Runpod

Runpod is an American cloud-computing company that provides GPU infrastructure for artificial-intelligence workloads. Its services include persistent GPU containers known as Pods, autoscaling Serverless workers, hosted model endpoints, and multi-node clusters.¹

Runpod operates as both a cloud GPU provider and an inference provider. Pods and clusters expose accelerator environments on which customers run their own software, while public endpoints provide access to models managed by Runpod. Serverless workers occupy an intermediate position because customers supply the containerized application and Runpod manages its workers and request queue.

History

Runpod was founded in 2021 by Zhen Lu and Pardeep Singh. The founders began with GPU systems assembled in a basement in New Jersey, initially renting the machines during a period of limited accelerator supply.²³ They later added capacity from outside operators and developed a common platform for provisioning it.

Lu serves as chief executive and Singh as chief technology officer. Runpod describes itself as a remote-first company, with a legal and contact address in Moorestown, New Jersey, and a presence in San Francisco.⁴⁵

On 24 June 2026, the company announced a US$100 million Series A financing led by Summit Partners. Runpod also said that more than one million developers had used the platform by that date.⁶

The company previously styled its name RunPod. It adopted Runpod, with a lowercase “p,” as its preferred spelling in February 2026.⁷

Pods

A Runpod Pod is a containerized environment with allocated GPU, CPU, memory, and storage. It can be created from a template or a custom Docker image and accessed through SSH, JupyterLab, Visual Studio Code, or exposed web ports. Pods support training, inference, rendering, and other compatible GPU applications.⁸

Pod capacity is divided between Secure Cloud and Community Cloud. Secure Cloud comprises data-center capacity operated under additional infrastructure and security requirements. Community Cloud capacity is supplied by approved third-party operators. Both use the Runpod control plane, but their locations, host arrangements, and available features can differ.⁸

Compute charges accrue while a Pod is running. Container volumes and network volumes have separate lifecycles, so stopping or terminating an instance can have different effects on stored data. Runpod does not charge a platform fee for ordinary Pod ingress or egress, although services at the other end of a transfer may do so.⁸

Serverless and endpoints

Runpod Serverless runs a customer-supplied container as one or more workers behind a managed endpoint. The platform receives requests and adjusts the number of workers in response to demand. Workers can be invoked through a job queue or used behind a load-balancing endpoint.⁹

A deployment can retain active workers or scale to zero. Scaling to zero reduces idle compute charges but can introduce a cold start while the container and model are loaded. The customer continues to maintain the code and container image, while Runpod manages worker provisioning and request distribution.

Runpod also operates Public Endpoints containing models deployed by the company and Instant Clusters for multi-node computing.¹ Public Endpoints require no customer container and are the most fully managed inference product in the company’s catalog. Instant Clusters provide coordinated accelerator capacity for distributed workloads.

Infrastructure

Runpod’s inventory varies by region, hardware type, and supply class. A deployment’s performance depends on the accelerator as well as host memory, CPU, storage, networking, and interconnect. Pods also leave operating responsibilities such as container security, exposed ports, model licenses, monitoring, and checkpointing with the customer.

Templates reduce the work required to create an environment, but may include software maintained by third parties. In Serverless deployments, image size and model-loading time affect cold-start latency. Keeping workers active avoids part of that delay at the cost of continuing compute usage.

References

Footnotes

1. Runpod overview, Runpod documentation. ↩ ↩²

2. A note to the developers who built Runpod with us, Runpod. ↩

3. Founder series: origin story, Runpod. ↩

4. About Runpod, Runpod. ↩

5. Terms of service, Runpod. ↩

6. Runpod raises $100M after reaching one million developers, Runpod, 24 June 2026. ↩

7. It’s Runpod, not RunPod, Runpod, 25 February 2026. ↩

8. Pods overview, Runpod documentation. ↩ ↩² ↩³

9. Serverless overview, Runpod documentation. ↩