Modal

Serverless cloud compute for AI and data workloads, in pure Python

Modal is a New York-based serverless cloud compute platform that lets Python developers run GPU and CPU workloads in the cloud — model training, fine-tuning, inference, batch jobs, and data pipelines — by writing standard Python functions decorated with simple annotations, with zero infrastructure management required. The platform handles containerization, GPU allocation, autoscaling, and billing automatically; developers never interact with YAML, Kubernetes, or cloud IAM configuration. Modal raised an $87M Series B in September 2025 at a $1.1B unicorn valuation, co-founded by Erik Bernhardsson (former Spotify CTO) and Akshat Bubna. Its key technical differentiator is GPU memory snapshots (currently in alpha), which capture the full GPU state including model weights in VRAM to enable up to 10x faster cold starts — addressing the main performance limitation of serverless GPU platforms. Modal is developer-first and positioned for ML engineers and data scientists at AI-native companies, not for business teams seeking no-code tools.

Compliance

SOC 2

Visit Modal

Key Features

Python-native infrastructure-as-code: Compute requirements are defined via decorators directly in Python source files (@app.function(gpu="A100")) — keeping infrastructure definitions alongside application code rather than in separate YAML or Terraform configurations.
GPU and CPU workloads from one API: Fine-tune a model on H100s, run inference on A10Gs, and process data on CPU nodes from the same Python code and billing relationship — without switching between cloud services or compute abstractions.
Automatic scaling and sub-second cold starts: Functions scale from zero to thousands of concurrent instances automatically based on demand, with standard cold starts in seconds and GPU memory snapshots (alpha) reducing warm-up to milliseconds for pre-loaded models.
Web endpoints and scheduled jobs: Modal functions can be exposed as HTTPS endpoints for serving inference APIs, or scheduled with cron syntax for periodic training, evaluation, and batch processing jobs — covering the full ML pipeline lifecycle.
Persistent volumes and secrets management: Secure storage for model weights, datasets, and credentials, persisted across function runs without re-downloading on each invocation.
Usage-based billing with $30/month free credit: Pay only for active compute time with per-second GPU billing, a free $30 monthly credit, and no idle charges — eliminating the cost of reserved instances during development and low-usage periods.

Use Cases

For ML teams running fine-tuning experiments: An ML engineering team runs multiple concurrent fine-tuning jobs on H100s by decorating their training scripts with @app.function, scaling to 64 GPUs without provisioning a single instance.
For AI companies serving model inference APIs: A company that has trained a custom model uses Modal to deploy a scalable inference endpoint that scales to zero between requests and scales up automatically during traffic peaks.
For data pipelines requiring burst GPU compute: A team processing millions of documents for embeddings, OCR, or classification uses Modal to run batch jobs that burst to hundreds of GPU cores in parallel and cost nothing when idle.

Pricing MODELS

Usage-based

Pricing Summary

Usage-based with a free $30/month compute credit on the Starter plan (no credit card required). Team plan at $250/month plus usage, adding higher concurrency limits, longer log retention, and custom domains. Enterprise is custom-quoted. GPU pricing starts at approximately $1.10/hr for A10G through $4.29/hr for H100 SXM, charged per second of active compute. Regional and non-preemptible multipliers apply — buyers should review the pricing page carefully for effective production rates.