the $2.6 billion DDTL 5.5 facility dedicated to short-duration contracts.
This is consistent with our call in The optical selloff is about the cost of capital, not orders: the bottleneck in AI infrastructure is shifting from "is there supply" to "who can afford the interest." CoreWeave wrote out its answer this quarter — it can, but only if every dollar of CapEx is backed first by a customer contract and a prepayment.
5.2 What It Means for the Optical Supply Chain
Demand is locked in: 4.6 GW contracted against 1.7 GW active leaves about 2.9 GW still to build. STT's rough estimate: at about 130–140 kW per GB300 NVL72 rack, 1 GW holds 7,000-plus racks and more than 500,000 GPUs; at roughly 2–3 800G/1.6T optical modules per GPU in the scale-out network, that is more than 1 million optical modules per GW. That excludes the data center interconnect needed for the 1.3 GW international footprint — CoreWeave also launched a global AI backbone linking its regions this quarter, which is demand for coherent optics.
But the pace depends on capital, not technology. CoreWeave's "success-based CapEx" means no contract and no financing, no purchase order. The CFO was clear that purchase orders are placed and pricing is locked when customer contracts are signed. So what optical module makers should watch is not CoreWeave's $132.5 billion backlog but whether the DDTL closes smoothly at 7–7.5% — that is the switch for the next batch of purchase orders.
A reverse risk suppliers should note: management's hedge is that "if industry construction slows, installed compute becomes scarcer and the old fleet more valuable." That is good for CoreWeave but the opposite for optical and component suppliers — slower construction means fewer new shipments. CoreWeave makes money in both scenarios; the supply chain makes money in only one.
5.3 Shifts in the Customer Base
An analyst raised a sensitive point: frontier AI labs are starting to lease sites and build compute themselves. CoreWeave's answer is to expand enterprise customers, stressing that more than 75% of committed 2027 CapEx is tied to investment-grade and enterprise customers. Customers named this quarter include Canva, Wayve, Substack and Serve Robotics, along with expanded relationships with OpenAI, Jane Street, Perplexity and Cohere.
This echoes what we saw in Microsoft's Q1 FY2027 earnings call: hyperscaler CapEx came in below expectations partly because "third-party compute deliveries slipped" — the delivery capacity of third-party compute providers now shows up directly in hyperscaler results.
6. Risks: Interest, Renewal Pricing and Regulation
Interest burden: Q4 net interest of $1.05–1.15 billion still exceeds adjusted operating income. If the new DDTL prices above expectations or SOFR does not fall, the profitability timeline slips.
Renewal pricing assumptions: the "about $4 billion in 2027, nearly $10 billion in 2028" renewal revenue is calculated at today's prices. If a wave of new capacity arrives in 2027–2028, old-GPU pricing may not hold.
Short-contract mix: short contracts lift margins but reduce revenue visibility. Management did not give a ceiling on the short-contract share.
Regulatory and political variables: the U.S. regulatory environment has shifted after the elections. Management said the overwhelming majority of its Texas power is online or fully permitted and under construction, and it does not expect an impact on backlog delivery; capacity beyond 2028 will still depend on the stance of new state governments.
Capital markets window: with an ATM program in place plus convertible notes, shareholder dilution is an ongoing cost.
7. Conclusion
CoreWeave's Q3 is a quarter where "operations passed the test, financials are still climbing." Operating leverage, pricing power and the renewal value of the old fleet all beat expectations, but interest still exceeds operating income and profitability waits until 2028.
Three metrics to track next quarter:
The actual pricing of the investment-grade DDTL: whether it lands at 7–7.5% and finances more than 90% of CapEx. This sets the 2027 build pace — and the order pace for the optical supply chain.
Whether the Q4 adjusted operating margin reaches 16–17%, and whether interest expense stays within $1.05–1.15 billion.
When 2028 capacity formally goes on sale and at what price, and the actual re-contracting prices for Hopper-generation GPUs.
In one line: CoreWeave has rewritten the story from "borrow money to build compute" to "customers pay first, then banks lend" — a safer model for CoreWeave, and a slower, more selective ordering rhythm for its suppliers.
Further reading: for hyperscaler CapEx pacing in the same quarter, compare Amazon's Q3 FY2026 earnings call and Meta's Q3 2026 earnings call.
This article is for technology and industry trend analysis only and does not constitute investment advice.
Comments