Modal Clusters GA: multi-node GPU behind `@modal.clustered`

Modal’s blog (2026-10-01, Peyton Walters) made Modal Clusters generally available through a single decorator, @modal.clustered, and says they are available to all workspaces today (Modal blog).

The product is serverless multi-node GPU orchestration as a decorator on Modal functions—not a managed agent-loop platform and not a mega silicon announce.

The decorator

Primary sample pattern: decorate a GPU function with @modal.clustered(size=…, rdma=…), then read placement from modal.Cluster.from_context() (private IPs, container rank). The SDK documents rdma=False by default—with that setting, containers can still talk over Modal’s private IP network without RDMA-capable placement (Modal blog, Clusters guide).

RDMA is optional. Enable it with rdma=True. Do not read every Cluster run as an RDMA fabric job.

What stays in the Modal loop

Modal says Clusters integrate with existing primitives: write checkpoints to Volumes, load data through Cloud Bucket Mounts, and orchestrate jobs with Queues (Modal blog).

Networking and customer color (vendor-attributed)

Modal claims InfiniBand verbs at up to 6.4 Tbps, automatic PyTorch/NCCL setup when RDMA is on, and cluster acquisition “within seconds,” billed by the second. Treat bandwidth, acquisition-time, and “fastest / truly serverless” comparative lines as Modal’s vendor framing, not independent newsroom measurements (Modal blog).

The same post includes customer testimonials from Decagon, 1x, and Runway (fine-tunes, world-model pretrain, multi-node inference). Short attributed color only—not independent case studies.

Pricing (as stated—no invented rates)

Modal frames Clusters as billed by the second / pay for what you use, with cluster size bounded by your plan’s GPU limits and a “reach out” path for large jobs. The GA post does not publish dollar-per-GPU-hour rates; this write-up invents none (Modal blog).

Who should care

Teams already on Modal who need multi-node GPU training or inference without owning the fabric should start at the GA post and the multi-node Clusters guide. Lead with @modal.clustered and optional RDMA; keep Volumes / buckets / Queues in the same mental model.