CoreWeave: How a Former Crypto-Mining Business Became a Key AI Compute Provider

CoreWeave is one of the clearest examples of how quickly the economics of computing can change when the right hardware, expertise and market timing come together. Its story began with Ethereum mining rather than artificial intelligence: the founders bought their first graphics processing unit, or GPU, in 2016 and used it to mine Ethereum. The activity developed into a commercial operation during the cryptocurrency boom, but the company did not remain dependent on digital-asset mining. It gradually redirected the same type of high-performance hardware towards visual effects, machine learning and other demanding computing tasks, launched its cloud computing service in 2020 and eventually discontinued its former crypto-mining offerings. By 2026, CoreWeave had become a major supplier of computing capacity to companies including Microsoft, OpenAI, Meta and Anthropic. The scale of the change is striking. Revenue increased from $229 million in 2023 to $1.915 billion in 2024 and $5.131 billion in 2025, while second-quarter 2026 revenue reached $2.575 billion. CoreWeave’s history is therefore not simply a story about a cryptocurrency company changing industries; it shows how experience operating large fleets of GPUs became unexpectedly valuable when artificial intelligence created an unprecedented need for specialised computing infrastructure.

From Ethereum Mining to a Specialist GPU Cloud Business

The origins of CoreWeave pre-date the current generative-AI boom by several years. Co-founders Michael Intrator, Brian Venturo and Brannin McBee began experimenting with GPU-based cryptocurrency mining when Ethereum was still attracting a rapidly expanding community of miners. CoreWeave has said that its first GPU was purchased in 2016 and placed on a pool table in a Lower Manhattan office, where it was used to mine an Ethereum block. What started as an experiment soon became considerably larger. During the cryptocurrency expansion of 2017, the founders obtained additional capital and bought more hardware, eventually moving the operation from an office to a garage and then into a data centre in New Jersey. The company itself was founded in September 2017. SEC records also show that it was previously known as Atlantic Crypto Corp, with the corporate name changing in March 2019. That earlier name reflects how closely the original business was connected to cryptocurrency, although the technical assets being accumulated — GPUs, data-centre capacity and the operational knowledge required to keep thousands of machines running — would later prove useful far beyond mining.

The cryptocurrency downturn of 2018 and 2019 could easily have ended that first business model, but it instead gave the company an opportunity to increase its hardware holdings at lower prices. CoreWeave later said it acquired distressed equipment during the so-called crypto winter, increasing its fleet from hundreds of GPUs to tens of thousands and expanding to seven facilities. This was an unusual position to occupy. A falling cryptocurrency market reduced the immediate economics of mining, but it also meant that equipment acquired during the boom was being sold by operators that could no longer justify the cost of running it. For a company willing to think beyond cryptocurrency, those GPUs still represented valuable computing equipment. Unlike specialised cryptocurrency hardware that may be useful for only a narrow set of calculations, GPUs can perform many highly parallel computing tasks. The same general category of processor used for Ethereum mining could therefore be redirected towards image rendering, scientific work and machine learning. CoreWeave’s future business was not fully established at that stage, but the hardware accumulated during the crypto downturn gave it a substantial physical base from which to experiment with other sources of demand.

The first important diversification came before artificial intelligence became a mainstream business priority. CoreWeave began receiving enquiries from organisations that required GPU computing but found it difficult or expensive to obtain the configurations they wanted from large general-purpose cloud providers. In 2019, the company purchased Leonardo Render and subsequently introduced Concierge Render, moving into services for visual-effects and rendering customers. These workloads were different from cryptocurrency mining, yet they depended on the same fundamental capability: large quantities of GPUs had to be available, connected and managed reliably. The company then extended its focus to machine learning and other high-performance computing applications. CoreWeave formally launched its cloud service in 2020. Its later SEC filings make an important distinction between this new direction and its origins: before 2022, CoreWeave still generated limited revenue and most of it came from its former cryptocurrency-mining activities, but those mining offerings were eventually discontinued. The transition was therefore gradual rather than an overnight decision. CoreWeave first built expertise through crypto, tested other commercial uses for its equipment and only then developed a business centred on providing computing resources to outside customers.

Why the 2018–2019 Crypto Winter Became a Turning Point

The crypto downturn mattered because it forced CoreWeave to confront a basic weakness in mining economics: owning powerful hardware does not guarantee predictable income if that hardware is tied to the price and profitability of a digital asset. Cryptocurrency miners are exposed to several variables they cannot control, including token prices, network difficulty, electricity costs and changes in mining technology. A sharp decline in mining returns can therefore turn an expensive GPU fleet from a productive asset into a financial burden. Renting computing capacity to businesses offered a different economic logic. Instead of using GPUs to compete for cryptocurrency rewards, CoreWeave could sell access to the equipment to customers that already had valuable workloads waiting to run. Rendering studios, machine-learning researchers and other organisations were less concerned about cryptocurrency prices; they cared about whether computing resources were available, reliable and fast enough for their work. This shift did not remove commercial risk, but it allowed the company to seek demand across several industries rather than remain dependent on the economics of one cryptocurrency network.

CoreWeave also benefited from skills that a casual description of “crypto mining” can overlook. Running thousands of GPUs is an infrastructure problem as much as a financial one. The equipment consumes substantial electricity, produces considerable heat and must be monitored continuously. Hardware failures need to be identified quickly, workloads need to be distributed efficiently and facilities require sufficient power, cooling and network connectivity. The founders had accumulated practical experience with these problems before AI companies started requesting enormous clusters of accelerators. CoreWeave also developed an early relationship with NVIDIA, whose GPUs were becoming increasingly important not only for graphics but for machine learning. That relationship helped CoreWeave focus purchases on hardware that could support different types of accelerated computing rather than equipment with only one economic purpose. When demand for AI computing later increased dramatically, the company was not starting from zero. It already understood how to purchase, install and operate GPU fleets at a scale that many conventional software businesses had never needed to manage.

The transition also created a more useful identity for the business. CoreWeave was not trying to reproduce every service offered by Amazon Web Services, Microsoft Azure or Google Cloud. Instead, it concentrated heavily on demanding workloads where GPU availability and performance mattered. This specialisation became increasingly valuable because AI developers often need thousands of accelerators working together for extended periods. Supplying that type of capacity involves more than putting individual GPUs in a data centre: the servers, networking, storage, power systems and management software all have to work together consistently. CoreWeave’s earlier experience with mining and rendering gave it a relatively narrow but relevant set of capabilities. The company could therefore position itself around accelerated computing at a time when GPUs were moving from a specialist component to one of the most sought-after resources in technology. What had initially been an attempt to find additional uses for mining equipment eventually became a business whose principal product was access to the computing power itself.

The AI Boom Turned CoreWeave into a Major Compute Supplier

The decisive change came as machine learning moved from a specialist field into a major source of technology investment. The release and rapid adoption of generative-AI services from 2022 onwards dramatically increased demand for the hardware required to train and operate large models. Training a modern model can require thousands of high-end GPUs working together for weeks or months, while serving a successful model to millions of users creates a continuing requirement for inference — the computing performed each time the model generates an answer, image, prediction or other output. This created a shortage not merely of individual chips but of complete, usable computing environments. CoreWeave entered this period with an unusually relevant combination of GPU experience, an established NVIDIA relationship and a business already designed around renting accelerated computing capacity. Its financial growth illustrates how quickly that demand developed. Revenue rose from $229 million in 2023 to $1.915 billion in 2024, an increase of more than eightfold, before reaching $5.131 billion in 2025. For the first six months of 2026 alone, revenue was $4.653 billion, more than double the $2.194 billion reported for the same period of 2025.

The physical expansion behind those figures was equally rapid. At the end of 2023, CoreWeave operated 10 data centres with approximately 70 megawatts of active power. By the end of 2024, the company reported 32 data centres and roughly 360 megawatts. At the end of 2025, it had 43 data centres and more than 850 megawatts of active power, alongside approximately 3.1 gigawatts of contracted power intended for future deployment. Expansion continued during 2026. CoreWeave passed one gigawatt of active capacity during the first quarter and reported approximately 1.5 gigawatts at the end of the second quarter, with contracted power reaching about 3.7 gigawatts. These power figures help explain why AI infrastructure has become closely connected to energy and property development. Adding computing capacity requires suitable buildings, electrical connections, cooling equipment and network infrastructure long before a customer can begin using a GPU. CoreWeave’s growth therefore depends not only on obtaining NVIDIA hardware but on securing enough power and data-centre space to put that hardware into operation.

CoreWeave has also moved beyond simply renting servers. In May 2025 it completed its acquisition of Weights & Biases, a widely used developer business focused on tracking AI experiments, evaluating models and monitoring AI applications. The acquisition reflected a broader strategy: CoreWeave wanted a larger role in the process through which customers build, test and operate AI systems rather than supplying computing capacity alone. It later added further software and services covering areas such as inference, model evaluation, reinforcement learning and infrastructure monitoring. For customers, the appeal is straightforward. A research team does not want to spend most of its time diagnosing hardware failures or manually coordinating thousands of processors; it wants to train a model, evaluate the results and move the finished system into production. CoreWeave’s business increasingly attempts to reduce the amount of infrastructure work required between those stages. That strategy also gives the company another way to differentiate itself from much larger cloud competitors, whose product catalogues cover almost every category of business computing rather than concentrating so heavily on AI workloads.

How NVIDIA, Microsoft, OpenAI and Other Customers Changed the Scale

Large customers were crucial to CoreWeave’s acceleration. A master services agreement with Microsoft dates from February 2023, and Microsoft subsequently became by far the company’s largest source of revenue. In the second quarter of 2025, Microsoft accounted for approximately 71% of CoreWeave’s revenue, illustrating both the value and the risk of winning a customer capable of buying computing capacity at enormous scale. By 2026, the revenue mix had become more diversified, although concentration remained significant. CoreWeave’s second-quarter 2026 filing did not publicly identify customers in its revenue table, but its three customers above the 10% reporting threshold represented 36%, 26% and 10% of quarterly revenue respectively. That is a meaningful change from the previous year because the largest customer no longer represented a clear majority of revenue. The company nevertheless continues to acknowledge that a relatively small number of large customers account for a substantial portion of its business. Microsoft and OpenAI are specifically identified in CoreWeave’s 2026 risk disclosures as significant customers.

OpenAI became another major source of contracted demand in 2025. In March of that year, CoreWeave announced an agreement under which OpenAI could pay up to $11.9 billion for dedicated computing capacity through 2030, while OpenAI received $350 million of CoreWeave shares as part of the arrangement. Additional agreements followed. CoreWeave said in September 2025 that its total contract value with OpenAI had reached approximately $22.4 billion after a further commitment of up to $4 billion in May and another agreement worth up to $6.5 billion in September. Meta also became an important customer. A 2025 agreement initially involved a commitment of approximately $14.2 billion, and in April 2026 CoreWeave announced an expanded long-term agreement to provide Meta with approximately $21 billion of AI computing capacity through December 2032. During the same month, Anthropic signed a multi-year agreement to use CoreWeave resources for the development and deployment of the Claude family of models. These contracts demonstrate why CoreWeave is now considered part of the core infrastructure behind several of the world’s largest AI developers rather than a small specialist cloud supplier.

Demand has also broadened beyond the companies most closely associated with large language models. In April 2026, quantitative trading firm Jane Street committed approximately $6 billion to CoreWeave services across multiple facilities and separately invested $1 billion in CoreWeave shares. Other customers and expanded relationships reported during 2026 included businesses working in software, engineering, financial markets, robotics and scientific research. NVIDIA has played an especially unusual role in CoreWeave’s development because it is simultaneously a critical technology supplier, strategic collaborator and investor. In January 2026, NVIDIA invested $2 billion in CoreWeave Class A shares at $87.20 per share. The two companies also announced plans intended to support the development of more than five gigawatts of AI-focused computing facilities by 2030. CoreWeave’s close relationship with NVIDIA has helped it obtain and deploy new generations of AI hardware quickly, but it also illustrates an important feature of the modern AI supply chain: the companies designing chips, operating data centres, building models and financing infrastructure are increasingly connected through large commercial and investment relationships.

CoreWeave in 2026: Rapid Growth Comes with Significant Financial Risks

By mid-2026, CoreWeave’s financial scale was dramatically different from only three years earlier. The company completed its public listing in March 2025, pricing 37.5 million Class A shares at $40 each before trading began on Nasdaq under the CRWV ticker. In June 2026, it was selected for inclusion in the Nasdaq-100. Second-quarter revenue reached a record $2.575 billion, compared with $1.212 billion in the same quarter of 2025. Even more striking was the amount of future contracted business. CoreWeave reported approximately $104 billion of revenue backlog as of 30 June 2026. The company also stated that this figure did not include more than $25 billion of additional customer commitments secured in early in the third quarter. Backlog is not the same as recognised revenue: CoreWeave must still deliver the contracted capacity and meet the relevant service requirements. Nevertheless, the size of the commitments provides unusually long visibility into expected demand and explains why the company continues to invest aggressively in new equipment, power capacity and data centres.

That expansion is extremely expensive, and CoreWeave’s financial statements make the cost visible. As of 30 June 2026, the company reported $35.6 billion of total indebtedness. During the first six months of the year, it used approximately $14.9 billion of cash in investing activities, largely because of infrastructure expansion, while property and equipment on its balance sheet increased to approximately $46.7 billion. The company reported a second-quarter net loss of $626 million despite its rapid revenue growth, while net interest expense for the quarter was $640 million. These figures do not mean the underlying business is failing; they show that CoreWeave is financing a very large amount of infrastructure before receiving all of the revenue those assets are expected to generate. Management has frequently used long-term customer commitments to support debt financing for GPU purchases and new facilities. The approach can accelerate growth when contracted demand materialises as planned, but high debt also increases sensitivity to interest costs, construction delays, weaker AI spending or customers reducing future commitments.

Customer concentration remains another important risk, even though it has improved from the unusually high levels seen in 2024 and 2025. In the second quarter of 2026, CoreWeave’s three customers above the reporting threshold accounted for a combined 72% of revenue. The company also reported that 98% of quarterly revenue came from committed contracts. Long-term commitments provide greater revenue visibility than purely on-demand usage, yet they create a different form of exposure: CoreWeave may build or finance capacity specifically to satisfy large contracts, leaving it vulnerable if a major customer cannot meet its obligations or if deployment schedules change. There are also physical constraints. New data centres require sufficient electricity, suitable sites and large quantities of advanced hardware, while new generations of GPUs arrive quickly enough that older equipment can lose relative value. CoreWeave must therefore balance several goals at once — installing capacity rapidly, keeping existing infrastructure highly utilised, maintaining strong customer relationships and ensuring that the debt used to fund growth does not become a disproportionate burden.

Why CoreWeave Matters to the AI Infrastructure Market

CoreWeave matters because its rise illustrates a structural change in the computing industry. For many years, the largest cloud businesses succeeded by offering very broad collections of services to almost every type of organisation. The AI boom created room for a more specialised model. Large AI workloads require unusually dense concentrations of expensive accelerators, substantial power and infrastructure designed to keep those processors working efficiently. CoreWeave concentrated on that requirement at exactly the time when demand began to exceed available supply. Its crypto-mining history, which might once have appeared unrelated to enterprise computing, gave the company relevant experience in operating large GPU fleets. Its rendering work then demonstrated that those assets could support commercial customers outside cryptocurrency, while the growth of generative AI created a far larger addressable market. This sequence is important because it shows that the company’s competitive position was not created by one decision made after ChatGPT became popular. The underlying transition had been under way for several years before AI computing demand accelerated.

The next stage of CoreWeave’s development will depend on how successfully it converts infrastructure scale into a broader and more durable business. AI training remains important, but inference is becoming an increasingly large source of computing demand as companies place models into everyday products and services. CoreWeave has responded by adding tools for model monitoring, inference, evaluation and reinforcement learning, particularly through Weights & Biases and other software acquisitions. It is also extending its geographic footprint. In 2026, the company announced additional capacity in Sweden and plans for its first data-centre presence in the Asia-Pacific region through three facilities in Indonesia with a combined 360 megawatts of contracted IT power expected to come online in 2028. International expansion could put computing resources closer to customers and help organisations meet local data requirements, but it adds further construction, regulatory and financing complexity. CoreWeave therefore needs to demonstrate that the operating discipline developed in the United States can be repeated across a much larger international network.

The transformation from cryptocurrency mining to AI infrastructure is ultimately more substantial than a change of branding. CoreWeave took assets and skills developed for one volatile use of GPUs and redirected them towards a different market whose demand turned out to be vastly larger. The company survived the crypto downturn, entered rendering and machine learning, launched its cloud business in 2020, discontinued its former mining offerings, secured relationships with leading AI companies and expanded into a publicly traded business generating billions of dollars in quarterly revenue. By 2026, the combination of major contracts, approximately $104 billion of reported backlog and 1.5 gigawatts of active power provides strong evidence that CoreWeave has become an important supplier in the AI computing market. The same figures also show why its future cannot be judged on revenue growth alone. CoreWeave must convert its huge contractual commitments into delivered services while controlling debt, maintaining reliability and keeping expensive infrastructure relevant as AI hardware evolves. Its first transformation was from mining Ethereum to selling computing capacity; the next challenge is proving that this extraordinary period of expansion can develop into a sustainable long-term business.