COREWEAVE INC

Business Model and Scope

CoreWeave builds and operates a specialized cloud for artificial-intelligence workloads. It buys or finances GPUs, servers, networking, storage, power distribution, and related equipment; leases or develops high-density data-center capacity; connects facilities through a network backbone; and layers proprietary orchestration, observability, security, storage, and developer software over the infrastructure. Customers use the system for model training, inference, data movement, agent development, and other high-performance workloads.

The offering spans Infrastructure Services, Managed Software Services, and Application Software Services. CoreWeave Mission Control provisions and monitors distributed infrastructure; its Kubernetes and Slurm capabilities schedule large clusters; storage products feed data to GPUs; Weights & Biases supplies model-development tools. The intent is one AI-native stack rather than separately purchased colocation, hardware, networking, cluster software, and operations.

Revenue is primarily from committed cloud contracts, supplemented by on-demand usage. At year-end 2025, CoreWeave operated 43 data centers with more than 850 megawatts of active power, up from 10 centers and 70 megawatts two years earlier. It had approximately 3.1 gigawatts of contracted power to deploy. Remaining performance obligations, or RPO, were $60.7 billion, up from $15.1 billion a year earlier. RPO is contracted consideration net of estimates for service credits, delays, and resalable capacity; it is not cash, profit, or guaranteed revenue on the original schedule.

Customers and Purchasing Decisions

Frontier AI labs and large technology companies buy rapid access to clusters that would take time, power rights, hardware allocations, networking expertise, and operations teams to assemble internally. The economic need is time-to-compute with high accelerator utilization and reliability. A model developer can reach training or inference capacity sooner, avoid owning hardware, and scale across locations. Enterprises also buy tools that simplify model experimentation, monitoring, and deployment.

Alternatives include Amazon Web Services, Microsoft Azure, Google Cloud, Oracle Cloud, other specialized GPU clouds, colocation with self-owned equipment, and building private data centers. Large customers can negotiate bespoke pricing and long commitments because their purchases justify entire facilities. They can also internalize infrastructure or shift workloads if competing capacity becomes available.

Concentration is extreme. Microsoft supplied approximately 67% of 2025 revenue. Another customer supplied 15%; Microsoft and another customer accounted for 68% and 11% of year-end receivables. OpenAI committed up to approximately $6.5 billion through May 2031 under a 2025 order, and Meta initially committed up to approximately $14.2 billion through December 2031. These contracts improve visibility while increasing counterparty, delivery, and bargaining risk.

Switching costs come from workload tuning, data placement, network topology, scheduling, operational familiarity, and the risk of interrupting large training runs. They are meaningful during an active project but may weaken between model generations. Zero-egress migration and multicloud connections lower barriers to adopting CoreWeave; the same interoperability can help customers leave. Customer loyalty is unproven beyond the short evidence period.

Profit Creation and Value Capture

CoreWeave seeks to earn the spread between contracted compute revenue and the full cost of delivering capacity. That cost includes GPU and server depreciation, data-center rent, power, networking, storage, operations, support, software engineering, customer credits, interest, and eventual hardware replacement. Profit requires high utilization, reliable service, disciplined contract pricing, and an asset life long enough to recover acquisition and financing cost before chips become uneconomic.

Revenue grew from $229 million in 2023 to $1.915 billion in 2024 and $5.131 billion in 2025. In 2025, cost of revenue was $1.453 billion, technology and infrastructure expense $2.929 billion, sales and marketing $144 million, and general and administrative expense $651 million. Operating loss was $46 million. After $1.229 billion of net interest expense, other income, and tax benefit, net loss was $1.167 billion.

The income statement's gross-looking spread is incomplete because server depreciation sits mainly in technology and infrastructure rather than cost of revenue. Depreciation and amortization reached $2.454 billion, while property and equipment purchases were $10.309 billion. Technology expense rose faster than revenue in 2025 as infrastructure depreciation increased about $1.5 billion. A narrow operating loss after heavy depreciation suggests contracted assets are generating contribution, but the large interest burden shows that capital providers currently capture more than common shareholders.

Upfront customer payments also affect cash interpretation. Operating cash flow was $3.058 billion despite the net loss, largely because deferred revenue increased $4.174 billion. That cash finances construction before service, but it creates an obligation to deliver compute. It is not free cash available after the corresponding infrastructure investment.

NVIDIA and other chipmakers capture scarce-hardware economics; OEMs and lenders earn financing returns; data-center owners receive rent; utilities receive power revenue; network suppliers provide essential equipment; employees received $630 million of stock-based compensation. CoreWeave retains value only after all those stakeholders and customers' negotiated service protections are satisfied.

Industry Structure and Capital Cycle

AI cloud infrastructure is growing rapidly but has high fixed capital, long power and construction lead times, fast hardware generations, and concentrated suppliers and buyers. General clouds—AWS, Azure, Google, and Oracle—have broader customer relationships, cheaper funding, mature software ecosystems, and internal AI demand. Specialized clouds compete on early hardware availability, cluster performance, flexibility, engineering support, and price. Some major customers and suppliers are also competitors.

Supplier bargaining power is unusually high. Leading GPUs are available from a limited supplier base, with NVIDIA central to current systems. OEM financing, advanced networking, liquid cooling, transformers, generators, construction labor, and grid interconnection can each constrain deployment. Utilities and regulators decide when power becomes usable. Land and a signed lease do not create revenue without energized, networked, tested equipment.

Customer power is also high because a few buyers contract at enormous scale and can compare clouds or build internally. Long contracts reduce immediate price competition but can lock CoreWeave into economics set before power, construction, and financing costs are fully known. Availability failures create credits and can reduce RPO. Substitutes include more efficient models and chips that require less compute, on-premise capacity, and slower AI adoption.

The capital cycle is decisive. Scarce GPUs and power can support attractive pricing and large commitments. Those returns attract tens of billions of dollars into data centers, accelerators, custom chips, and competing clouds. Capacity arrives with delay, potentially after customer growth slows or model efficiency improves. Because hardware depreciates quickly and debt continues, providers may cut prices to keep clusters utilized, destroying the scarcity margin that justified construction.

Entry requires supplier access, power, financing, engineering, and customer credibility at scale. These barriers are significant today. They may be temporary if hardware supply normalizes, open-source orchestration improves, and capital markets fund rivals. One filing cannot distinguish a durable operating advantage from early-cycle scarcity.

Sources and Durability of Competitive Advantage

CoreWeave's proposed advantage is coordinated speed: procure new accelerators early, design high-density facilities, connect clusters with fast networking and storage, and operate them through software built for AI. If Mission Control, Kubernetes, Slurm, and observability sustain higher GPU utilization and faster deployment than general clouds, customers receive more useful compute per dollar and reach market sooner.

Scale can reinforce the system. Large commitments justify hardware allocations and power projects; more clusters generate operating knowledge; a distributed footprint supports training, inference, and latency needs; software can make capacity fungible. Close supplier and customer relationships can improve planning. Contracted cash flows then support asset-backed financing, which funds additional capacity.

Evidence includes rapid scaling to 850 active megawatts, deployment of newer NVIDIA systems, large RPO, and repeated expansion by existing customers—about 85% of 2025 revenue growth came from existing customers. Yet each indicator has an alternative interpretation. Expansion may reflect industry shortage, RPO may require unbuilt assets, and concentrated customers may have extracted favorable terms. Third-party benchmark claims do not establish through-cycle returns.

The mechanism would endure if CoreWeave consistently deploys faster, operates at higher utilization, and earns lower total cost per useful workload after financing and replacement. It would weaken if general clouds match performance, NVIDIA allocates broadly, customers internalize capacity, software becomes portable, service quality falls during expansion, or capital cost remains structurally above larger competitors.

Operating System and Strategic Trade-offs

The operating chain begins before a customer receives compute. Sales and engineering define a committed workload; finance raises equity, delayed-draw debt, notes, OEM credit, or lease funding; procurement secures chips and network gear; real-estate and power teams obtain sites and interconnection; contractors build cooling and electrical systems; engineers commission clusters; software schedules and monitors workloads; operations maintain uptime and customer success.

The company funds delayed-draw facilities against contractual cash flows and underlying infrastructure. This matches some borrowing to customer commitments but does not eliminate timing risk: debt and construction can precede revenue, and service credits can reduce consideration. OEM arrangements finance specific equipment over one to three years, shortening amortization relative to many customer contracts. Operating and finance leases add site-level fixed obligations.

CoreWeave also acquired Weights & Biases, OpenPipe, Marimo, and Monolith in 2025. Software can increase switching costs and move the company higher in the developer workflow, but acquisitions consume equity and management attention while infrastructure is scaling. Goodwill rose to $1.101 billion from $20 million. The single filing offers no integration record.

Trade-offs define the model: first-to-market hardware versus purchase price and obsolescence; owned control versus lease flexibility; committed utilization versus customer concentration; vertical integration versus managerial complexity; rapid construction versus controls and uptime. A good system must synchronize all layers. Excess GPUs without power, power without customers, or contracts without financing each produce poor returns.

Financial Resilience

At December 31, 2025, CoreWeave had $3.127 billion of cash, $34 million of marketable securities, and $3.701 billion of unused facility commitments, for stated liquidity of $6.862 billion. Against that, current debt was $6.708 billion and noncurrent debt $14.665 billion. Operating lease liabilities were $8.195 billion and finance-lease liabilities $254 million. Deferred revenue totaled $8.185 billion, representing delivery obligations rather than ordinary financing.

Property and equipment was $30.557 billion, more than six times annual revenue. Accounts receivable was $3.169 billion, heavily concentrated. Total liabilities were $45.967 billion against $49.302 billion of assets and $3.335 billion of equity. Interest paid was $869 million in 2025, and net interest expense was $1.229 billion. The capital structure has little room for a simultaneous revenue, asset-value, and refinancing shock.

Debt included $9.7 billion outstanding under delayed-draw facilities, senior and convertible notes, and $3.8 billion of OEM financing, among other obligations. Some loans amortize as contracted cash flows arrive, which improves alignment. It also means customer delay or cancellation can impair the same cash flows supporting debt. Hardware collateral can depreciate rapidly if newer chips deliver better economics.

A severe scenario combines slower AI adoption, a large customer reducing demand or disputing delivery, grid delays, lower spot compute prices, GPU obsolescence, and tighter capital markets. Upfront cash might remain restricted economically by customer obligations, while construction, leases, and interest continue. CoreWeave could slow uncommitted expansion and raise equity, but dilution or distressed asset sales may be required. One filing and three comparative financial years are insufficient to establish survival through that scenario.

Capital Allocation and Shareholder Outcomes

CoreWeave's dominant capital decision is infrastructure deployment. Cash property-and-equipment purchases were $10.309 billion in 2025 and $8.702 billion in 2024, far above operating cash flow. Management expected 2026 investment to increase further. Each cluster should be judged on contracted cash after power, rent, support, interest, expected utilization, residual value, and replacement—not revenue attached to the contract.

Financing has expanded faster than retained earnings. The March 2025 IPO and over-allotment produced about $1.47 billion of net proceeds and converted preferred shares, but 2025 financing cash was $9.308 billion, driven mainly by debt. In December 2025, CoreWeave issued $2.6 billion of convertible notes and spent $340 million on capped calls intended to reduce potential dilution. Capped calls reduce, rather than eliminate, the economic cost of convertible financing.

Weighted-average basic shares doubled from 218 million in 2024 to 435 million in 2025. Year-end common shares were 502 million across classes. Acquisitions, the IPO, preferred conversion, employee awards, and a contract incentive all contributed issuance. Stock-based compensation was $630 million, including $177 million triggered by the IPO, and $1.3 billion remained unrecognized for RSUs and restricted shares.

The company has not paid common dividends and does not expect to in the foreseeable future. Common shareholders receive value only if asset-level cash after financing, replacement, and dilution eventually exceeds the capital supplied. Revenue and RPO growth can coincide with falling value per share when each incremental dollar requires disproportionate debt, leases, or equity.

Legal and Regulatory Exposure

CoreWeave is exposed to export controls and sanctions affecting advanced chips, customers, locations, and technical services. AI regulation can impose model, data, transparency, security, or energy requirements on the company or its customers. Privacy and cybersecurity rules matter because the platform handles valuable models and data; a breach or prolonged outage can create contract credits, claims, regulatory action, and customer loss.

Data-center development requires land, power, environmental, construction, and operating approvals. Grid constraints and energy policy can delay facilities or impose procurement and collateral obligations. Tariffs and trade restrictions can raise equipment cost after a customer price is fixed. Supplier license and intellectual-property disputes could restrict essential software or hardware.

Management identified material weaknesses in internal control over financial reporting and was pursuing remediation as of the filing. Rapid acquisitions, complex financing, leases, revenue estimates, and public-company reporting increase the consequence of weak controls. The multi-class stock structure concentrates voting power with co-founders, limiting ordinary shareholders' influence over financing, acquisitions, and control transactions.

Litigation was incorporated through the commitments note; even absent a quantified material claim, contractual disputes are economically important because a single customer or facility represents substantial value. Regulation can also alter competitive structure: power and export barriers protect scarce capacity while restricting growth and raising capital needs.

Conclusion, Uncertainties and Disconfirming Evidence

CoreWeave creates customer value by assembling scarce accelerators, power, networking, storage, and AI-specific software faster than customers could build them alone. It seeks to retain value through early hardware access, integrated operations, high utilization, workload expertise, and long committed contracts. The 2025 revenue expansion and RPO show demand; they do not yet establish normalized common-shareholder profit.

Contrary evidence dominates the uncertainty assessment. Microsoft supplied 67% of revenue, suppliers are concentrated, interest exceeded operating profit, infrastructure purchases dwarfed operating cash, total debt exceeded $21 billion, lease liabilities exceeded $8 billion, and share count expanded sharply. Operating cash relied heavily on advance customer payments that carry performance obligations. Hardware scarcity and favorable financing may be temporary.

The thesis would be invalidated by major customers not expanding or honoring commitments; persistent construction or power delays; service credits materially reducing RPO; general clouds matching performance and price; GPU useful lives proving shorter than depreciation assumptions; utilization falling as industry capacity arrives; software failing to create measurable retention; debt and lease service consuming cluster cash; or equity issuance outpacing normalized per-share cash generation. With only one annual filing, the appropriate conclusion is conditional: operational capability is visible, but durability, through-cycle return on capital, and ultimate value transfer to common shareholders remain unresolved.