← All organizations

Cloud infrastructure for AI

Modal

Modal lets developers configure cloud hardware and software dependencies in Python. With Modal Functions, a team can serve a model from an HTTPS endpoint that scales with traffic and returns to zero workers when idle. For batch work, it can split a job across thousands of GPUs. Modal manages the machines and bills for compute by the second.

Erik Bernhardsson and Akshat Bubna started building Modal in 2021 after seeing data teams spend hours configuring even an hourly batch job. They wanted developers to edit code locally and run it in the cloud within seconds. That goal led Modal to build its own container runtime, scheduler, image builder and distributed filesystem. The filesystem delivers container images, model weights and datasets wherever a function runs. It fetches files on demand, so a container can start before its entire image has downloaded.

Teams can run their own training code on anything from a single GPU to a multi-node cluster. Modal’s open-source training server, Spindle, lets researchers write the training loop while the server handles distributed training. Several experiments can share the same trainer GPUs. While one experiment waits for an agent to finish a task, the GPUs can run a training step for another. In Modal’s tests, sharing lowered the cost per run, though individual runs took longer.

In 2023, customers began running AI-generated code on Modal, prompting the company to build Modal Sandboxes, isolated environments where agents can execute untrusted code. Cognition uses Modal for reinforcement learning and for serving models in production. Snorkel serves the open-weight models it evaluates on Modal and runs millions of agent simulations a month in CPU and GPU sandboxes. It equips each environment for the task. An agent learning to train a model needs a GPU, for example. Without one, a failed run tells Snorkel little about the agent’s ability.

8 talks

Newest first

Loading more talks and speakers…

5 speakers at AIE

Affiliations reflect their AIE event listings and recorded appearances, not necessarily current employment.

Company sources · checked 2026-11-13