CoreWeave to acquire data center builder Core Scientific for $9B in stock

Confirming recent rumors, cloud operator CoreWeave Inc. today announced plans to buy data center builder Core Scientific Inc. for $9 billion in stock.

Shares of both companies declined on the news.

Core Scientific launched in 2017 to build data centers optimized for cryptocurrency mining. It went public in 2022 through a merger with a special purpose acquisition company, or SPAC, and filed for bankruptcy the following year. Core Scientific exited Chapter 11 protection last January with a plan to repurpose its crypto mining facilities for artificial intelligence workloads.

Today, the company operates a data center network with 1.3 gigawatts’ worth of compute capacity. One gigawatt corresponds to the power usage of more than 700,000 homes. Core Scientific’s facilities support direct liquid cooling, which is used by large graphics card clusters to remove heat, and offer 99.999% of uptime. That translates into about five minutes of downtime per year.

CoreWeave inked an infrastructure hosting deal with Core Scientific last February, a month after the latter company exited bankruptcy. The initial contract covered 15 megawatts of data center capacity. Over the following year, the companies’ collaboration expanded to cover billions of dollars’ worth of infrastructure. 

CoreWeave relies on Core Scientific data centers to power its eponymous public cloud, which is optimized to run AI workloads. The platform offers access to more than 250,000 Nvidia graphics cards. Customers also have access to other hardware and software tools that ease tasks such as deploying AI models to production.

Earlier this month, CoreWeave became the first cloud provider to deploy a GB300 NVL72 system (pictured). It’s an AI appliance based on the Blackwell Ultra, Nvidia’s newest and most advanced graphics card. The chip can run some workloads 50% faster than the earlier Blackwell B200.

Each GB300 NVL72 combines 72 Blackwell Ultra chips with 36 of Nvidia’s Grace central processing units. There are also 18 BlueField-3 data processing units, or DPUs. The latter chips offload some auxiliary processing tasks from a system’s CPUs, which leaves more compute capacity for applications.

CoreWeave has developed custom software to manage the graphics cards in its cloud. An application called NLCC detects when one of the servers in an AI cluster malfunctions, removes it from the cluster and adds a new machine. This automated workflow is designed to reduce downtime for customers’ AI workloads.

About 40% of Core Scientific’s data center capacity, or 500 megawatts, is currently dedicated to cryptocurrency mining. CoreWeave plans to either sell that infrastructure or adapt it to AI workloads after the deal closes in the fourth quarter. Over the subsequent two years, the company hopes to realize $500 million in annual savings by boosting operational efficiency.

Photo: CoreWeave