CoreWeave just posted the best quarter in its history. Revenue hit $3.66 billion, up 168% year over year, and the backlog climbed to roughly $132.5 billion. In the same quarter it paid $846 million in net interest and lost $709 million.

That's the whole CoreWeave story in two lines. It's the biggest independent AI cloud, or "neocloud," renting NVIDIA GPUs to Microsoft, OpenAI, Meta and a growing list of others. And it's financing that build with one of the largest debt stacks in tech.

This guide covers what CoreWeave actually sells, how big the GPU fleet and power footprint are, every major contract, how the debt works, what happened with Core Scientific, how it compares with rivals, and the honest limits investors and buyers should weigh. Last updated: November 21, 2026.

Key Takeaways

CoreWeave at a glance

Here's the short version, pulled from CoreWeave's own filings and its Q3 2026 earnings release.

ItemDetail
Founded2017, as a crypto miner; now an AI cloud
HeadquartersLivingston, New Jersey
ListingNasdaq: CRWV, IPO March 2025
CEOMichael Intrator (co-founder, chairman)
Q3 2026 revenue$3.661 billion (Q3 2025: $1.365 billion)
2026 revenue guidance$13.0 billion to $13.3 billion
Revenue backlogAbout $132.5 billion (Sept 30, 2026)
Active powerOver 1.7 GW; more than 1.85 GW targeted by year end
Contracted powerOver 4.6 GW, including more than 1.3 GW outside the U.S.
GPU supplierNVIDIA only (per the 10-Q)
2026 capex guidance$37 billion to $39 billion

Source: CoreWeave Q3 2026 results and Form 10-Q for the quarter ended September 30, 2026.

💡
What CoreWeave sells, in one line: dedicated, liquid-cooled NVIDIA GPU clusters plus the software to run them, mostly on multi-year take-or-pay contracts. The 10-K says committed contracts typically run one to six years and require payment regardless of utilization.

What CoreWeave actually is

CoreWeave started in 2017 mining crypto on GPUs. It pivoted those GPUs, and its relationship with NVIDIA, into renting compute for AI training and inference. The company now calls itself "The Essential Cloud for AI" and says it's trusted by 9 of the 10 leading foundation model providers.

The label you'll see most often is neocloud. That just means a cloud built only for GPU workloads, without the databases, email and thousands of general services that AWS, Azure and Google Cloud sell.

Here's why that matters. A hyperscaler builds AI capacity for its own models and its own customers first. A neocloud sells almost all of its capacity to others, including to those same hyperscalers when they run short.

That's exactly what happened. Microsoft, a direct competitor in cloud, became CoreWeave's biggest customer by a wide margin. The 10-K names AWS, Google Cloud, Microsoft Azure and Oracle as key competitors, and notes several are also current customers.

CoreWeave by the numbers: Q3 2026 results

The Q3 2026 quarter ended September 30 and was reported on November 10, 2026. It beat the LSEG consensus of $3.55 billion in revenue, and management raised full-year guidance again.

MetricQ3 2026Q3 2025
Revenue$3,661 million$1,365 million
Operating income$207 million$52 million
Interest expense, net$846 million$310 million
Net loss$709 million$110 million
Adjusted EBITDA$2,256 million (62% margin)$838 million (61% margin)
Adjusted operating income$379 million (10% margin)$217 million (16% margin)
Revenue backlogAbout $132.5 billionn/a

Source: CoreWeave Q3 2026 earnings release, November 10, 2026.

Look at the gap between the lines. Adjusted EBITDA margin was 62%, but GAAP net loss margin was negative 19%. Depreciation and interest eat almost everything a GPU earns in its first years.

For context, full-year 2025 revenue was $5.131 billion with a $1.167 billion net loss. CoreWeave now guides 2026 to $13.0 billion to $13.3 billion and an exit run rate of $20 billion to $21 billion.

The backlog roughly doubled in nine months, from $66.8 billion at the end of 2025 to about $132.5 billion. The 10-Q says 36% of remaining performance obligations should convert to revenue within 24 months and the rest over months 25 to 86.

The GPU fleet: power, data centers and NVIDIA hardware

CoreWeave no longer leads with a GPU count. It reports capacity in megawatts, which is the real constraint on how many GPUs it can switch on.

DateData centersActive powerContracted power
Dec 31, 202310About 70 MWn/a
Dec 31, 202432About 360 MWn/a
Dec 31, 202543Over 850 MWAbout 3.1 GW
Sept 30, 2026Not disclosedOver 1.7 GWOver 4.6 GW
Dec 31, 2026 (target)n/aOver 1.85 GWn/a

Source: CoreWeave 2025 Form 10-K for 2023 to 2025; Q3 2026 results for 2026.

Every GPU in the fleet is NVIDIA. The 10-Q says customer contracts require it, and it lists that as a supplier risk. CoreWeave typically ships NVIDIA platforms first: it was early on GB200 NVL72 and GB300 NVL72, and in Q3 2026 it brought online multi-rack Vera Rubin NVL72 deployments.

The first production Vera Rubin workload, for Cognition, delivered 4.8x the token throughput of GB200 NVL72, according to CoreWeave. That's a vendor number on one workload, not an independent benchmark. For the rack itself, see Inside Deep Tech's NVIDIA Vera Rubin NVL72 guide.

International expansion is real but early. CoreWeave says it has more than 1.3 GW of contracted power outside the U.S., including India, Indonesia and Spain.

The software layer: Mission Control, Weights & Biases and Forge

CoreWeave's argument against "GPU landlord" is software. Its stack includes Kubernetes-native orchestration, CoreWeave Mission Control for cluster health and node failure prediction, AI Object Storage, and a growing set of developer tools.

The biggest piece came in May 2025, when it bought Weights & Biases for aggregate consideration of about $1.0 billion. It has since added Marimo and Monolith, launched Serverless RL, and in Q3 2026 launched CoreWeave Forge, an environment for training, inference, evaluation and agent workflows.

Third parties do rate the stack well. CoreWeave says it's the only provider to earn SemiAnalysis' Platinum ClusterMAX rating three consecutive times.

CoreWeave's contracts: Microsoft, OpenAI, Meta, Jane Street and NVIDIA

CoreWeave's backlog is built on a handful of very large, multi-year commitments. These are the ones disclosed publicly.

CustomerCommitmentTermSigned
OpenAIUp to about $22.4 billion totalThrough May 2031 (latest order)Mar 2025, May 2025, Sept 2025
MetaAbout $14.2 billionThrough Dec 2031Sept 2025
MetaAbout $21 billion (includes Vera Rubin)Through Dec 2032Apr 2026
Jane StreetAbout $6 billionNot disclosedApr 2026
NVIDIA$6.3 billion capacity backstopThrough Apr 2032Sept 2025
MicrosoftNot disclosed as a single figureNot disclosedNot disclosed

Source: CoreWeave OpenAI release for the $22.4 billion total; CNBC and CoreWeave Meta release for Meta; CoreWeave Jane Street release; Data Center Dynamics and the September 2025 8-K for NVIDIA.

The NVIDIA deal is the unusual one. Under the $6.3 billion initial order, NVIDIA agreed to buy any residual capacity CoreWeave can't sell to customers through April 2032. It works as a floor under utilization, from the company that also sells CoreWeave its GPUs and owns part of its stock.

NVIDIA also put in equity. In January 2026 it bought about 23 million shares in a $2.0 billion private placement at $87.20 per share. Jane Street bought $1.0 billion of stock at $109 per share in April 2026, alongside its cloud deal.

Smaller logos matter too. Q3 2026 wins included Canva, Wayve, Substack and Serve Robotics, with expansions at Cohere, Perplexity, Jane Street and OpenAI, per the Q3 release.

If you want to see where that OpenAI and Microsoft compute ends up, Inside Deep Tech's OpenAI Stargate guide and Microsoft Fairwater guide cover the first-party campuses CoreWeave capacity sits alongside.

How CoreWeave makes money: take-or-pay and price per megawatt

Most revenue comes from committed contracts. Customers reserve capacity for a fixed term, often prepay part of it, and pay whether they use it or not, per the 10-K. On-demand GPU rental exists, but it's a small slice.

Pricing moved sharply in 2026. On the Q3 call, CoreWeave said it raised prices across its offerings by about 25% in July and another 10% later in the quarter.

CEO Michael Intrator also gave rare per-megawatt figures. Short-term capacity is selling at more than $40 million per megawatt annually, while new four- and five-year contracts price at $20 million to $25 million per megawatt.

📊
Why the per-megawatt number matters: it's the cleanest way to compare neocloud economics. A 100 MW hall on a five-year contract at $20 million to $25 million per MW is $2 billion to $2.5 billion a year of revenue, before power, depreciation and interest, based on the pricing CoreWeave disclosed.

The trade-off is simple. Short contracts earn more per megawatt, but lenders prefer long contracts because the cash flow is predictable. CoreWeave is now trying to do both.

The debt machine: how CoreWeave finances its GPU fleet

This is the part of CoreWeave most readers actually want explained. The company buys GPUs and builds out leased data centers years before customers finish paying. Debt fills the gap.

As of September 30, 2026, total debt principal was $47.5 billion. The repayment schedule in the same filing shows about $2.4 billion due in the rest of 2026, $11.0 billion in 2027 and $19.6 billion after 2030.

Financing toolWhat it isRecent example
Delayed draw term loans (DDTL)Secured loans backed by GPUs and specific customer contracts$2.6 billion DDTL 5.5 for shorter contracts
Convertible notesParent-level notes that can convert to stockAbout $4.2 billion raised in Q3 2026
Contract-backed loanSecured loan tied to a highly rated customerAt least $4 billion, Meta-backed, in market Nov 2026
EquityPrivate placements and an at-the-market programATM for up to 35 million shares
Customer prepaymentsCash paid up front on new contractsAbout $9 billion expected over three months

Source: CoreWeave Q3 2026 results for DDTL 5.5, the converts, ATM and prepayments; Semafor and Quartz for the Meta-backed loan.

The key split is secured versus unsecured. CoreWeave's unsecured bonds sold off hard this fall, with some yields topping 16%. Its GPU-backed loans kept trading around par, according to CFO Nitin Agrawal.

So CoreWeave changed course. Agrawal told analysts it won't issue new parent-level unsecured debt through 2027 at current yields, and expects under $5 billion of parent-level financing in that period. The new Meta-backed loan is targeting an all-in rate around 7.5%, roughly half its junk bond yields.

Interest is still climbing. CoreWeave guided Q4 2026 interest expense to $1.05 billion to $1.15 billion, up from $846 million in Q3.

What happened with Core Scientific

In July 2025 CoreWeave agreed to buy Core Scientific, a former bitcoin miner and one of its biggest data center landlords, in an all-stock deal valued at about $9.0 billion. The point was to own roughly 1.3 GW of gross power instead of leasing it.

Core Scientific shareholders voted it down. The company terminated the merger agreement on October 30, 2025 and stayed independent on Nasdaq as CORZ.

That leaves CoreWeave on the leasing model. It still rents most of its powered shells from developers like Core Scientific, which keeps capex lower per site but leaves lease obligations and landlord risk in place.

How CoreWeave compares with rival AI clouds

CoreWeave isn't alone, and its edge is narrowing as everyone buys the same NVIDIA racks. Here's how the main groups compare, based on public positioning rather than audited capacity.

Provider typeExamplesWhere they beat CoreWeaveWhere CoreWeave leads
HyperscalersAWS, Microsoft Azure, Google CloudBalance sheets, full service catalogs, custom chipsSpeed to new NVIDIA platforms, GPU-only focus
OracleOCILarger parent balance sheet, OpenAI Stargate roleTrack record running dense GPU clusters
Other neocloudsNebius, Lambda, Crusoe, NscaleSome own more of their power and landScale, backlog, ClusterMAX rating
First-party buildsxAI Colossus, Meta HyperionNo margin paid to a middlemanFlexible capacity without owning the site

Two first-party projects show the alternative to renting. xAI built its own Colossus 2 supercomputer in Memphis, and Meta is building Hyperion in Louisiana while still signing two contracts worth about $35 billion combined with CoreWeave.

The whole group trades together. MarketWatch reported that CoreWeave, Iren and Nebius had each shed more than half their value over three months as of November 18, 2026.

Inside Deep Tech's take

🧭
Inside Deep Tech's take: CoreWeave is a real business with real customers, not a bubble artifact. A 62% adjusted EBITDA margin on $3.66 billion of quarterly revenue, per the Q3 release, proves the demand side. The open question is the right side of the balance sheet. CoreWeave only works if GPUs keep earning long enough to repay the debt that bought them, and if customers renew at today's prices. Bet on the operator, but watch the 2027 maturities and the Microsoft contract renewals more closely than the backlog headline.

Honest limits: where the CoreWeave story can break

Customer concentration

Microsoft alone was 67% of 2025 revenue, and the top two customers were 77% in 2024. Diversification is happening: in Q3 2026 the top two were 30% and 24%. But the backlog leans on OpenAI and Meta, and the 10-Q says concentration will likely continue because of long contract terms.

A contract is only as good as the customer's ability to pay. CoreWeave's own filing flags non-payment and non-performance risk from its top customers, and says it plans to add more early-stage and private customers over time.

Debt load and GPU-backed financing

Interest expense ran $846 million in Q3 2026 against $207 million of operating income. With $11.0 billion of principal due in 2027, per the 10-Q, CoreWeave needs open credit markets and steady customer prepayments. A bad quarter for AI sentiment raises its cost of capital fast, as this fall showed.

Depreciation and GPU useful life

CoreWeave depreciates data center computing equipment over six years, up from five before 2023. Depreciation on property and equipment was $1.9 billion in Q3 2026 alone.

NVIDIA now ships a new rack generation roughly every year. If Hopper and early Blackwell systems rent for much less in year four or five than the six-year schedule assumes, reported margins are too high today.

Power and construction risk

Contracted power of 4.6 GW against 1.7 GW active means most of the capacity still has to be built, energized and cooled. Revenue in the backlog is conditional on delivery: the 10-Q defines backlog as subject to delivery and availability of service requirements. Late data centers mean late revenue and possible credits.

NVIDIA dependence

NVIDIA is CoreWeave's only GPU supplier, a shareholder and a capacity backstop. When NVIDIA's board announced Jensen Huang's immediate resignation on November 12, 2026, CoreWeave stock fell with it. On November 18, interim CEO Colette Kress said NVIDIA would be disciplined and selective about financing support for neoclouds.

What CoreWeave means for you

  • If you buy AI compute: CoreWeave is often first to new NVIDIA platforms. Negotiate credits for late delivery and check which data center your cluster lands in.
  • If you sell into data centers: CoreWeave's $37 billion to $39 billion 2026 capex guide makes it one of the largest buyers of power, cooling and networking gear outside the hyperscalers.
  • If you invest: read the debt maturity table and customer concentration note in each 10-Q before the revenue line.
  • If you build AI infrastructure strategy: neocloud capacity is a hedge, not a free option. Price the counterparty risk.

FAQ

What does CoreWeave do?

CoreWeave rents NVIDIA GPU clusters and AI cloud software to AI labs, hyperscalers and enterprises. Its Q3 2026 release says it serves 9 of the 10 leading foundation model providers.

Who are CoreWeave's biggest customers?

Microsoft, OpenAI and Meta are the largest. Microsoft was 67% of 2025 revenue, and the 10-Q names Microsoft and OpenAI as significant customers alongside a $21 billion Meta order.

How much debt does CoreWeave have?

Total debt principal was about $47.5 billion as of September 30, 2026. Roughly $11.0 billion of that is scheduled to be repaid in 2027.

Is CoreWeave profitable?

Not on a GAAP basis yet. It reported a $709 million net loss in Q3 2026 despite $207 million of operating income, because of $846 million in net interest expense.

Did CoreWeave buy Core Scientific?

No. Core Scientific shareholders rejected the deal and the company terminated the merger on October 30, 2025.

What GPUs does CoreWeave use?

Only NVIDIA GPUs, according to its 10-Q. It runs Hopper and Blackwell systems including GB200 and GB300 NVL72, and brought Vera Rubin NVL72 racks online in Q3 2026.

What to watch next

  • Whether CoreWeave closes the Meta-backed loan of at least $4 billion near its 7.5% target.
  • Q4 2026 results: did active power clear 1.85 GW and did the run rate exit at $20 billion to $21 billion, per guidance?
  • Microsoft contract renewals and any new hyperscaler deals.
  • NVIDIA's financing posture under its new leadership.
  • Rental pricing for Hopper and early Blackwell systems as Vera Rubin ramps.

If you're tracking where AI compute actually gets built, start with the campuses CoreWeave competes and partners with. Inside Deep Tech's OpenAI Stargate guide is the best next read.

Microsoft Fairwater: A Full Guide to the AI Superfactory in Wisconsin and Atlanta
How Microsoft's first-party AI superfactory works, and how it compares with Stargate, Hyperion and Colossus 2.

Cover image: generated GPU data center hall, blurred with title overlay for this post.