Modal Adds Clustered Computing for HPC

Clustered computing turns Modal from a convenient serverless GPU layer into infrastructure for jobs that break on a single machine. Protein folding, weather simulation, and similar workloads need many GPUs running at once, fast links between nodes, and simple ways to launch and monitor long jobs. Modal now supports multi node clusters with RDMA, up to 64 H100 GPUs, and integration with its existing Functions, storage, and notebook workflows, which lets the same platform cover both everyday AI inference and heavier scientific compute.

• This matters because HPC buyers are not just buying raw GPUs. They need several machines reserved together, low latency networking between them, shared storage, and a clean way to move data and logs around. Modal already had the developer workflow piece, and clustered computing adds the missing hardware coordination layer.
  1 sacra 2 modal 4 modal
• The closest comparison is not Replicate style single endpoint inference, but systems built for full cluster jobs. Replicate emphasizes simple API calls for packaged models, while Modal now reaches into multi node training and scientific workloads that look more like research lab or advanced enterprise jobs than standard app inference.
  3 sacra 4 modal 5 sacra
• There is already proven demand for this kind of compute in science. Cerebras found early traction with national labs and enterprises working on protein folding, climate prediction, and molecular dynamics. That shows the adjacent market is real, even if winning it depends on packaging cluster grade performance in a much easier developer product.
  5 sacra 6 nvidia 7 nvidia

The next step is for Modal to become the default place where a team starts with one Python function, then grows into multi GPU training, simulation, and lab scale compute without changing platforms. If that handoff works, Modal can expand from AI startup infrastructure into a broader scientific and enterprise compute layer.