Lambda Labs

Provisions NVIDIA H100, A100, and A10G GPUs on-demand with per-second granularity billing, enabling users to spin up single or multi-GPU instances without long-term commitment. The platform abstracts away bare-metal provisioning complexity through a web dashboard and API, handling resource allocation, networking, and billing calculation automatically. Users can scale from single-GPU development instances to multi-node clusters for distributed training without manual infrastructure management.

Delivers a proprietary, pre-installed software stack (Lambda Stack) on GPU instances containing optimized ML libraries, CUDA drivers, and frameworks, eliminating the need for manual dependency installation and environment configuration. The stack is pre-baked into instance images, reducing time-to-training from hours (manual setup) to minutes. Specific contents of Lambda Stack are not documented, but the platform claims it includes 'pre-configured ML software' suitable for training and inference workloads.

Launches a Jupyter notebook server on a GPU instance with a single click, automatically configuring GPU access, kernel selection, and persistent storage mounting. Users access notebooks via web browser without SSH or CLI knowledge. Persistent storage is mounted to the notebook environment, enabling data and model checkpoints to survive instance termination. The implementation abstracts away Jupyter server configuration, SSL certificate management, and storage binding.

Provides persistent block storage volumes that survive instance termination, allowing users to store training data, model checkpoints, and logs independently of compute instance lifecycle. Storage is mounted to instances via a documented mount point, enabling seamless data access across multiple training runs. The implementation decouples storage from compute, enabling cost optimization (stop instances, keep data) and disaster recovery (reattach storage to new instance).

Provisions multi-GPU clusters (via '1-Click Clusters™') that abstract away distributed training setup, enabling users to scale from single-GPU to multi-node training without manual NCCL/Horovod configuration. The platform handles GPU-to-GPU networking, collective communication initialization, and cluster topology discovery. Users submit training scripts that automatically detect available GPUs and scale across the cluster. Implementation details (NCCL version, collective communication backend, topology discovery mechanism) are not documented.

Deploys trained models on GPU instances for real-time or batch inference, leveraging GPU acceleration for low-latency predictions. The platform enables users to serve models via HTTP endpoints (implementation details not documented) or batch inference jobs. GPU instances can be sized independently of training, enabling cost optimization (smaller GPUs for inference than training). Inference-specific features (batching, quantization, model serving frameworks) are not documented.

Provisions dedicated, single-tenant GPU clusters isolated from other customers, enabling compliance with data residency, security, and regulatory requirements (SOC 2 Type II claimed). The platform isolates compute, storage, and networking at the cluster level, preventing data leakage or cross-tenant interference. Specific isolation mechanisms (hypervisor-level, network segmentation, storage encryption) are not documented. Marketed for enterprises in regulated industries (healthcare, finance) requiring data sovereignty.

Sells pre-configured GPU workstations (desktop/tower systems with NVIDIA GPUs) for on-premises ML development and training. The platform bundles hardware with Lambda Stack software and support services, enabling teams to run ML workloads locally without cloud dependency. Workstation specifications, pricing, and support SLA are not documented. This is a secondary business line alongside cloud GPU rental.

Provides programmatic API for provisioning, monitoring, and managing GPU instances and clusters (implementation details not documented). The API enables infrastructure-as-code workflows, CI/CD integration, and automated scaling. Specific API endpoints, authentication mechanisms, rate limits, and response formats are not documented. The platform likely supports REST or gRPC, but this is not confirmed.