CoreWeave adds Flex Reservations and Spot to fix GPU capacity tradeoffs

The gap between reserved and on-demand capacity

CoreWeave argues that the cloud's old promise of starting small, scaling without limits and paying only for use no longer holds once teams scale with GPUs. AI demand runs in waves, the company says: training and evaluation come in iterative cycles, while inference rises and falls through the day, across regions and around unpredictable product moments.

The problem, according to CoreWeave, is not the variability itself. It is that capacity cannot be counted on when demand rises unless a customer pays for it all the time, and "on-demand" does not always feel on-demand when it matters most. The company says this changes how teams innovate: they either over-protect early ideas with always-on commitments or keep projects small because scaling feels risky.

Customers on CoreWeave already use two modes. Reservations cover a steady baseline, and On-Demand covers incremental best-effort top-up. The gap, CoreWeave says, is variable and still-growing workloads where paying full 24×7 economics is wasteful but best-effort availability is not an acceptable risk.

Flex Reservations and Spot

Flex Reservations, launching in preview, provide guaranteed capacity for workloads that rise and fall. Customers choose a ceiling and CoreWeave guarantees access up to that level. Instead of full 24×7 economics, Flex uses a holding fee that maintains the guarantee, plus usage charges when instances are actually running. CoreWeave says it is setting a new standard by guaranteeing capacity without requiring customers to pre-schedule exact windows or overbuy full-time reservations.

Spot is now generally available as a lower-cost option for work that can tolerate interruption. According to CoreWeave, interruption is explicit and comes with several minutes of notice and clear signals, so teams can move work off a node before it is terminated. Spot is delivered as a node pool type in CoreWeave Kubernetes Service (CKS), with clear preemption signaling and operational guardrails meant to support checkpoint-and-recover workflows. The company points to experiments, backfills, batch work and overflow as fitting uses.

CoreWeave expects most teams to run a mix: reserve what is steady, use Flex when peaks must be protected, and use Spot or On-Demand when work can move or best-effort is acceptable. It also says CoreWeave ARENA supports the proof step by letting teams run real workloads end to end under real conditions, while Flexible Capacity Plans target the ramp and scale step that follows.

Inference options across the lifecycle

The post ties the plans closely to production inference, which CoreWeave describes as something operated day to day under changing traffic, tighter latency targets and evolving models. The company warns that abstracting away infrastructure can cost teams clarity into GPU classes, runtime behavior, time-to-serve expectations and the real drivers of cost and performance.

CoreWeave AI Inference is presented as covering the full lifecycle through three entry points:

• Serverless Inference in Weights & Biases, built on CoreWeave infrastructure, lets developers run open-source models on demand and pay in tokens without managing infrastructure. Teams can bring LoRA-style adapters to base models for light customization.
• Dedicated Inference (Preview) is a newly announced offering for running inference on chosen GPU classes and open runtimes, with infrastructure-aligned economics and operational visibility.
• Inference on CKS gives the most control, letting customers deploy directly on hardware and run additional services next to their models.

CoreWeave says a common production pattern is to reserve a baseline, protect peaks with Flex, and use Spot for tolerant background work such as batch scoring, evals, backfills or overflow.

What it means for AI teams

For teams with uneven GPU utilization, the plans offer a middle path between paying for always-on reservations and relying on best-effort scaling. Details such as pricing levels, holding fee amounts and availability dates for the preview offerings are not given in the announcement.

Cline, which describes itself as an open-source agentic coding platform for large global enterprises, is cited as a customer moving inference workloads to CoreWeave. Founder Saoud Rizwan said the company uses Serverless Inference in Weights & Biases "to support enterprise-ready environments for our most security-sensitive customers."