Overview

Lightning AI offers a high-performance cloud at different levels of control, all the way from bare-metal to fully-managed Studios and clusters.

On-demand

By default, everything you run on Lightning AI runs on-demand.

When you start a workload (like a Studio or Job) without reserving machines in advance, it runs on-demand. You pay only for the time you use. Availability for on-demand depends on supply and demand - popular GPU types may be temporarily sold out.

Reserved

When workloads need guaranteed capacity or will run continuously, reserve instances in advance. You’ll get dedicated resources at reduced rates.

Contact our team to reserve instances ([email protected])

GPU marketplace

Lightning AI's GPU marketplace lets you run any type of workload (Studio, Job, Pipeline, Deployment, etc...) on any cloud.

Clusters

Lightning AI offers high-performance GPU clusters with SLURM or Kubernetes.

Clusters allow teams to migrate their stacks as-is without changing workflow or code changes. Our managed clusters allow teams to maximize GPU utilization with full RBAC, observability, logs and more.

Bring your own cloud

Lightning AI lets organizations connect their cloud VPCs. This allows data to never leave your organization and to consume your cloud credits. This is great for startups with AWS/GCP credits and enterprises with pre-existing cloud commits.

Key benefits

Bring your own platform

We find that mature ML teams have already built internal tooling stacks (i.e., ‘ML platforms’). Lightning AI lets you bring those platforms to our managed Kubernetes or SLURM clusters. We can help augment and simplify your workflows with ML-specialized runtimes - similar to pods in Kubernetes, but purpose-built for ML workloads.

Our ML specific runtimes (similar to K8s pods) are:

Studios - development environment pod
Jobs - generic script submission
Deployments - deploys models with autoscaling
MMT (multi-node training) - a specialized workload for training models on multiple-nodes
LitData - a specialized workload for doing map operations with data

When you bring your own platform, Lightning AI gives you the ability to:

Run jobs exactly as you do today - no migration required
Bring your full stack
Onboard teams without introducing new tools
No new workflow required
Add RBAC, observability, alerts, and cost controls instantly
Alerts and observability layers on top of existing SLURM/Kubernetes

You get the familiar interface and control - with enterprise-grade management layered on top.

Maximize GPU utilization

Lightning AI can help you get the most out of your GPUs, no matter which orchestration system you use. Our tools layer on top of Kubernetes or SLURM to surface where utilization is falling short, and provide controls to manage usage at the cluster, workload, or even individual user level.

We often see teams averaging 50–60% utilization. With Lightning AI, they typically reach 70–80%+ by mapping the right workload to the right runtime. For example, using on-demand resources for development instead of keeping GPUs idle, we typically help customers save 40% or more on compute costs. The result is higher throughput for the same budget, and infrastructure that scales with your needs instead of holding you back.

Networking

In large-scale training, the wrong network setup can silently cut your throughput in half and waste millions in GPU time. Lightning AI’s team learned this firsthand building PyTorch Lightning in 2019 and running thousands of GPUs at Facebook AI. By 2020, we were among the first to understand that cloud training required co-located compute and storage workloads with high-speed interconnects,* years before this became standard practice*, because we saw how critical it was for scaling. Today, that experience is baked into Lightning AI’s platform.

We design networking from the ground up to keep GPUs fully saturated, even when training trillion-parameter models.

Storage

Lightning AI created PyTorch Lightning in 2019 to train large-scale models at Facebook AI on thousands of GPUs. We’ve spent years tuning data pipelines where any slowdown in disk or network I/O can waste millions in GPU hours. This expertise is built into Lightning AI’s platform - so your foundation model training jobs never get bottlenecked by storage.

Was this page helpful?