Modal

Modal is a Python-native serverless cloud platform that lets developers run AI/ML inference, training, fine-tuning, batch jobs, and sandboxed code on GPUs and CPUs with sub-second cold starts and instant autoscaling. Hardware and container images are defined directly in Python — no YAML, Dockerfiles, or DevOps required.

Category

Infrastructure & Cloud

Pricing

usage-based, per-second compute (separate GPU, CPU, memory, storage rates)

Target Market

AI startups, ML/AI engineering teams, Data teams, Generative-AI app builders (LLM, image, audio, video), Research labs and academics, Enterprises running production AI inference

Founded

2021

Headquarters

New York, NY, USA

Customers

Thousands of customers (no exact public count)

• ✓
  Serverless GPU access with instant autoscaling

  On-demand access to T4, A10G, L40S, A100, H100, H200, and B200 GPUs; scales from 0 to 1000+ GPUs in seconds with no capacity planning or commitments.

• ✓
  Sub-second cold starts

  A proprietary Rust container runtime and image builder boots containers and loads large models in under a second, even on in-demand GPU types.

• ✓
  Python-native infrastructure-as-code

  Define container images, hardware, secrets, and scaling in pure Python via decorators, with no YAML, Dockerfiles, or separate DevOps scripts.

• ✓
  Broad workload support

  LLM and multimodal inference, fine-tuning (SFT, LoRA, full), multi-node training, batch jobs, scheduled/cron jobs, web endpoints, and Sandboxes for untrusted code and agents.

• ✓
  Built-in storage and data primitives

  High-performance distributed Modal Volumes plus the ability to mount S3, GCS, and other cloud storage directly into functions, with integrated logging and observability.

• •
  AI music and media generation at scale

  Suno generates AI music and Substack transcribes podcasts at scale on Modal.

• •
  LLM inference and fine-tuning for production apps

  Ramp deploys open-source LLMs and Lovable runs tens of thousands of concurrent containers on Modal.

• •
  Sandboxed execution of AI-generated code

  Developers and agents run untrusted, AI-generated code in isolated sandboxes; over a billion sandboxes have been launched, including thousands of concurrent sandboxes for reinforcement learning.

Pay only for active compute time, billed per second with no charge for idle resources. Sample rates: B200 $0.001736/sec, H100 $0.001097/sec, A100-80GB $0.000694/sec, T4 $0.000164/sec; Volumes $0.09/GiB/month with 1 TiB free. All new accounts get $30/month in recurring free compute credits; startup grants up to $25K and academic grants up to $10K are available.