In the story most people tell about artificial intelligence, the heroes build models: the chatbots, the image generators, the systems that pass bar exams. But every one of those models runs on a machine somebody had to buy, power, cool, and keep busy. CoreWeave sells the machines. It does not train frontier models or ship a consumer app. It rents out fleets of Nvidia graphics processors, wrapped in software that keeps them running, to the companies that do - and in doing so it has become one of the most consequential firms of the AI era that most consumers have never used.
The pitch is narrow on purpose. CoreWeave describes itself as "The Essential Cloud for AI," a phrase that doubles as a strategy. Where Amazon, Microsoft, and Google build general-purpose clouds that can host anything from a payroll database to a video game, CoreWeave does one thing: it runs AI and high-performance computing workloads on Nvidia hardware, at scale, and claims to do it faster and cheaper than clouds designed for everything else.
01 / OriginsFrom Ethereum rigs to AI racks
CoreWeave did not start as an AI company, or even as a cloud company. In 2017, three commodities traders - Michael Intrator, Brian Venturo, and Brannin McBee - founded a business called Atlantic Crypto to mine Ethereum. The bet was simple: banks of GPUs could throw off strong returns during the crypto boom. The 2018 downturn ended that thesis and left the founders with a warehouse's worth of graphics cards and a question about what else those cards could do.
The answer, in 2019, was to rent the GPUs out. The first customers were visual-effects and animation studios that needed to render frames fast. The company rebranded as CoreWeave and, almost by accident, positioned itself at exactly the right spot for what came next. When generative AI demand exploded, CoreWeave already knew how to buy, rack, and operate GPUs by the thousands - and it already had a relationship with the company whose chips everyone suddenly wanted.
The founders' background matters to how the company operates. Coming from commodities trading rather than Big Tech, they treated GPUs the way a trader treats a scarce asset: something to secure early, contract long, and deploy efficiently. That instinct - buy capacity before the market realizes it needs it - is stamped on the business to this day. The old rendering roots still show, too; VFX and graphics workloads remain part of the platform alongside AI training and inference.
A pile of leftover mining hardware became the foundation for a business now measured in gigawatts.The CoreWeave pivot, 2019
02 / The productNot just chips - the software that keeps them busy
Anyone with enough capital can buy Nvidia GPUs. The harder problem is keeping tens of thousands of them working together without stalling, and that is where CoreWeave concentrates its engineering. The company organizes its offering around a metric it calls "goodput" - the share of time hardware spends actually training or serving a model rather than waiting on data, recovering from a failed node, or sitting idle. On clusters that can cost millions of dollars a day, small improvements in goodput translate into real money.
At the center sits CoreWeave Kubernetes Service (CKS), a GPU-native orchestration layer designed to schedule high-performance workloads across clusters that can exceed 100,000 GPUs. Around it, Mission Control monitors GPU fleets, manages node and cluster lifecycles, and speeds up the detection and repair of the inevitable hardware faults that come with running hardware at this density. Tensorizer, an open-source tool, streams model weights straight from storage so that large models load in seconds instead of minutes. Underneath, purpose-built storage and high-speed InfiniBand and Ethernet networking tie the machines into a single supercomputer-scale system.
03 / The customersThe labs everyone thinks own the AI boom
CoreWeave's client list reads like a who's-who of frontier AI. Microsoft became an anchor customer early, accounting for the majority of 2024 revenue. In 2025, OpenAI signed a set of multi-year infrastructure agreements that grew to roughly $22.4 billion in committed value - a striking arrangement, given that Microsoft is both OpenAI's largest backer and one of CoreWeave's own biggest customers. Meta, Anthropic, IBM, Mistral AI, and the trading firm Jane Street round out a roster that has pushed the company's contracted backlog to about $99.4 billion.
That backlog is the number to watch. These are largely long-term, reserved commitments - customers booking capacity years in advance because GPUs, and the power to run them, are scarce. It is also the reason a company still posting net losses can raise billions: lenders and investors are underwriting future contracts, not this quarter's profit.
04 / The edgeWhy not just use AWS?
CoreWeave's argument against the hyperscalers is specialization. Because its entire stack is tuned for AI on Nvidia silicon - down to the networking topology and the failure-recovery tooling - the company claims workloads can run substantially faster and cheaper than on general-purpose clouds. It is often among the first to deploy each new generation of Nvidia hardware at scale, a timing advantage that matters enormously to labs racing to train the next model.
CoreWeave decided to do exactly one thing, and do it faster than anyone. Focus, it turns out, scales to a $99B backlog.
The trade-off is dependence. CoreWeave is deeply tied to a single supplier in Nvidia - which is also, notably, one of its shareholders - and its fortunes rise and fall with the AI capital-expenditure cycle. The company's answer has been to build moats where it can: securing power, hardware allocation, and long contracts, and expanding up the stack so that customers get more than raw compute.
05 / The strategyBuying the whole stack - and the electricity
Compute is nothing without two things around it: developer tools and power. CoreWeave has been acquiring both. In 2025 it bought Weights & Biases, a widely used platform for tracking AI experiments and managing models, for roughly $1 billion, folding it into the CoreWeave AI Cloud. It followed with OpenPipe, a reinforcement-learning specialist, extending the platform from raw GPUs up to model iteration and deployment.
Then came the least glamorous and possibly most important move: securing electricity. CoreWeave agreed to acquire Core Scientific, a former Bitcoin-mining operation, in an all-stock deal that brings roughly 1.3 gigawatts of data-center power under CoreWeave's control, with expansion room beyond that. In an industry where the binding constraint is increasingly the grid rather than the chip, owning power is owning the future.
06 / The businessSelling shovels, financed by contracts
The business model is capital-intensive and, in its way, elegant. CoreWeave buys hardware and power at scale, wins long-term reserved contracts from customers who need certainty of supply, and uses those contracts - backed by billions in debt raised against them - to fund the next round of buildout. Revenue grew from $229 million in 2023 to $1.9 billion in 2024 to $5.13 billion in 2025. The company is not yet profitable; the losses reflect the enormous cost of building data centers and servicing debt ahead of the revenue those assets will produce.
CoreWeave listed on Nasdaq under the ticker CRWV in March 2025 at roughly a $23 billion valuation, in one of the year's largest US tech IPOs. The public-market debut turned a niche GPU renter into a barometer for how investors feel about the durability of AI spending itself. Early private backers had included Nvidia, Magnetar Capital, Coatue, Fidelity, and Jane Street, and the company had raised more than $10 billion in debt on top of its equity before ever ringing the opening bell.
There is risk baked into all of it. Heavy customer concentration, dependence on a single chip supplier, and a debt load tied to the assumption that AI demand keeps climbing are real exposures, not footnotes. The company's counter is the backlog: tens of billions in contracts signed years out, increasingly spread across more customers. Whether that cushion holds through a downturn is the wager both bulls and skeptics are making.
07 / The stakesWhere it fits in the market
CoreWeave sits in a category that has earned a nickname: the "neocloud," a wave of specialized GPU clouds - alongside names like Lambda, Crusoe, and Nebius - built to serve AI when the hyperscalers could not supply enough capacity fast enough. Its competitors are therefore two-sided: the giants it undercuts on focus, and the upstarts it outruns on scale. Recognition has followed, including a Visionary placement in Gartner's 2026 Magic Quadrant for Cloud AI Infrastructure.
Whether CoreWeave becomes a permanent fixture or a leveraged bet on a single cycle is the open question. What is not in doubt is its position today. When the largest AI labs in the world need somewhere to run their models, a surprising number of them rent from the same landlord. Understanding CoreWeave is, increasingly, a way of understanding the physical machinery beneath the software everyone is talking about.