A serverless cloud platform built specifically for AI workloads—inference, training, batch processing, and secure code execution. Developers write ordinary Python functions, decorate them with GPU specs, and Modal handles containerization, scheduling, autoscaling, and scaling to zero. Its core primitives are Functions (serverless compute), Sandboxes (isolated environments for untrusted or agent-generated code, with 1B+ launched and $300M annualized revenue run rate), and **Servers** (low-latency LLM inference). Its technical edge is cold-start performance: a custom container runtime boots in ~1 second, and GPU memory snapshotting cuts model cold starts by ~10x (e.g., vLLM from 45s to ~5s). Pricing is per-second with no minimum commitment on Starter; H100 at $3.95/hr, A100 80GB at $2.50/hr. Network egress fees began October 1, 2026 (Starter includes 1 TiB/month, then $0.04/GiB). Status note: Modal remains independent (unlike Replicate, acquired by Cloudflare), raised $750M in October 2026 at a $15.75B valuation, and reports $1B+ annualized revenue run rate. SOC 2 Type II from Starter; HIPAA and audit logs on Enterprise.

Pros & Cons:

Pros: True per-second serverless billing with scale-to-zero means you only pay while code runs, which is dramatically cheaper than hourly GPU rentals for bursty or spiky workloads—a published example showed 50 average GPUs against a 75-GPU peak costing less than a fixed fleet despite a higher per-hour rate; GPU memory snapshotting is a genuine technical differentiator, cutting cold starts ~10x and making scale-to-zero viable for production LLM inference rather than a latency trap; the Python-native developer experience (write a function, decorate it, deploy) removes Dockerfile and Kubernetes overhead, and the same primitives span inference, training, batch jobs, and Sandboxes; Sandboxes provide isolated, disposable compute for agent-generated code, which is increasingly relevant as coding agents and autonomous workflows proliferate; Modal remains independent, unlike some competitors acquired by larger platforms, and its rapid revenue growth suggests strong product-market fit.

Cons: Hourly-equivalent rates are higher than raw GPU rental—H100 at $3.95/hr vs. dedicated providers at ~$2.89/hr—so the economics only favor Modal when duty cycles are low enough that scale-to-zero savings outweigh the rate premium; long-running, steady-state workloads are cheaper elsewhere—if you need a GPU pinned near 100% utilization 24/7, a dedicated pod or reserved instance beats Modal; network egress fees introduced October 2026 add a new cost dimension that must be monitored, especially for applications transferring large model weights or outputs; no pre-built model catalog—unlike Replicate, Modal gives you GPUs and expects you to wire up your own serving code (vLLM, model loader, custom logic), which is more flexible but requires more setup; HIPAA and audit logs are Enterprise-only, and the platform is Python-first, limiting appeal for teams on other stacks.

Who is it for?:Teams building production AI applications with spiky or unpredictable traffic who want scale-to-zero without cold-start penalties; developers who want arbitrary Python code on GPUs (custom inference pipelines, batch jobs, training) rather than a fixed model catalog; teams building agent runtimes that need isolated sandboxes for untrusted or generated code; organizations that need SOC 2 from a low entry point and can accept Enterprise for HIPAA/audit logs; teams already using AWS or GCP Marketplace billing who want committed spend to apply toward Modal usage. Not suited for steady-state 24/7 GPU workloads (dedicated rental is cheaper), teams that want a managed model catalog (Replicate is a better fit), or non-Python stacks.