"AI infrastructure that developers love."
The New York company that rebuilt the cloud from the metal up - so running a GPU is as simple as writing Python.
Modal's brand mark and tagline. New York, NY · founded 2021
Modal is a serverless platform where developers run compute-intensive AI and data workloads - model inference, training, fine-tuning, batch jobs, and sandboxed code - without ever touching a server, a Dockerfile, or an orchestration layer.
The pitch is deceptively simple. You write ordinary Python, declare the hardware you want with a decorator, and Modal handles the rest: it provisions the GPUs, boots the containers, scales the workload out and back to zero, and bills you by the second for exactly what you used. There are no idle-GPU charges, no egress fees, and no long-term commitments.
What makes that possible is unusual for a cloud startup. Rather than reselling capacity from a hyperscaler and wrapping it in a nicer dashboard, Modal built the hard parts itself. The container runtime, the file system, the scheduler, and the GPU memory snapshotting were all written from scratch in Rust. That vertical ownership is what lets Modal promise something most serverless platforms cannot: cold starts measured in fractions of a second, even for GPU workloads.
For the machine-learning teams who live with the friction of infrastructure - waiting on containers, wrestling with CUDA drivers, paying for GPUs that sit idle - the appeal is immediate. The platform spans the full lifecycle of an AI application, from a researcher's first fine-tuning experiment to a production endpoint serving thousands of requests a second.
The company describes its own product plainly on its homepage: infrastructure that developers love. It is a claim that sounds like marketing until you notice how many AI teams have quietly moved their workloads onto it.
Modal serves thousands of ML and AI engineering teams, from two-person startups to established enterprises. Its named customers read like a map of the current AI build-out.
The customer list spans the categories where GPU compute is the bottleneck: Suno in audio generation, Runway in video, Mistral in foundation models, Harvey AI and Cognition in agents, Lovable in AI app creation, and Ramp and Substack in production consumer and B2B software. The common thread is teams that need serious compute occasionally, unpredictably, and without a platform team to babysit it.
The problems Modal removes are the ones every ML engineer knows by heart. GPUs are expensive and scarce, so leaving them running to avoid slow starts wastes money - but scaling from cold is painfully slow on most platforms. Packaging code for the cloud means Dockerfiles and dependency hell. And running AI-generated code safely, now a daily need in the agent era, requires isolation most teams do not want to build. Modal's answer to each is to make the hard thing the default.
Python-native cloud development. Define logic and hardware in one codebase and attach a GPU with a single decorator.
Low-latency, autoscaling deployment for LLMs, audio, image, video, and multimodal models with web endpoints and observability.
Single- and multi-node GPU training, reinforcement learning, and parallel hyperparameter sweeps on B200, H100, and A100.
Secure, ephemeral, gVisor-isolated environments for running AI-generated code, coding agents, and RL rollouts.
Massively parallel batch processing and distributed job queues for embeddings, data jobs, and bursty workloads.
Collaborative, GPU-backed notebooks for experimentation and research without local hardware limits.
A custom Rust runtime and GPU memory snapshotting attack serverless GPU's oldest problem head-on, so functions boot in seconds and scale to zero without penalty.
Where rivals rely on REST configs or YAML files, Modal integrates through the language ML engineers already write. Zero config files to ship a GPU function.
No egress fees, no storage surprises, no idle-GPU bills. Billing tracks consumption to the second - a sharp contrast to hyperscaler invoices.
Container runtime, file system, and scheduler are all in-house. Modal can promise things a company reselling AWS simply cannot.
The competitive field is crowded - Baseten, Replicate, RunPod, Together AI, Cerebrium, Beam, and Fal on one side, and hyperscalers like AWS SageMaker and Google Cloud Run adding scale-to-zero GPU features on the other. Modal's position is the one carved out by depth rather than breadth: it competes on developer experience and cold-start performance rather than on being the cheapest raw GPU-hour.
Modal was founded in 2021 by Erik Bernhardsson and Akshat Bubna - both gold medalists at the International Olympiad in Informatics.
Former ML lead at Spotify, where he shaped early music recommendation systems, and former CTO of Better.com. Creator of Annoy, the open-source nearest-neighbor library. IOI gold, Sweden, 2003. He built Modal as the tool he wished he'd had while fighting infrastructure to ship ML.
Formerly an engineer at Scale AI. IOI gold, India, 2014. He leads the systems work behind Modal's from-scratch Rust runtime - the container engine, file system, and scheduler that make sub-second GPU cold starts possible.
The company is headquartered at 222 Broadway in Manhattan, with additional offices in San Francisco and Stockholm, and a team of roughly 120. The culture is engineering-led and infrastructure-obsessed, defined by a willingness to build unglamorous plumbing rather than assemble someone else's.
Modal runs a consumption-based SaaS model - billing by the second for compute, with monthly free credits and premium options for guaranteed execution and specific regions.
| Round | Amount | Date | Lead Investors |
|---|---|---|---|
| Seed | $7M | 2022 | Amplify Partners |
| Series A | $16M | Oct 2023 | Redpoint Ventures |
| Series B | $87M | Sept 2025 | Lux Capital ($1.1B valuation) |
| Series C | $355M | May 2026 | General Catalyst, Redpoint, Menlo, Bain, Accel |
The trajectory is what caught investors' attention: annualized revenue climbed from roughly $119M at the end of 2025 to about $300M by April 2026. The $355M Series C valued the company at $4.65 billion - a signal that the market believes AI-native infrastructure is a durable layer, not a passing bet.
Bernhardsson and Bubna start the company in New York to make GPU workloads as easy as local Python.
Amplify Partners leads as the team builds its Rust-based serverless runtime.
Redpoint Ventures leads as Modal opens up inference, training, and batch workloads.
Secure, ephemeral environments arrive for running AI-generated code and coding agents.
Lux Capital leads at a $1.1B valuation; GPU Notebooks and multi-node clusters ship.
General Catalyst and Redpoint lead as annualized revenue reaches roughly $300M.
Both founders won IOI gold - Sweden 2003 and India 2014.
The container runtime was written from scratch to kill cold starts, not patched onto someone else's cloud.
A single function can scale from zero to 1,000+ GPUs and back, billed only for seconds used.
Modal is a serverless cloud platform for running compute-intensive AI, data, and GPU workloads - inference, training, fine-tuning, batch jobs, and sandboxed code - without managing servers. Developers write Python, attach GPUs with a decorator, and the platform autoscales and bills by the second.
Modal (Modal Labs) was founded in 2021 by Erik Bernhardsson (CEO, formerly Spotify and Better.com) and Akshat Bubna (CTO). Both are International Olympiad in Informatics gold medalists.
Modal has raised roughly $465M total, including a $355M Series C in May 2026 at a $4.65B post-money valuation, led by General Catalyst and Redpoint, following earlier rounds from Lux Capital, Redpoint, and Amplify Partners.
Modal built its entire stack - container runtime, file system, scheduler, and GPU memory snapshotting - from scratch in Rust, enabling sub-second cold starts and true scale-to-zero serverless GPUs. It integrates through pure Python rather than YAML or REST config, and charges per second with no egress fees.
Thousands of ML and AI teams, from startups to enterprises. Named customers include Suno, Mistral, Ramp, Substack, Harvey AI, Cognition, Lovable, Quora, Runway, and Physical Intelligence.