Revenue was in Q2 2026, against a year earlier, all of it AI cloud capacity and the services around it. Concentration eased: Customer A took , against a year earlier, Customer B and Customer C , and the 10-Q names Jane Street, Microsoft and OpenAI as significant customers. Customer A is Microsoft by arithmetic on the 2025 10-K, which names Microsoft at of 2025 revenue and no other customer at ; Customer B is OpenAI or Meta, and no stored source says which. The ledger’s payer split moves toward labs and enterprises: from hyperscalers and Meta, from AI labs and AI-native companies, from enterprises, every one of them its own judgment.
What changed in disclosure: the CFO’s Q1 figure for the backlog from non-investment-grade labs and AI natives was not repeated, and run rates appeared instead. Managed inference, priced by the token, reached more than of booked annual recurring revenue from , and storage, CPU, networking and software more than ; both sit inside the payer channels and are left out of the flow total. CoreWeave Omni signed its first deal, to scale in 2027. Prices rose about across SKUs in July, after the quarter.
The spend side grew with active power, which reached megawatts: cash paid for property and equipment was , management’s capital expenditure figure , depreciation , lease cost , interest on debt , and the ledger’s power estimate . Funding stayed mixed and partly vendor-financed on both sides of the company: equipment makers financed outstanding and a customer, Jane Street, invested , beside about of new debt, convertibles and equity. Debt principal reached . Other income included net unrealized gains of on strategic investments in privately held companies; the filing does not name the investees or describe them as AI, so the gain is recorded on no channel.
On who pays, the sources say more about willingness than about funding: the CEO says many customers now monetize their products and pay higher margins for compute, and the 10-Q still describes some customers as private, early stage or highly leveraged. The investor-funded half the roster expected to see is visible only as a description, not as a filed number, in this quarter.
Sized channels against the income statement, Q2 2026
10 of 11 channels sized
Each blue mark is one channel's dollars for the quarter; a bar is the range of an estimate. Grey marks are the company's reported lines. The distance between them is the point: how large the AI channel is next to the line it sits in.
Total operating expensesreported line
$2.62bn
Revenuereported line
$2.58bn
Cost of revenuereported line
$879mn
General and administrativereported line
$178mn
Sales and marketingreported line
$60mn
Capital expenditures on GPU servers, networking and data center fit-outspend · our inference
AI cloud capacity rented by hyperscalers and large technology platformsrevenue in · our inference
Depreciation of the GPU fleet and data center equipmentspend · our inference
AI cloud capacity rented by established enterprisesrevenue in · our inference
Interest on the debt that funds the fleetspend · our inference
Lease cost of the data centers that house the fleetspend · our inference
AI cloud capacity rented by AI labs and AI-native companiesrevenue in · our inference
Power and utilities for the data centersspend · our inference
Storage, CPU, networking and AI developer software sold beside GPU capacityrevenue in · our inference
Managed inference: serverless and dedicated model serving, priced by the tokenrevenue in · our inference
$1mn$10mn$100mn$1bn$10bn
AI channel, dollars for the quarter low to high of an estimate reported lineLog scale: each gridline is ten times the one before.
New money and old money, Q2 2026
1 new10 expanded
Each channel is tagged once for whether its money existed before language models, from the company's annual report and call at the start of the period. A bar splits one flow's sized dollars by that tag. The incremental total is the part that would not be there without the models: a new channel counts in full, an expanded one only for what AI added, a relabelled one at zero.
Flow, sized total
Split by novelty
Incremental total
Paid for AI$2.31bn to $2.80bn sized
expanded $2.31bn to $2.80bn
Incremental total $0 to $2.80bnpoint $0$2.59bn in 4 channels has no traced baseline
Revenue arriving through AI$1.57bn to $3.79bn sized
new $258mn to $1.08bnexpanded $1.31bn to $2.70bn
Incremental total $258mn to $3.79bnpoint $438mn$2.14bn in 2 channels has no traced baseline
The sized total counts every channel the company credits to AI, including relabelled money that existed before language models and the ledger's own estimates for it. The incremental total counts relabelled channels at zero. A flow is split by layer where its dollars sit at more than one: end use, compute sold to builders, and hardware. The same dollar can be a buyer's spend, a cloud's revenue and a chipmaker's revenue, so the layers are never added together.
Paid for AI5 channels · $2.31bn to $2.80bn sized · $0 to $2.80bn incremental
other
Capital expenditures on GPU servers, networking and data center fit-out
212% to 249.4% of the quarter’s revenue; capital spending on the build, not added into totals (its depreciation is)
our inferencedisclosure: quantified· motive: offensive· before LLMs: expanded· layer: compute
Cash paid for property and equipment was in the quarter, the first half less Q1, against a year earlier and times revenue; management’s capital expenditure figure was , and guidance for 2026 rose to to (c53, c55). The 10-Q again says the investments enable the development and deployment of AI models and are funded through a mix of debt and equity issuances, delayed draw term loans, OEM financing and cash from the balance sheet (c10). The size, , is the ledger’s own; it is capitalized and left out of totals. NVIDIA, whose GPUs are the whole fleet, is on this ledger as NVDA, where this spending reaches its revenue, partly through the equipment makers; this ledger does not net the two. The CFO says the build is financed by debt, customer prepayments and corporate capital (c48). Funding is mixed, with vendor financing on both sides: equipment makers financed outstanding, NVIDIA’s share purchase of January is repeated, and Jane Street, a customer, invested (c13, c12, c26). About of debt, convertibles and equity was raised in the quarter. Self-built sites are under way (c44).
Evidence: 16 quotes, 14 figures, 2 confounds, 5 from before coverage
Cash paid for property and equipment: GPU servers (all of them NVIDIA GPUs, which customer contracts specify), networking, storage and the data center equipment the company installs in leased shells. It is capitalized and reaches the income statement as depreciation, which counts in the spend total in its place, so this channel is traced and shown and left out of totals. The payees are NVIDIA, OEMs and ODMs. Every GPU in the fleet is an NVIDIA GPU, and NVIDIA is on this ledger as NVDA, where this spending reaches its revenue, partly through the equipment makers; this ledger does not net the two. It is funded by delayed draw term loans secured on customer contracts, senior and convertible notes, OEM vendor financing, customer prepayments and equity, including stock bought by NVIDIA, the main supplier, and by Jane Street, a customer: vendor financing on both sides. Management’s own capital expenditure figure differs from cash paid, because equipment bought on vendor financing or still payable is counted when received.
Why this motive
Capital committed against contracted AI capacity: higher capital expenditures reflect accelerating customer deliveries (c53, c10).
Before LLMs: expanded
Capital spending on a GPU fleet predates LLMs at the company, which ran GPUs for crypto mining before launching its cloud in 2020. At the anchor cash purchases of property and equipment were in 2024 and management capital expenditures in Q1 2025, all of it already the AI build; no quarter before AI can be traced. The size is the whole of an activity that existed before.
“We were founded in September 2017 and launched our CoreWeave Cloud Platform in 2020. Moreover, prior to 2022, we had limited revenue, most of which was derived from our crypto mining offerings, which we have discontinued.”
“In April 2023, we entered into a Master Services Agreement (the “Master Services Agreement”) with NVIDIA, a beneficial owner of more than 5% of our outstanding capital stock, pursuant to which we provide NVIDIA with our infrastructure and platform services through fulfillment of order forms submitted to us by NVIDIA.”
“Turning to capital expenditures, CapEx consists primarily of investments in property and technology and equipment that drive our platform expansion, amplify the value of our software and tech stack, and ultimately fuel our revenue growth.”
“For example, as a result of our obligations in our current customer contracts, all of the GPUs used in our infrastructure today are NVIDIA GPUs. Additionally, for the year ended December 31, 2025, three suppliers accounted for 23%, 20%, and 17% of total purchases”
Equity investment by Jane Street, a customer · 2026-CQ2
Debt, convertible notes and equity raised in Q2 (approximate) · 2026-CQ2
Total principal of debt (future principal payments) · as-of 2026-06-30
Deferred revenue (customer prepayments not yet earned) · as-of 2026-06-30
Cash paid for property and equipment as a multiple of revenue · 2026-CQ2
Active power, megawatts · as-of 2026-06-30
Active power added in the quarter, megawatts (stated as nearly this amount) · 2026-CQ2
What else could explain it
line composition: Cash paid excludes equipment bought on OEM financing until it is repaid; management’s capital expenditure figure is measured differently and is higher this quarter.
other: Component prices rose; management says increases are passed through.
Quotes
“Our capital investments in property and equipment consist primarily of technology and infrastructure, which consist of our investments in servers and network equipment for computing, storage, and networking requirements that collectively enable the development and deployment of AI models. We fund these capital investments through a mix of debt and equity securities issuances, delayed draw term loan facilities, OEM financing arrangements, and cash from our balance sheet.”
“In January 2026, we entered into a securities purchase agreement with NVIDIA Corporation for a private placement of approximately 23 million shares of our Class A common stock at a purchase price of $87.20 per share, for aggregate gross proceeds of $2.0 billion.”
“The Company had entered into various agreements with original equipment manufacturers (the "OEM Financing Arrangements"), whereby the Company obtained financing for certain equipment. The Company had an outstanding balance of $4.8 billion and $3.8 billion as of June 30, 2026 and December 31, 2025, respectively.”
“Moreover, our committed contracts typically include a prepayment from our customers prior to them receiving access to our services. As of December 31, 2025, 2024, and 2023, the weighted-average prepayment across all our active contracts was 15% to 25% of the TCV.”
“We continued to execute on our power strategy, reaching 1.5 GW of active power, adding nearly 500 MW, more than any quarter in our history, and more than tripling year-over-year.”
“Building on our close partnerships with NVIDIA and our OEM and ODM partners, our recent long-term agreement with Solidigm is one illustration of how we are de-risking access to the critical inputs needed to serve our customers.”
“This operating margin improvement came before our July pricing changes, which included an approximately 25% increase across SKUs in response to the current demand environment and the increasing ROI our customers are observing from their investments in the CoreWeave platform as they shift to inference. We are also passing through component price increases.”
“The cost, primarily in the form of CapEx, is front-loaded, requiring a combination of debt, customer prepayments, and other corporate-level capital to finance its build-out. Once the cluster is delivered, contracted revenue ramps, becoming predictable and highly cash flow generative.”
“CapEx in Q2 totaled $9.4 billion, slightly above the high end of our guided range. Higher CapEx in the quarter reflects customer deliveries accelerating.”
“In Q2, we made significant progress in strengthening our balance sheet and expanding the depth and breadth of our access to capital, raising approximately $18 billion across a combination of debt, convertibles, and equity.”
“As a result of our increased expectations around capacity to be delivered to customers this year, as well as some of our significant recent wins, we now expect 2026 CapEx in the range of $35 billion- $39 billion.”
Depreciation of the GPU fleet and data center equipment
48.7% to 54.1% of the quarter’s revenue
Incremental total: counts at zero at the point and in full at the high end, with no traced baseline.
our inferencedisclosure: quantified· motive: offensive· before LLMs: expanded· layer: compute
Depreciation and amortization was , the first half less Q1, against a year earlier; inside technology and infrastructure, , it rose on equipment placed in service (c15). Management defends the useful life it assigns to GPUs: an A100 contract signed into 2029 for a generation introduced in 2020, older generations priced at or above a year ago, and little of the fleet up for renewal (c50, c59, c16). The size, , is the ledger’s own. The company sells only AI cloud capacity and its services, so the AI share is the fleet share: the 10-Q describes revenue as compute optimized for AI and high-performance computing and the investments as equipment that enables the development and deployment of AI models (c1, c10).
Evidence: 9 quotes, 6 figures, 3 confounds, 2 from before coverage
Depreciation and amortization of servers, switches, networking equipment and power installations placed in service, inside technology and infrastructure and cost of revenue. It is the income-statement cost of the capital spending channel and counts in the spend total. Its size depends on the useful life the company assigns to GPUs, an estimate the 10-Q flags and management defends by recontracting older GPUs.
Why this motive
Carried from the prior quarter: the cost of delivering separately priced AI capacity, with older GPU generations recontracted at attractive prices (c50, c59).
Before LLMs: expanded
Depreciation was in 2024, from investments in servers, switches and networking equipment; the fleet before LLMs was a crypto-mining and early cloud fleet with no traceable quarter. The size is the whole of an activity that existed before.
“We were founded in September 2017 and launched our CoreWeave Cloud Platform in 2020. Moreover, prior to 2022, we had limited revenue, most of which was derived from our crypto mining offerings, which we have discontinued.”
“This increase was attributable to an increase in depreciation and amortization of approximately $742 million, from $101 million for the year ended December 31, 2023, to $843 million for the year ended December 31, 2024, resulting from investments in our platform and servers, switches, and other networking equipment fixed assets within our infrastructure”
Depreciation and amortization (first half less Q1) · 2026-CQ2
Depreciation and amortization, first half of 2026 (cash flow statement) · 2026-01-01..2026-06-30
Depreciation and amortization, prior-year quarter (first half less Q1) · 2025-CQ2
Increase in depreciation and amortization inside technology and infrastructure · 2026-CQ2
Technology and infrastructure · 2026-CQ2
Technology and infrastructure, prior-year quarter · 2025-CQ2
What else could explain it
line composition: Depreciation and amortization also covers offices, internal software and acquired intangibles.
line composition: Part of the fleet may serve high-performance computing that is not AI; the filing does not separate it.
other: A shorter useful life for GPUs would raise this cost; the estimate is the company’s.
Quotes
“We generate revenue by providing our customers with access to cloud computing services, including compute enabled by our software and infrastructure optimized for AI and high-performance computing.”
“Our capital investments in property and equipment consist primarily of technology and infrastructure, which consist of our investments in servers and network equipment for computing, storage, and networking requirements that collectively enable the development and deployment of AI models. We fund these capital investments through a mix of debt and equity securities issuances, delayed draw term loan facilities, OEM financing arrangements, and cash from our balance sheet.”
“This increase was primarily attributable to an increase in depreciation and amortization of approximately $752 million, from $537 million for the three months ended June 30, 2025, to approximately $1.3 billion for the three months ended June 30, 2026. These increases in depreciation and amortization were related to investments in our platform and servers, switches, and other networking equipment fixed assets within our infrastructure that were placed in service.”
“This requires us to make certain estimates with respect to the useful life of the components of our infrastructure and to maximize the value of the components of our infrastructure, including our GPUs, to the fullest extent possible.”
“Pricing and margins for our Blackwell and Vera Rubin SKUs are setting new highs, while pricing for prior generation SKUs is at or above where it was years ago. Our near-term capacity remains effectively sold out.”
“In terms of the capacity that is coming up for renewal, Samik, that is a very limited part of our fleet, and the ASPs on the older generation remain higher or at levels that we have seen about a year ago.”
“We really think about the fact that that is an incredible opportunity for us to offer the most bleeding-edge compute that we have, but also a wonderful way for us to access and use GPUs that are coming off contract in a way to extract maximum value for the company over time.”
Lease cost of the data centers that house the fleet
17.5% to 19.4% of the quarter’s revenue
Incremental total: counts at zero at the point and in full at the high end, with no traced baseline.
our inferencedisclosure: quantified· motive: offensive· before LLMs: expanded· layer: compute
Operating lease cost was , against a year earlier; rent in cost of revenue, in all, rose about (c17). Active power reached megawatts with nearly added in the quarter (c43). Leases signed and not yet commenced carry of future payments, a future cost and not this quarter’s, lower than at the end of Q1 as leases commenced (c18). The size, , is the ledger’s own. The company sells only AI cloud capacity and its services, so the AI share is the share that houses the fleet: the 10-Q describes revenue as compute optimized for AI and high-performance computing and the fleet as equipment that enables the development and deployment of AI models (c1, c10). Paid from operating cash flow, which the 2025 10-K attributes mainly to upfront payments under customers’ committed contracts (crwv-anchor-c19).
Evidence: 5 quotes, 5 figures, 3 confounds, 2 from before coverage
Operating lease cost of powered data center shells leased from third-party developers, recorded as rent in cost of revenue, with offices a small part. Leases signed and not yet commenced are recorded as a metric: they are future cost, not the quarter’s. The payees are data center developers and landlords, some of them financed through joint ventures and special-purpose entities the company invests in.
Why this motive
Carried from the prior quarter: the cost of delivering separately priced AI capacity, shells leased for contracted demand (c17).
Before LLMs: expanded
Operating lease cost was in 2024, with rent rising as new data centers were deployed. The size is the whole of an activity that existed before.
“This increase was attributable to higher costs directly related to running our data centers to support the significant increase in customer demand, driven by the successful deployment of new data centers, which resulted in an increase in rent expense of approximately $290 million, an increase in data center utilities and power spend of approximately $76 million”
Increase in rent expense over the prior-year quarter · 2026-CQ2
Future lease payments on leases signed and not yet commenced (future cost) · as-of 2026-06-30
Cost of revenue · 2026-CQ2
What else could explain it
line composition: Operating lease cost includes offices and storage space.
line composition: Part of the capacity housed may serve high-performance computing that is not AI; the filing does not separate it.
other: Rent starts on delivery of a powered shell, before revenue; power delivered late in the quarter front-loads the cost.
Quotes
“We generate revenue by providing our customers with access to cloud computing services, including compute enabled by our software and infrastructure optimized for AI and high-performance computing.”
“Our capital investments in property and equipment consist primarily of technology and infrastructure, which consist of our investments in servers and network equipment for computing, storage, and networking requirements that collectively enable the development and deployment of AI models. We fund these capital investments through a mix of debt and equity securities issuances, delayed draw term loan facilities, OEM financing arrangements, and cash from our balance sheet.”
“This increase was primarily attributable to the expansion of existing data centers and the significant increase in the deployment of new data centers, which resulted in an increase in rent expense of approximately $335 million, and an increase in data center utilities and power spend of approximately $87 million.”
“As of June 30, 2026, the Company executed additional lease agreements, primarily for data centers, equipment, and office buildings, that had not yet commenced. The aggregate amount of estimated future undiscounted lease payments associated with such leases is $35.5 billion.”
“We continued to execute on our power strategy, reaching 1.5 GW of active power, adding nearly 500 MW, more than any quarter in our history, and more than tripling year-over-year.”
Incremental total: counts at zero at the point and in full at the high end, with no traced baseline.
our inferencedisclosure: direction only· motive: offensive· before LLMs: expanded· layer: compute
The filing again gives only the increase in data center utilities and power spend, about on the year, about the same as in Q1 (c17). Active power reached megawatts, with nearly added in the quarter (c43). The flat increase against a fast-growing fleet suggests most new power is billed inside rent, so the growth ratio for this line is not tied to active power: the point and the high-cost end keep the prior quarter’s ratios, and only the low-cost end follows active power. The size, , is the ledger’s own: the level implied by the increase and that growth ratio, plus an assumed share of variable lease cost, , for utilities reimbursed to landlords. The company sells only AI cloud capacity and its services, so the AI share is the fleet’s share: the 10-Q describes revenue as compute optimized for AI and high-performance computing and the fleet as equipment that enables the development and deployment of AI models (c1, c10). Paid from operating cash flow, which the 2025 10-K attributes mainly to upfront payments under customers’ committed contracts (crwv-anchor-c19). The CEO says the company pays for grid upgrades at its sites, a capital cost outside this channel (c63).
Evidence: 5 quotes, 4 figures, 4 confounds, 2 from before coverage
Electricity and other utilities the company pays for its data centers, in cost of revenue, and utilities reimbursed to landlords inside variable lease cost. The filing gives only the year-over-year increase in utilities and power spend, so the level is the ledger’s estimate. Management says power costs are contracted for the term and passed through in pricing. The payees are utilities and landlords.
Why this motive
Carried from the prior quarter: the cost of delivering separately priced AI capacity, contracted and passed through in pricing.
Before LLMs: expanded
Data center utilities and power spend rose in 2024 as new data centers were deployed; the filing gives no level. The size is the whole of an activity that existed before.
“This increase was attributable to higher costs directly related to running our data centers to support the significant increase in customer demand, driven by the successful deployment of new data centers, which resulted in an increase in rent expense of approximately $290 million, an increase in data center utilities and power spend of approximately $76 million”
Increase in data center utilities and power spend over the prior-year quarter · 2026-CQ2
Variable lease cost (common area maintenance, utilities reimbursed to landlords, physical security) · 2026-CQ2
Active power, megawatts · as-of 2026-06-30
Active power added in the quarter, megawatts (stated as nearly this amount) · 2026-CQ2
What else could explain it
line composition: Power bought inside all-in rent is in the lease channel.
line composition: Adding a share of variable lease cost may double count if the filing’s utilities and power spend already includes utilities reimbursed to landlords.
line composition: Part of the fleet may serve high-performance computing that is not AI; the filing does not separate it.
other: The level is solved from a single increase and an assumed growth ratio.
Quotes
“We generate revenue by providing our customers with access to cloud computing services, including compute enabled by our software and infrastructure optimized for AI and high-performance computing.”
“Our capital investments in property and equipment consist primarily of technology and infrastructure, which consist of our investments in servers and network equipment for computing, storage, and networking requirements that collectively enable the development and deployment of AI models. We fund these capital investments through a mix of debt and equity securities issuances, delayed draw term loan facilities, OEM financing arrangements, and cash from our balance sheet.”
“This increase was primarily attributable to the expansion of existing data centers and the significant increase in the deployment of new data centers, which resulted in an increase in rent expense of approximately $335 million, and an increase in data center utilities and power spend of approximately $87 million.”
“We continued to execute on our power strategy, reaching 1.5 GW of active power, adding nearly 500 MW, more than any quarter in our history, and more than tripling year-over-year.”
Q1 2026direction only · our inference · offensive · $147mn to $334mn
Q2 2026direction only · our inference · offensive · $131mn to $351mn
other
Interest on the debt that funds the fleet
18.4% to 21.7% of the quarter’s revenue
Incremental total: counts at zero at the point and in full at the high end, with no traced baseline.
our inferencedisclosure: quantified· motive: offensive· before LLMs: expanded· layer: compute
Interest expense, net was , against a year earlier, attributed to higher borrowing; interest on debt obligations was after was capitalized (c19, c20, c52). Debt principal reached , with a publicly syndicated delayed draw loan and further unsecured debt and convertibles raised in the quarter (c27, c28). The delayed draw loans are secured on contracts mostly with investment-grade customers and amortize from their cash flows (c21). The size, , is the ledger’s own. The company sells only AI cloud capacity and its services, so the AI share is the fleet’s share: the 10-Q describes revenue as compute optimized for AI and high-performance computing and the fleet as equipment that enables the development and deployment of AI models (c1, c10). Paid from operating cash flow, which the 2025 10-K attributes mainly to upfront payments under customers’ committed contracts (crwv-anchor-c19).
Evidence: 14 quotes, 7 figures, 3 confounds, 3 from before coverage
Interest expense on delayed draw term loans secured on customer contracts and the equipment they finance, senior and convertible notes, the revolving facility and OEM vendor financing, net of interest capitalized into equipment under construction. The interest expense line also carries imputed interest on customer prepayments with a significant financing component and on finance leases, which are not owed to lenders and are left out of this channel. The payees are banks, bondholders and equipment vendors acting as lenders.
Why this motive
Carried from the prior quarter: the cost of capital raised for contracted capacity, repaid from contract cash flows (c49, c21).
Before LLMs: expanded
At the anchor the company had raised of debt commitments through 2024, interest expense, net was in Q4 2024 and interest expense in Q1 2025. The size is the whole of an activity that existed before.
“This increase was attributable to higher borrowing levels under our Credit Facilities, resulting in an increase in interest expense, partially offset by capitalized interest associated with investments in our platform and data center infrastructure.”
“Interest expense for Q1 was $264 million, higher than our expectations due to improvement in our vendor payment terms, which reduced the days between vendor payment and assets being put in service, hence reducing the amount of interest cost capitalized in the quarter.”
Debt, convertible notes and equity raised in Q2 (approximate) · 2026-CQ2
What else could explain it
line composition: Interest expense, net also carries imputed interest on customer prepayments and finance leases.
line composition: Part of the fleet the debt funds may serve high-performance computing that is not AI; the filing does not separate it.
other: Unsecured notes and convertibles raised in the quarter fund general corporate needs as well as the fleet.
Quotes
“We generate revenue by providing our customers with access to cloud computing services, including compute enabled by our software and infrastructure optimized for AI and high-performance computing.”
“Our capital investments in property and equipment consist primarily of technology and infrastructure, which consist of our investments in servers and network equipment for computing, storage, and networking requirements that collectively enable the development and deployment of AI models. We fund these capital investments through a mix of debt and equity securities issuances, delayed draw term loan facilities, OEM financing arrangements, and cash from our balance sheet.”
“The Company had entered into various agreements with original equipment manufacturers (the "OEM Financing Arrangements"), whereby the Company obtained financing for certain equipment. The Company had an outstanding balance of $4.8 billion and $3.8 billion as of June 30, 2026 and December 31, 2025, respectively.”
“These delayed draw term loan facilities are collateralized with the assets underlying the contributed contracts and the pledged contractual cash flows, generally from investment grade counterparties. They are drawn as we build infrastructure to support customer requirements, and amortize over time as contracted cash flows are generated in a regular and predictable manner, with excess cash made available to us.”
“The cost, primarily in the form of CapEx, is front-loaded, requiring a combination of debt, customer prepayments, and other corporate-level capital to finance its build-out. Once the cluster is delivered, contracted revenue ramps, becoming predictable and highly cash flow generative.”
“The deployment delivers attractive returns, fully repaying asset-level debt used to fund the CapEx while generating significant additional free cash flow.”
“Interest expense for Q2 was $640 million, compared to $267 million in Q2 of 2025. Driven by increased debt to support the continued scaling of our infrastructure and delivery of our contracted customer commitments.”
“In Q2, we made significant progress in strengthening our balance sheet and expanding the depth and breadth of our access to capital, raising approximately $18 billion across a combination of debt, convertibles, and equity.”
“These clients want to buy compute for two years or three years, and that is not a market that was easily accessible to us until we were able to bring DDTL 5.5 to market.”
“This financing is significant as it unlocks our ability to serve critical part of the enterprise market at scale, while also allowing us to accelerate the ramp of our managed inference platform and grow our exposure to shorter-dated contracts that typically come at a higher ASP and margins.”
Revenue arriving through AI6 channels · $1.57bn to $3.79bn sized · $258mn to $3.79bn incremental · 1 not sized
customer cohort
AI cloud capacity rented by hyperscalers and large technology platforms
36% to 67% of the quarter’s revenue
Incremental total: counts at zero at the point and in full at the high end, with no traced baseline.
our inferencedisclosure: described· motive: offensive· before LLMs: expanded· layer: compute
Revenue was , against a year earlier. Customer A took , Customer B and Customer C ; the 10-Q now names Jane Street, Microsoft and OpenAI as significant customers and still expects Meta to become one (c2, c3, c4). Customer A is Microsoft, by arithmetic on the 2025 10-K: it names Microsoft as the top customer at of 2025 revenue and says no other customer reached , a floor of (crwv-anchor-c18, crwv-anchor-c22), and the same row of this 10-Q gives Customer A of the prior-year quarter, , above that floor in a single quarter. Customer B is OpenAI or Meta: neither reached the floor in 2025, the 10-Q names OpenAI as significant and still expects Meta to become so, and no stored source says which, so the range runs from Customer A alone to Customers A and B together. The size, , is the ledger’s own, an assumed share of revenue. The CEO says hyperscalers, labs and enterprises all use the platform (c57). These payers fund the rent from operations; Microsoft’s and Meta’s payments are spend on the MSFT and META ledgers. Backlog, , and remaining performance obligations, , are recorded as metrics only. Counterparty mixed: cloud providers and Meta.
Evidence: 17 quotes, 12 figures, 3 confounds, 6 from before coverage
GPU clusters, with the networking, storage and software around them, rented on multi-year take-or-pay contracts by the largest cloud and technology companies: Microsoft, which the prospectus names as the largest customer, Meta, which signed new orders in Q1 2026, and other hyperscalers the calls refer to without naming. The company is AI by construction: the 10-Q describes its revenue as access to compute optimized for AI and high-performance computing, and its capital investments as equipment that enables the development and deployment of AI models. The filings give revenue shares by customer letter. Customer A is Microsoft: the 2025 10-K names Microsoft as the top customer and says no other customer reached the disclosure floor in 2025, and Customer A’s revenue in each prior-year quarter of the 10-Q tables is above that floor. Customer B is OpenAI or Meta, and no stored source says which, so this channel is sized as an assumed share of reported revenue of at least Customer A’s share. These payers fund their rent from their own operations; Microsoft and Meta are on this ledger as MSFT and META, where the same dollars are spend, and this ledger records each company’s own flows without netting them.
Why this motive
Buyers paying more for a separately priced AI product: prices for new GPU generations at new highs, a price increase of about across SKUs from July, and capacity sold out (c33, c47).
Before LLMs: expanded
At the anchor this money was already the largest part of revenue: Microsoft took of revenue in 2024 and in 2023, under its master services agreement in 2024, and the company added a further hyperscaler in Q1 2025. Renting GPU cloud capacity predates LLMs at the company itself, which launched its cloud in 2020 beside crypto mining, with revenue of in 2022; no quarter of the same activity before AI can be traced, so the level has no baseline. The size is the whole of an activity that existed before.
“largest customer accounted for 16% of our revenue. For the years ended December 31, 2023 and 2024, our largest customer was Microsoft, which accounted for 35% and 62% of our revenue, respectively.”
“Large companies whose business and product are not AI, but are being driven by AI. These include Fortune 500 companies and other large enterprises with business models that are increasingly AI-enabled through their own development efforts and/or that leverage AI directly to drive internal efficiency gains. We have seen initial traction in adopting our platform from big tech, financial institutions, and life sciences companies to date with customers such as Meta, IBM, Microsoft, and Jane Street.”
“We were founded in September 2017 and launched our CoreWeave Cloud Platform in 2020. Moreover, prior to 2022, we had limited revenue, most of which was derived from our crypto mining offerings, which we have discontinued.”
“A substantial portion of our revenue is driven by a limited number of customers. We recognized an aggregate of approximately 67% of our revenue from our top customer, Microsoft, for the year ended December 31, 2025.”
Share of revenue from Customer A (the filing does not name the customer) · 2026-CQ2
Share of revenue from Customer B (the filing does not name the customer) · 2026-CQ2
Share of revenue from Customer C (the filing does not name the customer) · 2026-CQ2
Share of revenue from the then Customer A, prior-year quarter · 2025-CQ2
Revenue from Customer A in the prior-year quarter (its share times revenue) · 2025-CQ2
Microsoft share of revenue, fiscal 2025 (2025-CQ4 reference 10-K) · FY2025
Revenue no customer other than Microsoft reached in fiscal 2025 (the floor in dollars) · FY2025
Share of revenue from the top three customers · 2026-CQ2
Revenue · 2026-CQ2
Revenue, prior-year quarter · 2025-CQ2
Revenue backlog (not revenue) · as-of 2026-06-30
Remaining performance obligations (backlog, not revenue) · as-of 2026-06-30
What else could explain it
other: Customer B is not resolved between OpenAI and Meta; whether it belongs in this channel or the lab channel is the ledger’s judgment, and the range spans both.
mix shift: Shares move with deliveries: Microsoft’s share has fallen each quarter as labs and enterprises ramp.
line composition: Part of the capacity may serve high-performance computing that is not AI.
Quotes
“We generate revenue by providing our customers with access to cloud computing services, including compute enabled by our software and infrastructure optimized for AI and high-performance computing.”
“We recognized an aggregate of approximately 36%, 26%, and 10% of our revenue from our top three customers for the three months ended June 30, 2026, with no other customer representing 10% or more of our revenue for that period.”
“For example, in April 2026, we announced that Jane Street, a private company, had committed approximately $6.0 billion to use our AI cloud platform. Other significant customers include Microsoft and OpenAI.”
“Similarly, in March 2026, we entered into an order form under an existing master services agreement pursuant to which Meta Platforms, Inc. (“Meta”) initially committed to pay us up to approximately $21.0 billion (inclusive of access to new computing capacity through December 20, 2032 and the exercise of an existing option to access additional computing capacity through April 10, 2032) and, as a result, we also expect Meta to be a significant customer in future periods.”
“Although we currently generate the majority of our revenue from large, established customers in the AI industry, we intend to increase the number of our customers over time, including customers in their early stages and/or private companies that may have increased risk of insolvency, bankruptcy, or other issues impacting their creditworthiness.”
“Moreover, our committed contracts typically include a prepayment from our customers prior to them receiving access to our services. As of December 31, 2025, 2024, and 2023, the weighted-average prepayment across all our active contracts was 15% to 25% of the TCV.”
“Approximately 93% of the revenue increase in both the three and six months ended June 30, 2026 was attributable to expansion within our existing customer base, with the remainder attributable to new customers.”
“These delayed draw term loan facilities are collateralized with the assets underlying the contributed contracts and the pledged contractual cash flows, generally from investment grade counterparties. They are drawn as we build infrastructure to support customer requirements, and amortize over time as contracted cash flows are generated in a regular and predictable manner, with excess cash made available to us.”
“We generated record revenue of $2.6 billion, up 112% year-over-year, increased revenue backlog to $104 billion, while driving rapidly expanding enterprise adoption. This figure does not include the over $25 billion of net new customer commitments added in the early weeks of Q3.”
“In Q2, the customer contracts we signed came with contribution margins we expect to be 5 percentage points- 10 percentage points above those added in recent quarters.”
“Pricing and margins for our Blackwell and Vera Rubin SKUs are setting new highs, while pricing for prior generation SKUs is at or above where it was years ago. Our near-term capacity remains effectively sold out.”
“This operating margin improvement came before our July pricing changes, which included an approximately 25% increase across SKUs in response to the current demand environment and the increasing ROI our customers are observing from their investments in the CoreWeave platform as they shift to inference. We are also passing through component price increases.”
“The second piece of it is many of our clients are monetizing their products, and so they are more aggressive about coming in and willing to pay us higher margins because they need access to the compute that will allow them to be successful.”
AI cloud capacity rented by AI labs and AI-native companies
10% to 42% of the quarter’s revenue
Incremental total: counts in full.
our inferencedisclosure: described· motive: offensive· before LLMs: new· layer: compute
The CFO’s Q1 bound on the backlog from non-investment-grade AI natives and labs is not repeated; the 10-Q again names OpenAI among the significant customers and repeats that customers may be early stage, private or highly leveraged (c3, c6, c7). New and expanding AI-native customers include Cognition, Periodic Labs, Runway ML and Sunday Robotics (c24, c25). The new tag rests on the LLM labs and LLM-native companies; customers whose models are not language models, such as Isomorphic Labs in drug discovery, Sunday Robotics and Runway ML, strictly fail that test, but none is a significant customer, their share appears small and inside the range, and the range is not changed. The CEO says many customers now monetize their products and are willing to pay higher margins for compute (c56), a description of payers with revenue, not of how they fund it. The size, , is the ledger’s own, an assumed share of revenue: Customer A is Microsoft (see the hyperscaler reading), and Customer B is OpenAI or Meta and is not resolved, so the range runs from OpenAI below the disclosure floor (Customer C is not pinned to a name either) to OpenAI as Customer B, and the point weights Customer B’s two readings equally. Funding stays read as investor capital, a judgment the stored sources do not confirm. The contract incentive paid in the company’s own shares with the first OpenAI agreement, recorded in Q1, was issued under the terms of that agreement for no proceeds (crwv-anchor-c5, crwv-anchor-c23): a price term, not money put into the lab, and the company holds no stake in any lab in the stored sources, so the lab rule does not make the funding mixed. Counterparty mixed: AI labs and AI-native companies.
Evidence: 20 quotes, 4 figures, 3 confounds, 7 from before coverage
Capacity rented by companies whose business is AI: foundation-model labs (OpenAI, Anthropic, Mistral, Cohere) and AI-native companies (Perplexity, Cognition, and physical AI and world-model developers). The 10-Q names OpenAI, a private company, among the significant customers, and the CFO bounds the backlog from non-investment-grade AI natives and foundation labs. The payer exists only because of LLMs, so the money is new even where the product is old. These customers are private and in the filing’s words may be early stage or highly leveraged; the stored sources do not say how each pays, and the ledger reads their funding as investor capital. With its first contract the company also issued Class A shares to OpenAI, under the terms of that contract and for no proceeds, which the filings book as a contract incentive, consideration paid to the customer. That is a price term on the company’s own revenue, not money put into the lab: the company holds no stake in OpenAI or any other lab and has extended them no credit in the stored sources, so the lab rule reads the funding as investor capital, not mixed. Sized as an assumed share of reported revenue, since Customer B is OpenAI or Meta and the filings do not say which. The new tag rests on payers that exist because of LLMs: the labs and LLM-native companies. Some AI-native customers do not rest on LLMs (Isomorphic Labs in drug discovery, robotics and physical-AI developers such as Sunday Robotics, world-model and video-model developers); strictly their part is expanded, not new. None is a significant customer, their share appears small and inside the range, and they are left in this channel.
Why this motive
Buyers paying more for a separately priced AI product and, in the CEO’s words, monetizing their own products (c56, c33).
Before LLMs: new
At the anchor the AI lab customers named were Cohere, Mistral AI and OpenAI, and OpenAI had committed up to under its first agreement. The payers are AI labs and AI-native companies, which exist because of LLMs, so the tag is new whatever the anchor shows; no revenue was given for them as a group.
“Our AI Lab customers include, among others, Cohere, Mistral AI, and OpenAI.”
“As of March 11, 2025, subject to any termination described below and satisfaction of delivery and availability of service requirements, OpenAI has committed to pay us up to approximately $11.9 billion through October 2030.”
“We have agreed to issue to OpenAI, in accordance with the terms of the OpenAI Master Services Agreement and pursuant to a stock issuance agreement (the “OpenAI Stock Issuance Agreement”) with OpenAI, a number of shares of our Class A common stock equal to $350.0 million valued at a price per share equal to the initial public offering price.”
“A substantial portion of our revenue is driven by a limited number of customers. We recognized an aggregate of approximately 67% of our revenue from our top customer, Microsoft, for the year ended December 31, 2025.”
Share of revenue from Customer B (the filing does not name the customer) · 2026-CQ2
Share of revenue from the top three customers · 2026-CQ2
Share of the year-over-year revenue increase from existing customers · 2026-CQ2
Revenue backlog (not revenue) · as-of 2026-06-30
What else could explain it
other: Customer B is not resolved between OpenAI and Meta; the point weights the two readings equally and the range spans both.
other: Funding is inferred from the filing’s description of these customers, not stated by them.
bundling: Consideration paid to a customer reduces revenue as it is recognized, so reported revenue from OpenAI may be net of the contract incentive’s amortization.
Quotes
“We generate revenue by providing our customers with access to cloud computing services, including compute enabled by our software and infrastructure optimized for AI and high-performance computing.”
“We recognized an aggregate of approximately 36%, 26%, and 10% of our revenue from our top three customers for the three months ended June 30, 2026, with no other customer representing 10% or more of our revenue for that period.”
“For example, in April 2026, we announced that Jane Street, a private company, had committed approximately $6.0 billion to use our AI cloud platform. Other significant customers include Microsoft and OpenAI.”
“In May 2025, we entered into a master services agreement with OpenAI OpCo, LLC ("OpenAI") and in September 2025, we entered into an order form under this master services agreement pursuant to which OpenAI committed to pay us up to approximately $6.5 billion through May 31, 2031 and, as a result, we expect OpenAI to be a significant customer in future periods.”
“Although we currently generate the majority of our revenue from large, established customers in the AI industry, we intend to increase the number of our customers over time, including customers in their early stages and/or private companies that may have increased risk of insolvency, bankruptcy, or other issues impacting their creditworthiness.”
“In addition, some of our customers may be highly leveraged and subject to their own operating, financial and regulatory risks and, even if our credit review and analysis mechanisms work properly, we may experience risks of non-payment and non-performance in our dealings with such parties.”
“Moreover, our committed contracts typically include a prepayment from our customers prior to them receiving access to our services. As of December 31, 2025, 2024, and 2023, the weighted-average prepayment across all our active contracts was 15% to 25% of the TCV.”
“Approximately 93% of the revenue increase in both the three and six months ended June 30, 2026 was attributable to expansion within our existing customer base, with the remainder attributable to new customers.”
“Expanded relationships with existing enterprise and AI native customers including Cognition, Databricks, Hudson River Trading, Periodic Labs, Rescale, and Runway ML”
“We generated record revenue of $2.6 billion, up 112% year-over-year, increased revenue backlog to $104 billion, while driving rapidly expanding enterprise adoption. This figure does not include the over $25 billion of net new customer commitments added in the early weeks of Q3.”
“AI is no longer confined to frontier model labs. It is becoming embedded in software, industrial systems, financial markets, enterprise workflows, and national security missions.”
“The future of AI is being built by a new class of innovators. Some are AI- native companies operating at the frontier. Others are change makers inside established enterprises who are willing to challenge the status quo. CoreWeave serves both.”
“Pricing and margins for our Blackwell and Vera Rubin SKUs are setting new highs, while pricing for prior generation SKUs is at or above where it was years ago. Our near-term capacity remains effectively sold out.”
“This operating margin improvement came before our July pricing changes, which included an approximately 25% increase across SKUs in response to the current demand environment and the increasing ROI our customers are observing from their investments in the CoreWeave platform as they shift to inference. We are also passing through component price increases.”
“The second piece of it is many of our clients are monetizing their products, and so they are more aggressive about coming in and willing to pay us higher margins because they need access to the compute that will allow them to be successful.”
“In Q2, the customer contracts we signed came with contribution margins we expect to be 5 percentage points- 10 percentage points above those added in recent quarters.”
AI cloud capacity rented by established enterprises
15% to 38% of the quarter’s revenue
Incremental total: counts at zero at the point and in full at the high end, with no traced baseline.
our inferencedisclosure: described· motive: offensive· before LLMs: expanded· layer: compute
The 10-Q names Jane Street, a private company, as a significant customer with about committed, and the release reports a equity investment from it, a customer financing its supplier (c3, c26). New enterprise customers include Caterpillar, Bentley Systems and further trading firms (c34, c35). The CEO says enterprises want contracts shorter than the usual term, which a new loan structure lets the company finance (c58). About of the revenue increase came from existing customers. The size, , is the ledger’s own, the remainder after the other two payer points, with Customer C possibly Jane Street; at most it is everything outside Customers A and B, since A is Microsoft and B is OpenAI or Meta. These payers fund from operations. IBM, a customer named at the anchor, is on this ledger as IBM, where its rent here is spend; this ledger does not net the two.
Evidence: 19 quotes, 5 figures, 2 confounds, 6 from before coverage
Capacity rented by companies whose business is not AI: systematic trading firms (Jane Street, Hudson River Trading, Flow Traders, IMC), industrial and life-sciences companies (Caterpillar), IBM, and public-sector work through partners. The CEO says the trading firms scale their core machine learning workloads and are not AI labs. These payers fund the rent from operations. IBM is on this ledger as IBM, where its rent here is spend; this ledger records each company’s own flows and does not net them. Sized as an assumed share of reported revenue, at most the revenue outside Customers A and B; some of the work is machine learning that predates LLMs.
Why this motive
Enterprises committing to a separately priced AI product: Jane Street’s commitment and investment, trading firms building AI models for quantitative trading, and Caterpillar training physical AI (c3, c35, c34).
Before LLMs: expanded
At the anchor the company already counted IBM and Jane Street among its AI Enterprises and had added new enterprise customers in Q1 2025, inside revenue of for that quarter. Enterprises renting GPU capacity for machine learning predate LLMs; no quarter of the same activity before AI can be traced. The size is the whole of an activity that existed before.
“Large companies whose business and product are not AI, but are being driven by AI. These include Fortune 500 companies and other large enterprises with business models that are increasingly AI-enabled through their own development efforts and/or that leverage AI directly to drive internal efficiency gains. We have seen initial traction in adopting our platform from big tech, financial institutions, and life sciences companies to date with customers such as Meta, IBM, Microsoft, and Jane Street.”
“We were founded in September 2017 and launched our CoreWeave Cloud Platform in 2020. Moreover, prior to 2022, we had limited revenue, most of which was derived from our crypto mining offerings, which we have discontinued.”
“In particular, we are excited to see the broad-based increase in demand for inference as well as the accelerating adoption of AI by our enterprise customers.”
“A substantial portion of our revenue is driven by a limited number of customers. We recognized an aggregate of approximately 67% of our revenue from our top customer, Microsoft, for the year ended December 31, 2025.”
Jane Street commitment (contract value, not revenue) · as-of 2026-06-30
Equity investment by Jane Street, a customer · 2026-CQ2
Share of revenue from Customer C (the filing does not name the customer) · 2026-CQ2
Share of the year-over-year revenue increase from existing customers · 2026-CQ2
Share of revenue from the top three customers · 2026-CQ2
What else could explain it
line composition: Quantitative trading and scientific computing include machine learning that predates LLMs.
other: The share is a remainder of two assumed shares.
Quotes
“We generate revenue by providing our customers with access to cloud computing services, including compute enabled by our software and infrastructure optimized for AI and high-performance computing.”
“We recognized an aggregate of approximately 36%, 26%, and 10% of our revenue from our top three customers for the three months ended June 30, 2026, with no other customer representing 10% or more of our revenue for that period.”
“For example, in April 2026, we announced that Jane Street, a private company, had committed approximately $6.0 billion to use our AI cloud platform. Other significant customers include Microsoft and OpenAI.”
“Approximately 93% of the revenue increase in both the three and six months ended June 30, 2026 was attributable to expansion within our existing customer base, with the remainder attributable to new customers.”
“Expanded relationships with existing enterprise and AI native customers including Cognition, Databricks, Hudson River Trading, Periodic Labs, Rescale, and Runway ML”
“We generated record revenue of $2.6 billion, up 112% year-over-year, increased revenue backlog to $104 billion, while driving rapidly expanding enterprise adoption. This figure does not include the over $25 billion of net new customer commitments added in the early weeks of Q3.”
“AI is no longer confined to frontier model labs. It is becoming embedded in software, industrial systems, financial markets, enterprise workflows, and national security missions.”
“We recently added Flow Traders and IMC to our growing roster of systematic trading firms. These customers are using CoreWeave to develop and deploy the next generation of AI models for quantitative trading.”
“The future of AI is being built by a new class of innovators. Some are AI- native companies operating at the frontier. Others are change makers inside established enterprises who are willing to challenge the status quo. CoreWeave serves both.”
“Pricing and margins for our Blackwell and Vera Rubin SKUs are setting new highs, while pricing for prior generation SKUs is at or above where it was years ago. Our near-term capacity remains effectively sold out.”
“This operating margin improvement came before our July pricing changes, which included an approximately 25% increase across SKUs in response to the current demand environment and the increasing ROI our customers are observing from their investments in the CoreWeave platform as they shift to inference. We are also passing through component price increases.”
“The second piece of it is many of our clients are monetizing their products, and so they are more aggressive about coming in and willing to pay us higher margins because they need access to the compute that will allow them to be successful.”
“These clients want to buy compute for two years or three years, and that is not a market that was easily accessible to us until we were able to bring DDTL 5.5 to market.”
“This financing is significant as it unlocks our ability to serve critical part of the enterprise market at scale, while also allowing us to accelerate the ramp of our managed inference platform and grow our exposure to shorter-dated contracts that typically come at a higher ASP and margins.”
Managed inference: serverless and dedicated model serving, priced by the token
0.1% to 0.64% of the quarter’s revenue; overlaps another channel, not added into totals
our inferencedisclosure: quantified· motive: offensive· before LLMs: expanded· layer: compute
The CEO says managed inference, serverless and dedicated, monetizes tokens, and its booked annual recurring revenue grew from to more than since launch, with at least expected at year end (c38, c39). Grammarly and You.com run production traffic on it (c40). A run rate is not the quarter’s revenue: the size, , is the ledger’s own, a quarter of the booked run rate at the call date, six weeks after quarter end, scaled at every end of the range for how little of it was in force during the quarter. The dollars sit inside the AI-native and enterprise channels, so this reading overlaps them and is left out of the flow total. Funding mixed: AI-native and established software companies. The funding reading, mixed, names payer groups, not a stake: AI labs and AI-native companies pay from outside funding rounds, with no money of the company’s own put into any of them in the stored sources (the shares issued to OpenAI were a contract incentive for no proceeds, crwv-anchor-c23), and established companies pay from operations.
Evidence: 6 quotes, 3 figures, 2 confounds, 1 from before coverage
Model serving the company runs for customers, serverless on open models or on dedicated deployments, which the CEO describes as monetizing tokens. Dedicated deployments offer production inference on GPU types the customer picks, work the anchor already shows customers running on rented capacity; serverless serving priced by the token is new. With both readings supported, the tie-break takes expanded, sized as a level with no traceable baseline. Introduced in Q1 2026 as Dedicated Inference; management reports booked annual recurring revenue from Q2 2026 and names Grammarly and You.com among the users, a mix of AI-native and established software companies. It runs partly on GPUs coming off their first contracts. Its dollars are inside the capacity revenue read by payer, so this channel overlaps those channels and is left out of the flow total.
Why this motive
A separately priced AI product with measured growth: booked run rate from to more than , priced on tokens (c38, c39).
Before LLMs: expanded
At the anchor customers already ran inference on rented capacity, priced by GPU and term, and the CEO described broad demand for it; no product was priced per token and no revenue was given for inference, so the level has no baseline. The size is the whole of an activity that existed before.
“In particular, we are excited to see the broad-based increase in demand for inference as well as the accelerating adoption of AI by our enterprise customers.”
Booked annual recurring revenue of managed inference (a floor) · as-of 2026-08-11
Booked annual recurring revenue of managed inference at launch (launch date not stated; taken as Q1 quarter end) · as-of 2026-03-31
Managed inference annual recurring revenue expected at the end of 2026 (a floor; a target) · as-of 2026-12-31
What else could explain it
other: Booked run rate counts contracts signed, not revenue earned; part may start after the quarter.
relabel: Dedicated deployments may be capacity the same customers rented before, now counted in this product.
Quotes
“We have seen an explosion of growth in our managed inference platform in the few months since its launch, with growth constrained only by our near-term capacity. Across serverless offerings and dedicated deployments, CoreWeave is monetizing tokens while giving customers flexibility in how they consume our platform.”
“In the past few months since its launch, booked ARR for our managed inference platform has grown from $1 million to more than $100 million. We expect to exit 2026 with at least $250 million of managed inference ARR.”
“Companies such as Grammarly and You.com are moving from experimentation to real production traffic, running AI coding agents, serving their own fine-tuned models, and deploying open weight models at scale.”
“This financing is significant as it unlocks our ability to serve critical part of the enterprise market at scale, while also allowing us to accelerate the ramp of our managed inference platform and grow our exposure to shorter-dated contracts that typically come at a higher ASP and margins.”
“We really think about the fact that that is an incredible opportunity for us to offer the most bleeding-edge compute that we have, but also a wonderful way for us to access and use GPUs that are coming off contract in a way to extract maximum value for the company over time.”
“When we think about the lessons that we have learned as we have gone through this unbelievable and unique scaling of our managed inference product, which went from $1 million to $100 million inside of a single quarter.”
Storage, CPU, networking and AI developer software sold beside GPU capacity
3.3% to 5.8% of the quarter’s revenue; overlaps another channel, not added into totals
our inferencedisclosure: quantified· motive: offensive· before LLMs: expanded· layer: compute
The CFO says storage, CPU, networking and software exceed of annual recurring revenue as customers consolidate spend (c46). The CEO says the AI development services carry higher margins and reach more customers than the core cloud, and that more than model training runs are tracked on the platform, a count of AI work (c41, c42). An AI research agent and sandboxes were launched (c29). The size, , is the ledger’s own, a quarter of the run rate scaled for growth. The buyers are the same payers, so the reading overlaps the payer channels and is left out of the flow total. The funding reading, mixed, names payer groups, not a stake: AI labs and AI-native companies pay from outside funding rounds, with no money of the company’s own put into any of them in the stored sources (the shares issued to OpenAI were a contract incentive for no proceeds, crwv-anchor-c23), and established companies pay from operations.
Evidence: 6 quotes, 2 figures, 2 confounds, 3 from before coverage
The services the company sells around its GPU clusters: AI object storage, CPU compute, networking and software, including Weights & Biases, the AI developer platform acquired in 2025, and the agent and evaluation tools built on it. Management gives annual recurring revenue for the group and adoption counts, not revenue. The buyers are the same labs, platforms and enterprises that rent capacity, so the channel overlaps the payer channels and is left out of the flow total. Weights & Biases is an AI company whose product is in this channel, so its revenue counts here; the 10-Q calls its results not material.
Why this motive
Customers consolidating spend on separately priced services beside GPUs, with a stated run rate (c46).
Before LLMs: expanded
At the anchor the company already sold managed software and application services and storage built for AI training and inference beside its GPU capacity, and had just acquired Weights & Biases with more than customers. No revenue was given for these services. The size is the whole of an activity that existed before.
“We generate revenue by offering world-class AI infrastructure and proprietary managed software and application services through our CoreWeave Cloud Platform.”
“As we continue to build our infrastructure and software differentiation, we released our next-generation CoreWeave AI Object Storage, purpose-built for the most demanding training and inference use cases.”
“We announced and completed our acquisition of Weights & Biases, one of the industry's leading platforms for AI developers with more than 1,400 customers.”
Annual recurring revenue of storage, CPU, networking and software (a floor) · as-of 2026-06-30
Model training runs tracked on the platform (a floor; a count of AI work) · as-of 2026-08-11
What else could explain it
bundling: These services are often sold inside the same reserved-capacity contracts; the split from GPU capacity is management’s.
acquisition: Weights & Biases, acquired in 2025, is inside software; an AI company whose product sits in this channel.
Quotes
“The transaction extended the Company's application software services offering to include additional developer-focused capabilities for the training of models and development of AI applications.”
“CoreWeave ARIA (AI Research and Iteration Agent) that reads experiment data, uncovers hidden insights, and drives continuous model and agent improvement”
“The scale of our AI products and services beyond GPUs also continues to ramp significantly as customers consolidate spend with us. These margin-accretive businesses including storage, CPU, networking, and software already exceed $400 million of ARR as of Q2. We expect they will continue to expand rapidly.”
“CoreWeave Sandboxes, the execution layer that gives AI researchers and platform teams secure, isolated environments for running reinforcement learning, agent tool use, and model evaluation”
CoreWeave Omni: the cloud stack run in customers’ own data centers on their GPUs
Not sized
Not yet earning revenue: the first deal begins to scale in 2027, so the channel is read described and left unsized, never sized at zero.
described, no sizedisclosure: described· motive: exploratory· before LLMs: expanded· layer: compute
The CEO says the first CoreWeave Omni deal was signed in the weeks before the call and begins to scale in 2027, with interest from sovereign, enterprise and cloud customers (c37).
Evidence: 1 quote, 1 from before coverage
Software and operation of the company’s full cloud stack in a customer’s data center with the customer’s own GPUs, offered to sovereign, enterprise and cloud customers. Announced in Q1 2026; the first deal was signed in Q2 2026 and is to scale in 2027, so no revenue is in the covered quarters.
Why this motive
A first deal signed with scale from 2027 and no revenue in the quarter (c37).
Before LLMs: expanded
At the anchor the company already sold managed software and application services on its own platform; running that stack on hardware the customer owns is not in the anchor. The size is the change AI made, not the whole line.
“We generate revenue by offering world-class AI infrastructure and proprietary managed software and application services through our CoreWeave Cloud Platform.”
“Through CoreWeave Omni, we are seeing significant interest from sovereign, enterprise, and cloud customers alike. In the past few weeks, we signed our first deal, which will begin to scale in 2027.”
Q1 2026described · described, no size · exploratory
Q2 2026described · described, no size · exploratory
Reported lines, year-over-year growth
Revenue +112.3%
Q2 2026. Growing slower than revenue: sales and marketing (+63.0%), general and administrative (+2.2%). A displaced cost shows up as a line that stays under the dashed revenue line. These are the audited lines, as first reported; nothing here is attributed to AI by the filing. Not drawn: cost of revenue, sales and marketing, general and administrative, total operating expenses, which one-time items move by more than 60% in a quarter; the values are in the table below.