CoreWeave, Inc.

Abstraction. The process of reducing complexity to simplify the way a developer or user interacts with an underlying system, component, or particular layer of a technology stack. OpenAI’s ChatGPT API is an example of an abstraction layer as it allows organizations to integrate conversational AI capabilities into their applications without requiring developers to build and fine-tune the underlying model or manage its computing resources.

Artificial Intelligence (“AI”). A technology that enables computers and machines to simulate human learning, comprehension, problem solving, decision making, creativity, and autonomy.

AI Enterprises. Large companies whose business and product are not AI, but are being driven by AI. These include Fortune 500 companies and other large enterprises with business models that are increasingly AI-enabled through their own development efforts and/or that leverage AI directly to drive internal efficiency gains.

AI infrastructure. The combination of high-performance computing, networking, and storage components that collectively enables the development and deployment of AI models.

AI Natives. Organizations whose core competency and singular focus is AI. These companies serve as important layers of abstraction for the market and enable end customers across industries to unlock the power of AI, typically by building and delivering access to their own foundational models or by providing solutions to make AI adoption easier.

***Application Programming Interface (“API”).***A set of protocols and tools that enable different software systems to communicate and share data, often to abstract complex functionality into simpler commands.

***Bare Metal.***Computing nodes that have removed the virtualization layer to provide direct access to resources to run training, inference, and experimentation with maximum performance and low latency.

Burn-in. The testing process where hardware components, such as processors or servers, are operated continuously for an extended period under stress conditions to identify potential failures or defects before deployment.

***CoreWeave Kubernetes Service (“CKS”).***A fully managed, cloud-based container management deployment framework and orchestration solution with automated processes specialized for handling AI workloads.

***Central Processing Unit (“CPU”).***The primary processor of a computer that performs general-purpose tasks by executing the instructions of a computer program.

***Chips.***Integrated circuits or processors such as GPUs and CPUs that are designed to perform specific computational tasks or general-purpose processing.

Compute. The processing power required to execute software tasks and applications, such as running AI workloads, and typically provided by GPUs, CPUs, or other specialized chips.

Containerized. Applications or processes that are packaged in a container. A container is an abstract, stand-alone unit of software that has everything needed to execute a specific task, including the code, runtime, system tools, and system libraries. Containers may be scaled across computer nodes using orchestration tools and run independently from one another, isolating AI models and their dependencies from others on the same machine.

CoreWeave Cloud Platform. Our integrated and proprietary software solution that enables the provisioning of cloud AI infrastructure, the orchestration of AI workloads, and the monitoring of our hardware fleet that is deployed in active purpose-built data centers.

ii

***Cluster.***A collection of interconnected servers or nodes working together as a unified system on a single task simultaneously, distributing the workload across the system to improve computational efficiency.

Data centers***.*** References to our data centers refer to a data center, or portion of a data center, that we may lease or own, for which we are receiving active power.

***Data Processing Unit (“DPU”).***A specialized processor designed to handle data-intensive tasks such as network processing, storage management, and security. DPUs are becoming increasingly critical in high-performance computing settings where they are used to offload data-intensive tasks, which frees up GPU and CPU resources for more critical compute tasks.

***Data Loss Prevention (“DLP”).***Tools and policies designed to prevent sensitive data from being leaked, stolen, or accidentally shared outside of an organization.

Fleet LifeCycle Controller (“FLCC”). An observability solution in the Mission Control suite that automates node provisioning, testing, and monitoring to validate nodes and systems for Day 1 operations and beyond.

***Floating Point Operations per Second (“FLOPS”).***A metric commonly used by industry participants to measure the computer performance of computing, calculated as the number of floating point operations (i.e., addition, subtraction, multiplication, division) that can be calculated in a single second.

***Foundational models.***AI models designed for broad applicability, including large language models (“LLMs”), which are a subset of foundational models that are trained on massive data sets and designed to process and generate human-like language.

Generalized clouds. Cloud computing platforms that offer a range of services (including compute, storage, and networking) not tailored to a specific purpose. These platforms were designed to excel at general purpose, CPU-based use cases such ase-commerce, web hosting, and databases.

***Graphics Processing Unit (“GPU”).***A type of processor optimized for parallel data processing, widely used in graphics rendering and high-performance computing tasks. GPUs are built on underlying design architectures that define how the GPU will operate, including its processing cores, memory systems, data pathways, and features. NVIDIA A100s and H100s are examples of GPUs that are based on the Ampere and Hopper architectures, respectively.

***Hyperscaler.***A cloud provider or technology company that is capable of delivering computing infrastructure and services at massive scale, typically through large data centers and geographically distributed networks.

***Inference.***The process of using a trained AI model to generate predictions or outputs based on new input data.

***Instances.***Virtualized computing environments or virtual machines that run on cloud infrastructure and can be scaled up or down based on usage need. In high-performance cloud computing settings, instances can include a specific allocation of GPU resources, memory, and storage allowing users to rent hardware resources that map to a node, part of a node, or multiple nodes, depending on the configuration.

ISO 27001. An international standard for managing information security, which outlines best practices for protecting sensitive data through a structured framework of processes, policies, and controls.

***Kubernetes.***A container orchestration platform designed to automate the deployment, scaling, and management of containerized applications. Kubernetes is extensively used for inference workloads given its ability to scale and effectively handle dynamic, real time or batch requests across containers.

iii

Liquid cooling. A method of heat removal being applied in high-performance computing data centers that uses liquid, such as water or a coolant, to absorb and transfer heat away from critical hardware components. Liquid cooling is generally considered superior to air cooling in high-performance compute settings because liquid cooling allows for more efficient heat transfer due to greater cooling capacity that can support higher heat loads, and generally supports high compute densities due to less reliance on bulky airflow management infrastructure.

Mission Control. A proprietary suite of software and operational services designed to ensure optimal setup and functionality of infrastructure components throughout their entire lifecycle, preventing and rapidly remediating issues and enhancing performance. This includes our Observability solution, which provides access to a rich collection of node and system-level metrics that help prevent or quickly identify system faults and restore system-level performance.

MLPerf. An industry-standard benchmark suite to fairly evaluate AI and machine learning system performance across both hardware and software.

Model FLOPS utilization (“MFU”). A metric commonly used by industry participants that measures how efficiently a model is using the computational power of the hardware it is running on by dividing the observed throughput by the theoretical system maximum if it operates at peak FLOPs.

Networking technology. The hardware and protocols that enable communication and data exchange between devices, systems, or servers.

***Nimbus.***A control and data-plane software service running on DPUs inside Bare Metal instances, performing the typical role of a hypervisor in enabling security, flexibility, and performance.

Nodes. Individual computing devices or units within a distributed compute system, often part of a cluster or network. Nodes typically contain one or more processors along with memory, storage, and a high-speed network connection to communicate with other nodes.

***Node LifeCycle Controller (“NLCC”).***An observability solution in the Mission Control suite that proactively assesses ongoing node health and performance to ensure problematic nodes are replaced with healthy ones before they cause failures.

Non-Volatile Memory Express Solid-State Drives (“NVMe SSDs”). High-performance storage devices that are designed to deliver ultra-low latency and high throughput, making them ideal for high-performance computing workloads. They use the NVMe protocol to access flash memory via the PCI-Express interface.******

***Observability.***Tools and practices used to understand the state of a system by collecting and analyzing metrics, data logs, and health indicators to monitor performance and detect issues.

***Orchestration.***Refers to the automated coordination, management, and optimization of containers, services, or complex workloads across distributed computing environments to efficiently run AI applications.

PCI-Express (“PCIe”). A high-speed interface standard used to connect hardware components like GPUs, NVMe SSDs, and network cards directly to a system’s motherboard. It has a scalable architecture that allows for significant bandwidth, making it highly effective for data intensive tasks such as AI training.******

***Provisioning.***The act of allocating and configuring computing resources such as GPUs to meet the needs of an application. This entails activities such as defining the type and quantity of compute resources needed based on the workload, installing the required operating system and necessary software frameworks, setting up networking rules and access permissions, and configuring firewalls and authentication protocols.

iv

Rack. A standardized enclosure used in data centers to house servers, networking devices, and other hardware.

Remaining Performance Obligations (“RPO”). The aggregate total of our committed contract obligations that are yet to be delivered to our customers, inclusive of both billed and unbilled amounts.

Serving models. Providing access to deployed models in a production environment where they fulfill or “serve” real-time or batch processing of user requests.

Simple Linux Utility for Resource Management (“Slurm”). An open-source workload manager used in high-performance computing environments to allocate resources and schedule tasks across clusters. Slurm is particularly effective for model training, where large-scale parallelized computations need to be efficiently scheduled and distributed across compute nodes.

Service Organization Control 2 (“SOC2”). A cybersecurity compliance framework that ensures third-party service providers securely store and process client data.

***Slurm on Kubernetes (“SUNK”).***A proprietary offering that optimizes workload fungibility and enables greater resource flexibility by allowing customers to run Slurm jobs on Kubernetes and colocate jobs on the same cluster.

Take-or-pay. Contract arrangements where the customer commits to paying for a minimum specified amount of services or usage, regardless of whether they fully utilize those services.

***Technology stack.***The totality of the infrastructure hardware, software, and services that together support the operation of a system.

Tensorizer. A training and inference optimization solution that loads a model rapidly from storage directly into GPU memory from a variety of different endpoints.

Training. The iterative process of optimizing an AI model’s algorithm by exposing it to data and adjusting parameters to improve its accuracy or performance.

Validation. The process of confirming that a system or component, such as GPU node, meets predefined requirements and functions correctly under specified conditions.

***Virtualization.***The creation of virtual instances of computing resources (e.g., servers or storage) that operate independently of the underlying physical hardware.

***Virtual Private Cloud (“VPC”).***A private, isolated network hosted within a public cloud, allowing organizations to manage and secure their resources with the flexibility of cloud computing while maintaining control over access and traffic.

***Workloads.***The computational tasks or applications, such as training AI models or running inference, that consume resources like GPU, memory, and storage in a computing environment.

***Extended Detection and Response (“XDR”).***A security tool that combines and analyzes data from multiple sources to detect, investigate, and respond to cyber threats in an integrated manner.

v