CoreWeave adds Spot and Flex Reservations to its Capacity Plans

Flex Reservations: guaranteed capacity with a lower holding fee

CoreWeave argues that traditional reservations create a problem of overprovisioning. When GPUs are reserved for 24/7 access, the customer pays for them whether or not they are fully used. The company cites a Harness press release from February 2025, which estimated that 21% of enterprise cloud infrastructure spend is wasted on underutilized resources.

Flex Reservations are designed to separate guaranteed access from full 24/7 run-rate pricing. According to CoreWeave, they provide:

• Guaranteed access up to a Flex Reservations ceiling
• A lower holding fee to keep the capacity reserved, whether idle or running
• A complementary usage rate that applies only when GPU nodes are in use

In practice, a customer commits to a defined peak capacity, for example 200 GPUs. That capacity is reserved up to the ceiling throughout the term, as with a traditional Reservation. When the customer scales down below the peak, the usage rate stops on idle capacity and only the holding fee remains. CoreWeave says that deployment of workloads does not change, because the difference lies in the commercial structure underneath.

Spot: cheaper GPUs for interruptible work

Spot instances give lower-cost access to GPUs with no long-term commitment. They are meant for workloads that can tolerate interruption, such as fault-tolerant, experimental, batch-based or opportunistic jobs.

Like other Spot-style offerings, CoreWeave Spot instances can be preempted. CoreWeave says it surfaces preemption as a clear signal with advance notice, so workloads have time to checkpoint and shift work before a node is terminated. The company lists explicit termination signaling and a defined preemption notice window. It also claims that other clouds typically offer shorter or variable notice periods, depending on product tier and instance type.

Teams that want to try Spot first can use CoreWeave ARENA, which the company describes as a controlled environment for running end-to-end workloads under production-like conditions. For production inference, CoreWeave points to Weights and Biases Inference, built on its infrastructure, which includes explicit GPU type selection.

Designing for interruption and mixing capacity types

CoreWeave says Spot works best for workloads that are checkpointable, retry-safe, distributed or fault-tolerant, and non-urgent or batch-based, such as backfills, experiments and evaluation runs. It advises treating interruption as a design constraint. Its recommendations include:

• Checkpoint frequently enough to limit lost progress
• Separate control and worker nodes so control logic survives instance termination
• Use orchestration frameworks that automatically reschedule terminated nodes
• Blend capacity types, with critical components on guaranteed capacity and elastic components on Spot

Many teams, the company says, combine Flex or Reservations for baseline and peak guarantees, while moving less critical, fault-tolerant work to On-Demand or Spot to reduce total cost.

Which model fits which workload

CoreWeave offers a simple guide to the four models:

• Spot: interruptible work where savings matter and the system can tolerate preemption
• On-Demand: unpredictable work that can wait for available capacity but cannot handle interruptions
• Flex Reservations: guaranteed access for peaks without paying full 24/7 run rates when not at peak
• Reservations: steady baseline workloads that need predictable, always-on capacity

For customers, the change is a wider set of options for matching cost and certainty to how their workloads behave. Flex Reservations are still in preview, and the company directs interested customers to contact its team. The source gives no pricing figures, so the actual savings from either model are not specified.