Modal

A cloud platform for AI workloads — inference, training, notebooks, batch — whose relevant product here is Sandboxes: containerized environments where customers run agent and LLM-generated code. In this spoke Modal is not a vendor to compare on price but a published scheduler design: modal-1m-sandboxes is one of the few first-party accounts of what a container scheduler looks like when it is built for burst scale instead of cluster consistency.

What it argues

Modal’s platform thesis, as far as this spoke cares, is that central coordination is the ceiling. Its sandbox-creation path holds no data store, treats workers as the source of truth publishing state asynchronously to a Redis stream, and lets scheduling servers load-balance against a stale in-memory view, with worker rejection as the correction. That buys the claimed million-sandboxes-in-a-minute burst and costs the consistent cluster view that Kubernetes spends etcd writes to maintain — the trade is the whole design (container-scheduling).

Reading it honestly

Every figure comes from Modal about Modal (T3, modal-1m-sandboxes): the sub-half-second median start, the tens of thousands of creations per second, the “no practical ceiling.” No neutral benchmark of sandbox providers exists in the corpus, so the numbers stand unchecked. The mechanism is inspectable and the admissions are the credible part — the rtnl kernel-lock contention, the deliberately unsharded Redis stream, the Beta status.

Adjacency

Modal-as-a-provider (pricing, tiers, how it stacks against E2B or AWS’s Firecracker sandboxes) belongs to cloud-wiki, whose synthesis already names Modal in its open question on where agent-written code executes. The practice of operating a fleet at this scale is here. Same company, different altitude.

Related

modal-1m-sandboxes · compute-sandbox · container-scheduling · serverless · kubernetes · platform-ops