Pricing Comparison
CoreWeave charges $2.65/hr per H100 on-demand with a 1-month commit and $2.15/hr with a 12-month reservation. Lambda lists H100 at $2.49/hr on-demand with no commitment required, dropping to $1.99/hr at 3-month reserved terms. RunPod offers H100 community cloud at $1.49/hr on the spot-like tier and $2.29/hr for secure cloud with guaranteed availability.
At the H200 level, CoreWeave commands $3.40/hr on-demand reflecting their early access to large clusters. Lambda prices H200 at $3.19/hr on-demand. RunPod does not yet offer H200 at scale as of mid-2026, capping their secure cloud at H100. For B200, only CoreWeave and Lambda have public pricing: $5.80/hr and $5.49/hr respectively.
| GPU Type | CoreWeave (on-demand) | Lambda (on-demand) | RunPod (secure) |
|---|---|---|---|
| H100 80GB SXM | $2.65/hr | $2.49/hr | $2.29/hr |
| H100 80GB PCIe | $2.10/hr | $1.99/hr | $1.79/hr |
| H200 141GB SXM | $3.40/hr | $3.19/hr | N/A |
| B200 180GB SXM | $5.80/hr | $5.49/hr | N/A |
| A100 80GB SXM | $1.50/hr | $1.39/hr | $1.19/hr |
Interconnect & Networking
CoreWeave operates a custom-designed fabric with NVLink 4 at 900 GB/s within nodes and InfiniBand NDR400 at 400 Gbps between nodes. Their topology uses a 3-tier Clos network with 48-port Quantum-2 switches, delivering sub-2 microsecond latency between any two GPUs in a 1024-GPU cluster.
Lambda offers InfiniBand NDR200 at 200 Gbps between nodes in their larger clusters, with NVLink 4 within 8-GPU nodes. Their smaller clusters (under 64 GPUs) use 200 Gbps Ethernet with RDMA over Converged Ethernet, which adds approximately 3-5 microseconds of latency compared to InfiniBand.
RunPod uses 200 Gbps InfiniBand NDR200 exclusively, but only on their secure cloud tier. Community cloud pods are limited to 25 Gbps Ethernet, which makes multi-node training impractical for models requiring frequent all-reduce. For single-node inference jobs, this is not a limitation.
Availability & Lead Times
CoreWeave claims 300 ms to spin up an H100 pod through their Kubernetes-native provisioning. In practice, large reservations require 2-4 weeks lead time for clusters above 256 GPUs. They maintain the largest publicly accessible H200 fleet at roughly 18,000 GPUs as of Q2 2026.
Lambda typically provisions H100 within 24 hours for clusters up to 64 GPUs and within one week for up to 512 GPUs. Their H200 availability is tighter at approximately 6,000 GPUs total, with 2-3 week lead times. Lambda posts live inventory counters on their dashboard, a transparency feature others do not offer.
RunPod secure cloud spins up single-GPU instances in under 60 seconds and clusters up to 8 GPUs within 5 minutes. Larger allocations require written requests with 48-hour turnaround. Community cloud is fully automated with no lead time but no availability guarantees.
Support & SLAs
CoreWeave provides 24/7 Slack-based support with a 15-minute response SLA for critical incidents. Their enterprise tier includes a named solutions architect and monthly capacity reviews. Standard support tickets average 4-hour response time during business hours.
Lambda offers email and ticket-based support with a 2-hour SLA for critical issues on reserved instances. On-demand instances get best-effort support only. Lambda maintains a public status page and publishes post-mortems for all major outages, which is rare in the neocloud space.
RunPod operates a community Discord for tier-1 support and a ticket system for tier-2. Their SLA guarantees 99.9% uptime on secure cloud only, with 5% service credit for each hour below the threshold. Community cloud has no SLA.
Ecosystem & Tooling
CoreWeave provides a Kubernetes-native platform with their own Container Runtime Interface and integrated object storage based on Rook/Ceph. Their CLI tool handles GPU pod lifecycle, volume management, and VPC configuration. They also support RunPod-style serverless endpoints through their Inference Platform.
Lambda offers a more traditional cloud experience with a web dashboard, API, and Terraform provider. Their stack includes managed Slurm and Kubernetes through their Lambda Cloud API. Lambda also sells bare-metal servers for teams that want full control, a differentiator from CoreWeave's purely virtualized approach.
RunPod leans heavily into the developer experience with a simple API, Python SDK, and pre-built templates for popular models. Their serverless endpoints bill per second of compute, making them ideal for burst inference workloads. Training features are more limited, with no managed Slurm or advanced scheduling.
Which Provider Fits Your Workload
CoreWeave wins for teams running large-scale distributed training (256+ GPUs) with InfiniBand requirements. Their network fabric maturity and large H200 fleet make them the default choice for foundation model training. The premium over Lambda is justified by the interconnect quality and support SLA.
Lambda is the strongest middle-ground option for teams needing 8-128 GPU clusters with reserved pricing. Their transparent inventory, Terraform integration, and bare-metal options make them a strong fit for production AI teams that value predictability over peak performance.
RunPod is the right choice for inference workloads, fine-tuning jobs under 8 GPUs, and teams that want the lowest possible barrier to entry. The community cloud tier is the cheapest way to run single-GPU experiments, but multi-node training at scale is not viable on their current infrastructure.