How we built a cloud GPU notebook that boots in seconds (modal.com)

Modal announced Modal Notebooks — a cloud-hosted, collaborative Jupyter environment that boots GPU-backed kernels and arbitrary custom images in seconds. Rather than keeping expensive dev machines “warm,” Modal runs kernels inside Modal Sandboxes (isolated, high-performance containers) and uses a modal-kernelshim to translate Jupyter’s ZeroMQ kernel protocol into HTTP over the control plane so cell execution, interrupts, and streaming outputs feel immediate in the browser. Notebooks run on the same large, autoscaled pool of CPUs/GPUs as Modal’s functions, support automatic pausing/resuming to avoid idle GPU costs, and are already powering other products (Lovable, Marimo) as a reusable primitive. The speedup comes from systems work: a lazy-loading, content-addressed FUSE filesystem (Rust) mounts container images by metadata and fetches files on demand through a tiered cache (RAM, local SSD, zonal cache, regional CDN, blob storage), avoiding lengthy image unpack phases. Persistent global storage is provided by VolumeFS, and scheduling balances instant placement across thousands of GPUs/CPUs. Real-time collaboration is implemented with Rushlight (OT over Redis Streams) and CodeMirror; large outputs are offloaded to S3 Express One Zone. Developer ergonomics include LSP integrations with Pyright, Ruff via WASM, and experimental AI-assisted edit prediction (Claude 4 and an internal Zeta-on-H100 path). The result is a low-latency, cost-efficient notebook platform that preserves interactive feedback loops crucial for modern ML development and makes GPU workflows more accessible and shareable.