“Imagine a caravan manufacturer and an RV park that buys only one kind of caravans.”
In March last year, the Financial Times opened its analysis of the CoreWeave case with this analogy. The American company, which describes itself as an AI hyperscaler, was about to go public. It is back in the news today thanks to announcements, within days of each other, of contracts with OpenAI and then Meta. Or rather, extensions of those contracts.
OpenAI had committed, in March 2025, to procure at least $11.9 billion of compute capacity from CoreWeave. Initially raised to $15.9 billion in May, the cap has just been lifted again to $22.4 billion. The contract runs through May 31, 2031.
Meta had signed with CoreWeave in December 2023. It too has enlarged the envelope allocated to that contract. It now stands at $14.2 billion.
The compute capacity in question takes, at the hardware level, one form and one form only: GPUs, exclusively supplied by NVIDIA. NVIDIA is also a customer of CoreWeave. The contract, signed in April 2023, had an initial value of $6.3 billion. It covered on-demand capacity. The two parties have now extended it, with NVIDIA committing to purchase, through April 2032, all unused capacity across CoreWeave’s entire infrastructure.
Vertical integration… up to buying its host
Around these GPUs, CoreWeave has pursued a vertical integration strategy. It has funded this in part through acquisitions, including, most recently, Weights & Biases and Core Scientific.
The acquisition of Weights & Biases brought a range of software bricks (inference, observability…) to CoreWeave’s AI stack, built around a homegrown Kubernetes distribution that incorporates the Slurm workload manager.
Core Scientific sits at the other end of the stack. It is the long-standing infrastructure provider for CoreWeave. Its main client for a time was Celsius Network. This crypto-mining platform, launched in 2017, collapsed five years later. Its founder ultimately pleaded guilty to fraud and market manipulation.
CoreWeave, a crypto ex
CoreWeave also dabbled in crypto, starting in 2017. The company was then called The Atlantic Crypto Corporation. It would not take on its current name until late 2019, after the Bitcoin/Crypto boom had peaked. Its GPUs—about half a thousand at the time—were subsequently repurposed for 3D rendering. A pivot would follow in November 2022, after the launch of ChatGPT: CoreWeave began to talk about machine learning and AI. By May 2024, it no longer described itself as “The GPU Cloud,” but as “The AI Hyperscaler.”
Before the crypto era, there was trading. CoreWeave is effectively the continuation of Hudson Ridge Asset Management, a firm that specialized in natural-gas futures. The same founder ran both ventures: Michael Intrator. Brian Venturo followed the move as well; he is now Chief Strategy Officer at CoreWeave.
In the shadows of CoreWeave, an obscure hedge fund
Until 2022, crypto activity accounted for most of CoreWeave’s revenue. That year, the company signed its first committed resource-supply contract. In parallel, it began to integrate the foundations of what now underpins its training and inference platform, including its tensorizer, designed to speed up loading models as serverless functions.
The homegrown Kubernetes stack followed in 2023, after a Series B round of $421 million… and a debt raise of $2.3 billion. In both funding moves, Magnetar Capital figures prominently. This hedge fund, formed two decades ago, became known for bets during the 2008 financial crisis. It invested in CoreWeave from the outset of the crypto-to-AI pivot. Since then, the ties have tightened. Last year Magnetar Capital became a customer of CoreWeave for certain companies in its MagAI Ventures portfolio. It provided a $230 million repayable advance to be offset against the price of resources actually used. In addition, CoreWeave invested in another magnetar-managed fund.
NVIDIA at the cap… and a lot of debt
Most of CoreWeave’s capital raises have taken the form of debt, with notable interest rates (estimated between 10% and 17%). More than one analyst flagged this ahead of the IPO. One such analyst, advising against investing, argued that deals like the OpenAI arrangement would merely service that debt, over a cycle too long relative to the GPU’s rate of obsolescence.
The same analyst went further. In his view, the only reason Microsoft uses CoreWeave’s services (as its main client; 62% of revenue in 2024) is that NVIDIA did not supply enough GPUs to Microsoft. He contends that Jensen Huang’s strategy is designed to keep pressure on the hyperscalers and to demonstrate the diversification of NVIDIA’s sales channels.
NVIDIA joined CoreWeave’s cap table in 2023. Its stake stands at roughly 5%. It is also, de facto, its second-largest customer (about $320 million in revenue booked through end-2024).
A “neo-cloud” in the hyperscalers’ arena
Despite revenue growth, CoreWeave—often described by the market as a “neo-cloud”—is still unprofitable. Its losses have been widening:
- 2022: $31 million loss on $16 million revenue
- 2023: $594 million loss on $229 million revenue
- 2024: $863 million loss on $1.92 billion revenue
- H1 2025: $605 million loss on $2.19 billion revenue
Behind this lies the expansion of its infrastructure. By end-2024, it boasted 250,000 GPUs spread across 32 data centers: 27 in the United States and 5 in Europe (2 in the United Kingdom, 2 in Spain, 1 in Sweden, all opened that year). Of the 2.2 GW at its disposal since acquiring Core Scientific, CoreWeave operates roughly 500 MW. Cohere, IBM and Mistral AI count among the companies that use its services to access the GB200 NVL72 servers.
Google also knocked on CoreWeave’s door to acquire capacity it intends to resell to OpenAI. CoreWeave acts as a buffer allowing Google to modulate investments in its own data centers. In the same spirit, Microsoft, after shelving several data-center projects, invited OpenAI to turn to CoreWeave.
At its IPO, CoreWeave did not raise the amount hoped for ($1.5 billion vs $2.7 billion). Yet the stock nearly tripled, valuing the company at just under $70 billion (the $100 billion mark was breached in June–July).
* The way CoreWeave orchestrated its transition from crypto to AI drew accusations of pump and dump. In broad terms, it sold its mining infrastructure in exchange for equity in a company that, two weeks later, would merge with another. The stock price subsequently surged… and CoreWeave would sell its own shares in the wake.
