Compare capacity options
All CKS options require a configured sandbox runner and a policy that permits the workload. You manage the cluster’s capacity. Choosing CKS placement doesn’t provision Nodes. For the commercial terms of each plan, see About capacity plans. To run a sandbox on any of these CKS capacity options, follow Use your own compute.
Check access and workload requirements
Some requirements can exclude an option before you compare demand or cost:
The GPUs-per-sandbox limit applies to your organization in both serverless and CKS placement. A request above it fails with
CWSANDBOX_GPU_COUNT_EXCEEDED. To discuss a higher limit, contact support or your account team. See GPUs per sandbox.
GPU sandboxes require separate private-preview access for both serverless and CKS placement. For supported clients, hardware, and isolation requirements, see Run a GPU sandbox. For CKS runtime settings, see Configure a sandbox policy.
The credential determines which account owns the sandbox. For Forge SaaS and CoreWeave accounts, see CoreWeave Infrastructure and Forge billing. If you use a Dedicated Cloud instance key, confirm billing terms and effective limits with your account team. The Forge account defaults alone don’t establish your instance’s quota. See Use sandboxes with W&B Dedicated Cloud.
Before planning a burst or fallback path, check Choose and get credentials and Limits and quotas. CKS sandboxes don’t count toward the concurrent serverless sandbox quota.
Plan for steady demand and bursts
Separate the capacity you use consistently from occasional peaks. The following illustration shows a steady baseline with bursts above it. It represents a planning pattern, not measured usage, a price comparison, or automatic placement behavior.- Steady demand: Consider Reserved Instances for sustained use. If you already have suitable unused reserved capacity, evaluate using it before committing to more.
- Variable demand with a guaranteed peak: Consider Flex Reservations when holding peak capacity matters, even while some of it is idle.
- Occasional bursts: Consider serverless to avoid managing additional Nodes. On-Demand is another option when you need CKS placement and can tolerate capacity being unavailable.
- Work that can restart: Consider Spot when retries and flexible completion times are acceptable. Demand shape alone doesn’t establish whether interruptions are safe.
Understand scaling and interruption limits
Node Pool autoscaling doesn’t guarantee capacity for sandbox creation, and spillover doesn’t recover interrupted work.Node Pool autoscaling
CKS Cluster Autoscaler adds Nodes in response to unschedulable Pods. Sandbox placement can reject a request for insufficient capacity before creating a Pod. In that case, no pending Pod exists to trigger scale-up. A longer sandbox timeout doesn’t create that trigger. This limitation applies to sandboxes on your own CKS cluster, regardless of the Node Pool’s capacity plan. Don’t rely on incoming sandbox requests to scale a CKS Node Pool from zero. Provision capacity ahead of demand, keep sufficient ready capacity, or use serverless for bursts. A capacity reservation guarantees the reservation, not that Nodes are already ready to accept a sandbox request.Spot interruptions
Spot reclamation can interrupt a running sandbox. Use Spot for work your application can safely retry. Keep results or checkpoints needed for recovery outside the interrupted sandbox. For the Node termination sequence, see Instance preemption. A Spot Node Pool can replenish Nodes when capacity becomes available again. Replacing a Node doesn’t recover a running sandbox’s state.Spillover between placement modes
Python’s automatic spillover retries sandbox creation once in the other placement mode for eligible placement failures. It doesn’t migrate running sandboxes or recover work interrupted by Spot reclamation. Spillover requires a CoreWeave API access token for an organization with serverless access and a configured CKS runner. Each mode applies its own resource policy and limits. The TypeScript client supports both placement modes but doesn’t provide automatic spillover. Creating a sandbox from a template andrun_from_file() require strict placement without spillover.
For configuration, supported directions, and runner restrictions, see Spill over between modes.