What Does CoreWeave Do? AI Cloud Business Explained

CoreWeave is a specialized cloud computing company that rents access to NVIDIA GPU infrastructure and the software needed to train, fine-tune, and run AI models. It does not make chips or build a consumer chatbot. It operates the cloud layer between AI developers and the physical data centers, GPUs, storage, and networks their workloads require.

The simplest description is an AI-focused alternative to a general-purpose cloud. Amazon Web Services, Microsoft Azure, and Google Cloud support almost every kind of business software. CoreWeave concentrates on a narrower problem: running large, GPU-intensive AI and high-performance computing workloads efficiently at scale.

What is CoreWeave?

CoreWeave is often called a neocloud, meaning a newer cloud provider built around accelerated computing rather than general business IT. In its 2025 Form 10-K, CoreWeave describes its platform as infrastructure, proprietary software, orchestration, and managed cloud services designed for the full AI lifecycle.

That lifecycle includes four compute-heavy jobs:

1. Training: processing large datasets across many connected GPUs to create a model.
2. Fine-tuning and reinforcement learning: adapting a model for a domain, task, or agent workflow.
3. Inference: running a trained model to generate answers, images, predictions, or actions.
4. High-performance computing: using dense clusters for simulations, rendering, and other parallel workloads.

CoreWeave was founded in September 2017 and initially generated most of its limited early revenue from crypto-mining services. The company says those offerings were discontinued before 2022. It launched the CoreWeave Cloud Platform in 2020 and listed on Nasdaq under the ticker CRWV in March 2025.

The crypto origin matters because both businesses depend on operating large fleets of graphics processors. CoreWeave’s pivot was from using GPUs to perform mining calculations to renting GPU capacity and related services to other companies.

What does CoreWeave actually sell?

CoreWeave sells more than hourly access to a graphics card. Its May 2026 product documentation divides the platform into compute, storage, networking, and software services that work together.

Product layer	What CoreWeave provides	What the customer uses it for
GPU compute	Bare-metal NVIDIA GPU nodes and large connected clusters	Training, fine-tuning, reinforcement learning, inference, and HPC
Orchestration	CoreWeave Kubernetes Service and Slurm on Kubernetes, called SUNK	Scheduling work across many machines without building the control layer from scratch
Inference	Dedicated and self-managed model-serving options	Deploying custom or open-weight models into production
Storage	Object, distributed file, dedicated, and local storage	Feeding datasets and model weights to GPUs, saving checkpoints, and sharing data between nodes
Networking	Private cloud networks, high-speed GPU interconnects, and direct connections	Moving data quickly within a cluster and between clouds or facilities
Operations	Mission Control observability and automated cluster-health tooling	Finding failing nodes, monitoring GPU performance, and reducing interrupted jobs
Developer tools	Weights & Biases and other acquired model-development products	Tracking experiments, evaluating agents, fine-tuning models, and monitoring applications

The value of that bundle is easiest to see during a large training run. Renting thousands of GPUs is not enough. Those GPUs must communicate with low latency, receive data quickly, save checkpoints, recover from failures, and stay busy. An expensive accelerator that is waiting for storage or another node is paid-for capacity producing no useful work.

The CoreWeave AI cloud stack

Layer	What it controls
Developer and model tools	Track experiments, tune models, evaluate agents, and monitor production
Control and orchestration	Mission Control, managed Kubernetes, and Slurm schedule the work
Storage and networking	High-speed data services keep large GPU clusters supplied and connected
Accelerated compute	NVIDIA GPU clusters run training, fine-tuning, inference, and HPC
Power and data centers	Physical capacity, cooling, and fiber support every service above

How does CoreWeave make money?

CoreWeave makes money by charging customers for cloud infrastructure and platform services. The underlying economic unit is usually reserved computing capacity or metered usage, not a subscription to an AI model.

The company has two main routes to revenue:

• Committed capacity: Large customers sign multi-year contracts for a specified amount of infrastructure. CoreWeave’s 2025 annual report says these agreements are generally take-or-pay, so the customer commits to paying for capacity over the contract term. The weighted-average duration of committed contracts was about five years at the end of 2025.
• On-demand usage: Customers can rent available infrastructure without the same long commitment. CoreWeave also lists spot and flexible-capacity options for interruptible or variable workloads.

This structure helps explain how CoreWeave finances growth. A long-term customer commitment can support borrowing for the GPUs and data-center capacity needed to serve that contract. CoreWeave says it primarily funds infrastructure with asset-level debt supported by contracted cash flows, supplemented by corporate debt and equity.

The public price list shows what smaller units of that capacity can cost. As of July 16, 2026, CoreWeave’s North American on-demand pricing listed these eight-GPU nodes:

NVIDIA system	GPUs per node	On-demand node price	Effective price per GPU-hour
HGX A100	8	$21.60/hour	$2.70
HGX H100	8	$49.24/hour	$6.16
HGX H200	8	$50.44/hour	$6.31
HGX B200	8	$68.80/hour	$8.60

Those list prices are not a complete enterprise cost comparison. A real bill can include storage, networking, support, idle capacity, and different contract terms. Large customers may also negotiate arrangements that do not resemble the public on-demand page. For a broader comparison of hourly GPU markets, see the site’s analysis of decentralized GPU pricing versus specialist and hyperscale clouds.

Who uses CoreWeave?

CoreWeave serves three broad customer groups: AI labs building models, hyperscalers that need additional AI capacity, and enterprises running large computing workloads.

Its first-quarter 2026 results named agreements or relationships with Meta, Anthropic, Cohere, Jane Street, Mistral, Perplexity, and World Labs. The 2025 annual report also identifies Microsoft and OpenAI as significant customers.

That list reveals an unusual feature of the AI infrastructure market: a company can be both a CoreWeave customer and a competitor. Microsoft sells Azure cloud services but also buys capacity from CoreWeave. Large technology companies may need more GPU capacity, a different location, or faster access to a new NVIDIA system than their own data-center schedules can provide.

Customer concentration remains important. CoreWeave reported that Microsoft generated approximately 67% of its 2025 revenue. The company has since announced large agreements with additional customers, but a broader named customer list does not by itself show how evenly future revenue will be distributed.

What is the relationship between CoreWeave and NVIDIA?

CoreWeave and NVIDIA have a close supplier, investor, and technology-partner relationship. They are not the same company.

NVIDIA designs the accelerators and networking technology used in CoreWeave’s clusters. CoreWeave buys and operates that hardware, combines it with its cloud software and facilities, and sells computing services to customers. A 2026 CoreWeave filing said all GPUs then used in its infrastructure were NVIDIA GPUs because of requirements in customer contracts, and NVIDIA accounted for 17% of its 2025 supplier purchases.

The relationship extends beyond hardware sales. In January 2026, NVIDIA invested $2 billion in CoreWeave at $87.20 per Class A share. The companies also announced plans to collaborate on more than 5 gigawatts of AI-factory capacity by 2030 and to integrate future NVIDIA computing, CPU, storage, and networking platforms.

CoreWeave is still an independent public company, not an NVIDIA subsidiary. NVIDIA benefits when CoreWeave deploys more NVIDIA systems, while CoreWeave benefits from early technical alignment with the hardware supplier most requested by its customers. The same closeness creates dependency: a supply disruption, product delay, or change in customer hardware preferences could affect CoreWeave more than a cloud with a broader accelerator mix.

How is CoreWeave different from AWS, Azure, and Google Cloud?

CoreWeave’s main distinction is specialization. It builds its platform and operating processes around dense AI clusters, while the largest clouds must also support databases, office applications, web hosting, analytics, and thousands of other services.

Question	CoreWeave	General-purpose hyperscaler
Primary focus	AI and high-performance computing	Broad enterprise and consumer cloud workloads
Core compute product	Connected NVIDIA GPU capacity	GPUs plus a much wider catalog of CPUs and managed services
Workload control	Kubernetes, Slurm, and AI-specific cluster tooling	Provider-specific services across a larger ecosystem
Main appeal	Specialized operations, current GPU systems, and large AI clusters	Breadth, global enterprise relationships, and integrated services
Main tradeoff	Smaller ecosystem and greater supplier concentration	More platform complexity and infrastructure not exclusively designed for AI

CoreWeave itself names AWS, Google Cloud, Microsoft Azure, and Oracle as key general-purpose competitors in its annual report. Specialist GPU clouds also compete for parts of the same market. The right provider depends on the required GPU type, cluster size, region, contract length, data-transfer pattern, and software stack. The site’s AI inference provider directory maps CoreWeave alongside model APIs, inference platforms, and other infrastructure specialists.

Calling CoreWeave “cheaper AWS for GPUs” misses the important part. Sticker price matters, but the useful output is a completed training run or reliably served model. GPU utilization, network performance, checkpoint recovery, available capacity, and engineering time can outweigh a small difference in hourly price.

Why has CoreWeave grown so quickly?

CoreWeave’s growth comes from matching a scarce, expensive input with customers willing to commit years of spending to secure it. AI labs and technology companies need clusters of modern accelerators, sufficient power, high-speed networking, and operating expertise. Building that stack internally takes capital and time.

The scale is visible in the financial results. CoreWeave reported first-quarter 2026 revenue of $2.078 billion, up from $982 million a year earlier. Revenue backlog reached $99.4 billion, active power passed 1 gigawatt, and contracted power exceeded 3.5 gigawatts.

Those numbers show demand and planned capacity, but not an effortless business. CoreWeave must obtain land, power, buildings, GPUs, networking equipment, and financing before much of the contracted revenue can be recognized. The site’s guide to AI data-center financing explains why private credit, asset-backed structures, and long customer commitments have become central to this buildout.

What are the biggest risks in CoreWeave’s model?

CoreWeave’s advantage and its main risks come from the same decision: focus heavily on large-scale AI infrastructure.

1. Capital intensity. New capacity requires large up-front spending and financing. In Q1 2026, CoreWeave reported $2.078 billion in revenue but a $144 million operating loss, $536 million of net interest expense, and a $740 million net loss. Its adjusted EBITDA was positive, but that non-GAAP measure does not remove the cash and financing demands of expansion.
2. Customer concentration. A few very large contracts can support rapid construction, but losing or renegotiating one major relationship can leave expensive capacity underused.
3. NVIDIA dependence. Close alignment gives CoreWeave access and technical focus, yet it concentrates the supply chain around one accelerator vendor.
4. Power and construction constraints. Contracted demand becomes revenue only when sites, electrical capacity, hardware, and networks are ready and performing.
5. Powerful competitors. AWS, Azure, Google Cloud, and Oracle can invest more, bundle services, and use existing enterprise relationships to win workloads.
6. Changing AI economics. More efficient models or a slowdown in AI spending could reduce the amount of centralized compute customers need.

This is why CoreWeave should be understood as both a software platform and an infrastructure-financing business. The software helps make GPU fleets productive. The contracts and financing determine whether the company can build those fleets at the required speed and earn an adequate return on them.

The bottom line

CoreWeave rents the computing foundation used to build and operate AI. Its product is not one GPU or one model. It is an integrated stack of NVIDIA accelerators, data centers, storage, networking, orchestration, observability, and developer tools sold as cloud capacity.

For a customer, CoreWeave can replace years of building specialized AI infrastructure internally. For CoreWeave, the challenge is to keep expensive capacity financed, deployed, and utilized while technology and customer demand change quickly. That combination explains both the company’s rapid rise and the financial risks behind it.

Frequently asked questions

What does CoreWeave do in simple terms?

CoreWeave rents NVIDIA GPU computing and the cloud services around it to companies that train, fine-tune, and run AI models. It provides compute, storage, networking, orchestration, inference infrastructure, and developer tools.

Does CoreWeave make GPUs?

No. NVIDIA designs the GPUs. CoreWeave buys and operates NVIDIA systems in data centers, adds cloud software and managed services, and sells access to that infrastructure.

Is CoreWeave owned by NVIDIA?

No. CoreWeave is an independent company listed on Nasdaq under CRWV. NVIDIA is a major supplier, technology partner, and investor, including a $2 billion share purchase completed in January 2026.

How does CoreWeave make money?

CoreWeave charges for cloud infrastructure and platform services. Large customers often sign multi-year take-or-pay capacity contracts, while on-demand, spot, and flexible usage options serve shorter or variable workloads.

Who are CoreWeave competitors?

CoreWeave names AWS, Google Cloud, Microsoft Azure, and Oracle as key general-purpose competitors. It also competes with specialist GPU and AI cloud providers for training, inference, and high-performance computing workloads.

What did CoreWeave do before AI cloud computing?

CoreWeave says most of its limited revenue before 2022 came from crypto-mining offerings, which it discontinued. It launched its cloud platform in 2020 and shifted its GPU fleet and expertise toward AI and high-performance computing.

Sources

• CoreWeave (2026). Annual Report on Form 10-K for the year ended December 31, 2025. U.S. Securities and Exchange Commission. https://www.sec.gov/Archives/edgar/data/1769628/000176962826000104/crwv-20251231.htm
• CoreWeave (2026). CoreWeave Reports Strong First Quarter 2026 Results. CoreWeave Investor Relations. https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-First-Quarter-2026-Results/
• CoreWeave (2026). Product Home: Compute, Networking, and Storage Products for AI and HPC Workloads. CoreWeave Documentation. https://docs.coreweave.com/product-home
• CoreWeave (2026). CoreWeave Cloud Pricing. CoreWeave. https://www.coreweave.com/pricing
• NVIDIA and CoreWeave (2026). NVIDIA and CoreWeave Strengthen Collaboration to Accelerate Buildout of AI Factories. CoreWeave Investor Relations. https://investors.coreweave.com/news/news-details/2026/NVIDIA-and-CoreWeave-Strengthen-Collaboration-to-Accelerate-Buildout-of-AI-Factories/default.aspx