Modal

Modal is a serverless GPU cloud for Python teams: decorate a function and it runs in the cloud with sub-second cold starts, per-second billing and autoscaling from zero to thousands of GPUs. It removes the Docker, Kubernetes and capacity-planning work that normally stands between an ML engineer and production inference, training or sandboxed agent execution.

Modal is a Python-native serverless compute platform: developers add decorators to ordinary Python functions and Modal runs them in the cloud on GPUs, replacing the Dockerfile, Kubernetes manifest and cluster-management layer that normally sits between an ML engineer and hardware. It was founded in January 2021 by Erik Bernhardsson, previously ML lead at Spotify and CTO of Better.com, with co-founder Akshat Bubna, and is headquartered in New York with offices in San Francisco and Stockholm. The company wrote its own container runtime in Rust and virtualises every job with gVisor, Google's user-space kernel sandbox, which is what underpins its claim of sub-second cold starts on pay-per-second billing with no reserved capacity. A single SDK covers four workload shapes: online inference with sub-10ms platform overhead plus token streaming, WebRTC and WebSocket transports; batch and asynchronous inference for embeddings, evaluations and dataset generation; training that spans LoRA and full fine-tunes, reinforcement learning and multi-node runs of up to 128 B200s over 3200 Gbps InfiniBand; and Sandboxes, isolated environments for executing untrusted or agent-generated code with GPUs attachable on demand. Autoscaling runs from zero to more than a thousand GPUs, routed across multiple clouds and regions, with integrated logging across every function, sandbox and container. Named production users include Suno, Lovable, Quora, Substack, Decagon, Reducto, Runway, Chai Discovery and Physical Intelligence. Modal completed a SOC 2 Type 2 audit and offers HIPAA BAAs on Enterprise plans. Independent analysis put its annualised revenue near $300 million by April 2026, and its own site reports over $466 million raised; it competes directly with Baseten, Replicate, RunPod, Beam and the hyperscalers' own scale-to-zero services.

Category

Infrastructure & Cloud

Pricing

Usage-based, Subscription, Freemium, Contact for pricing

Target Market

CTOs, VP Engineering, ML Engineers, Platform Engineers, Enterprise Developers

Deployment

Cloud-only, Multi-cloud, API-based

Founded

2021

Headquarters

New York, United States

Team Size

51-200

• ✓
  Python decorator SDK

  Turns an ordinary Python function into a cloud workload with one decorator, so there is no Dockerfile, YAML or cluster to maintain.

• ✓
  Sub-second cold starts on a Rust runtime

  A custom Rust container stack starts containers in under a second, which is what makes scale-to-zero viable for user-facing traffic rather than batch only.

• ✓
  gVisor-sandboxed execution

  Every job is containerised and virtualised with gVisor, giving isolation strong enough to run untrusted or model-generated code safely.

• ✓
  Multi-node distributed training

  Gang-schedules up to 128 B200 GPUs over 3200 Gbps InfiniBand from a single line of code, covering fine-tuning, RL and hyperparameter sweeps.

• ✓
  Sandboxes for agents

  Programmatically spawned isolated environments for coding agents, background agents and RL rollouts, with GPUs attachable on demand.

• ✓
  Per-second billing with autoscaling

  Meters compute by the second and scales from zero to over a thousand GPUs, so idle reserved capacity is never paid for.

• ✓
  Integrated observability

  Logging and full visibility into every function, sandbox and container, so failures are debuggable without bolting on a separate stack.

• •
  Autoscaling LLM inference

  Serve open-weight or fine-tuned models behind an autoscaling endpoint with sub-10ms platform overhead and token streaming.

• •
  Agent code sandboxing

  Run untrusted, model-generated code from coding agents inside gVisor-isolated sandboxes rather than on shared application infrastructure.

• •
  Fine-tuning and reinforcement learning

  Launch LoRA runs, full fine-tunes or multi-node RL jobs without provisioning a persistent GPU cluster or scheduler.

• •
  Batch embeddings and evaluations

  Fan out embedding generation, model evaluations and synthetic dataset creation across hundreds of containers, paying only for seconds actually used.

• •
  Generative media products

  Back image, video and audio generation features where traffic is spiky and idle GPU capacity would otherwise dominate unit cost.

List pricing is fully published, which is unusual in this category. Compute is metered per second on top of a plan fee — Starter $0, Team $250/month, Enterprise custom — with GPUs ranging from $0.000164/sec for a T4 to $0.001097/sec for an H100 SXM5 and $0.001972/sec for a B300, CPU at $0.0000131 per core-second (0.125 core minimum) and memory at $0.00000222 per GiB-second. Storage is $0.09/GiB/month after a 1 TiB free allocation. Starter includes $30/month of free credits and Team $100/month; startup credits and academic grants up to $10,000 are applied for separately. Per-second granularity favours spiky traffic, but an independent 2025 comparison priced Modal's H100 at roughly $4.76/hour against RunPod's $4.47/hour and noted Modal bundles CPU and memory charges alongside GPU time where competitors separate them — so sustained high-utilisation workloads pay a premium. HIPAA BAAs and embedded ML engineering are Enterprise-only.

This page was written from 6 sources, 3 on domains other than modal.com.

1. 1.modal.com — pricingvendor
2. 2.modal.com — securityvendor
3. 3.modal.com — companyvendor
4. 4.sacra.com — modal labs
5. 5.introl.com — serverless gpu platforms runpod modal beam comparison guide
6. 6.hn.algolia.com — search