Paperspace

Allocates NVIDIA GPU compute instances (H100 and other SKUs) on-demand with per-second granularity billing rather than hourly minimums. Instances are provisioned within seconds via API or web console, with configurable auto-shutdown timers (12 hours free tier, configurable paid) and no long-term commitments. Users can change instance types mid-session without data loss via persistent storage integration.

Provides pre-configured Jupyter notebook environments (called 'Gradient notebooks') running on GPU instances with built-in automatic versioning, tagging, and lifecycle management. Notebooks persist across sessions via integrated storage, support pre-configured ML templates for rapid onboarding, and include configurable auto-shutdown to prevent runaway costs. Versioning mechanism (Git-based or custom) is not detailed but enables reproducibility and rollback.

Provides real-time cost tracking and billing transparency with per-second granularity for compute and storage. Displays estimated costs before instance launch, actual costs after execution, and cost breakdowns by resource type (GPU, CPU, storage). Supports cost allocation across team members via Insights dashboard. Billing model emphasizes cost savings vs. hourly competitors (claimed 'up to 70% savings').

Captures and stores execution logs (stdout, stderr) from notebooks and training jobs with full execution history including timestamps, resource utilization, and cell-by-cell output. Logs are searchable and filterable by date, job ID, or keyword. Execution history enables debugging failed runs and comparing outputs across multiple job executions.

Abstracts multi-GPU and multi-node training via a job scheduling system that handles resource provisioning, dependency management, and lifecycle orchestration. Jobs support distributed training patterns (data parallelism, model parallelism) across multiple GPU instances with automatic resource cleanup on completion. Job definitions specify training scripts, hyperparameters, and resource requirements; the platform provisions matching instances and monitors execution.

Packages trained models as HTTP API endpoints with automatic scaling based on request volume. Deployment abstracts containerization, load balancing, and instance management — users specify a model artifact and framework (PyTorch, TensorFlow, etc.), and the platform provisions inference instances, exposes a REST API, and scales replicas based on latency/throughput thresholds. Supports custom inference code via container images.

Provides persistent block storage (5GB-unlimited depending on tier) attached to GPU instances, surviving instance termination and enabling data reuse across training/inference jobs. Storage is automatically versioned and tagged alongside notebook/job artifacts, supporting reproducibility. Overage storage billed at $0.29/GB. Storage can be mounted across multiple instances within a region for data sharing.

Enables multi-user team workspaces with role-based permissions (likely admin, member, viewer roles) controlling access to notebooks, jobs, and deployments. Provides 'Insights' dashboard for team utilization tracking, permission auditing, and cost visibility across team members. Separate team billing tiers (T0-T2 at $0-$12/user/month) support scaling from individual to enterprise teams.

Integrates with version control systems (Git) to trigger automated training jobs and deployments on code changes. Workflows feature (part of Gradient) allows defining pipelines that execute training on push, run tests, and deploy models to endpoints — abstracting CI/CD infrastructure and providing ML-specific orchestration (vs. generic GitHub Actions or Jenkins).

Provides a centralized repository for storing trained model artifacts, notebooks, and datasets with automatic versioning and tagging. Models can be tagged with metadata (framework, dataset, hyperparameters, performance metrics) and retrieved for deployment or further training. Repository appears to support model discovery and sharing within teams, though marketplace/community sharing features are not detailed.

Provides curated Jupyter notebook and training job templates for common ML tasks (e.g., image classification, NLP fine-tuning, generative models) with pre-installed dependencies, sample datasets, and starter code. Templates enable users to 'go from signup to training a model in seconds' by eliminating environment setup and boilerplate coding. Templates likely include framework-specific examples (PyTorch, TensorFlow) and popular datasets.

Enables deployment of Paperspace infrastructure on private cloud environments (Azure, AWS, GCP) or on-premise hardware (DGX systems, custom clusters). Provides Gradient software stack that can run on customer-managed infrastructure while maintaining integration with Paperspace control plane for unified management, billing, and monitoring across hybrid environments.