Lambda Cloud

Provisions bare-metal or containerized NVIDIA H100 and A100 GPU clusters on-demand with sub-minute spin-up times through a cloud orchestration layer that manages hardware allocation, network configuration, and resource scheduling. Uses a capacity-pooling model where GPUs are pre-allocated across regional data centers and assigned to users via API or web dashboard, eliminating the multi-day wait times typical of reserved capacity models.

Provides pre-built container images and OS snapshots with PyTorch, TensorFlow, CUDA, cuDNN, and common training libraries (DeepSpeed, Hugging Face Transformers, vLLM) pre-installed and optimized for the target GPU. Users select a template at cluster creation time; the orchestration layer pulls the image and boots the cluster with all dependencies ready, eliminating 30-60 minutes of manual environment setup.

Provides NFS-mounted or block-storage volumes that persist across cluster termination and can be shared across multiple concurrent clusters. Storage is provisioned in the same region/availability zone as the cluster to minimize latency; the orchestration layer automatically mounts volumes at cluster boot via fstab or cloud-init, exposing them as standard Linux mount points accessible to training jobs.

Allocates clusters within isolated virtual private clouds (VPCs) with configurable security groups, allowing users to restrict inbound/outbound traffic and establish private connectivity between clusters. Clusters receive private IP addresses by default; public IPs are optional and can be disabled for security-sensitive workloads. VPC peering or VPN tunnels can be configured to connect Lambda clusters to on-premises infrastructure or other cloud providers.

Provides built-in support for distributed training across multiple GPUs and nodes via pre-configured NCCL (NVIDIA Collective Communications Library) settings, automatic rank assignment, and environment variable injection (MASTER_ADDR, MASTER_PORT, RANK, WORLD_SIZE). Users launch training scripts with a single command; the orchestration layer handles inter-node communication setup, GPU affinity, and collective operation optimization for the specific GPU topology.

Charges users per minute of GPU usage (not per hour or per node), with pricing differentiated by GPU type (H100 vs A100) and region. Billing starts when the cluster is in 'running' state and stops immediately upon termination; no minimum commitment or reservation fees. Costs are aggregated hourly and billed to the user's account; detailed usage reports are available via dashboard or API.

Provides REST API and web UI for creating, monitoring, and terminating clusters with full state tracking (provisioning, running, stopping, terminated). API supports programmatic cluster creation with configuration parameters (GPU type, count, region, image); dashboard provides real-time monitoring of GPU utilization, temperature, memory usage, and network I/O. Cluster state transitions are logged and queryable for auditing and automation.

Offers dedicated support for large-scale training runs (typically 16+ GPUs) with guaranteed uptime SLAs (e.g., 99.9%), priority access to GPU capacity during peak demand, and direct communication with Lambda engineers for troubleshooting. Support includes pre-flight cluster validation, performance tuning recommendations, and post-incident analysis for failed training runs.

Allows users to specify preferred regions and fallback regions at cluster creation time; the orchestration layer attempts to provision in the primary region and automatically falls back to secondary regions if capacity is unavailable. Users can query regional availability and pricing before cluster creation to make informed decisions about region selection.