2027 planning guide — evidence checked 20 November 2026

Lambda vs CoreWeave: which GPU cloud fits inference?

Both sell NVIDIA GPUs by the hour. Lambda publishes lower on-demand rates for the common H100 and B200 shapes; CoreWeave sells the cloud around the GPU, with managed Kubernetes, free egress and reserved capacity. The right answer depends on whether you are buying a server or a platform.

Short answer. For a self-serve H100 or B200 deployment with no platform team, Lambda's published on-demand rates are lower: $4.59 per H100 GPU-hour and $7.69 per B200 GPU-hour. For managed Kubernetes, free egress, spot and reserved capacity, CoreWeave is the stronger fit even at a higher on-demand rate. At 8×H100 for a month, the list-price gap is about $9,140 before egress and discounts.

This is a 2027 planning guide built from both vendors' public price pages on 20 November 2026. It is not a performance benchmark, and it does not assume either vendor has capacity in your region on the day you need it.

How we evaluated this

We weighted five criteria for a team running inference on rented GPUs rather than a managed model API.

  • Cost per GPU-hour, 30%. On-demand list rates for H100, H200, B200 and A100, converted to a common per-GPU basis.
  • Capacity access, 25%. Whether self-serve, spot, reserved or cluster capacity is published and how it is sold.
  • Platform services, 20%. Kubernetes, networking, storage and egress.
  • Operational burden, 15%. How much of the control plane you must build and run.
  • Contract clarity, 10%. Whether an SLA, credit schedule and exclusions are published.

We read the Lambda and CoreWeave pricing pages on 20 November 2026. We did not rent GPUs, run a benchmark, test interconnect throughput or verify capacity. Per-GPU figures are conversions from eight-GPU instance prices and may round.

On-demand, spot and reserved rates

Published GPU rates, USD per GPU-hour, checked 20 November 2026
GPULambda on-demandCoreWeave on-demandCoreWeave spot
H100 80 GB$4.59$6.16$2.46
H200 141 GBNot listed$6.31$2.62
B200 180 GB$7.69$8.60$4.26
A100 80 GB$3.19$2.70$1.21
L40S 48 GBNot listed$2.25$0.99
Reserved discountCluster pricing by termUp to 60% off on-demandNot applicable
Internet egressNot listed on the pricing pageFreeFree
KubernetesInstance-centric; confirm control planeManaged KubernetesManaged Kubernetes
GPU uptime SLANot publishedNot publishedNot published

CoreWeave prices are per eight-GPU instance on its public page: HGX H100 at $49.24 an hour, HGX H200 at $50.44, HGX B200 at $68.80. We divided by eight for the table. Lambda lists per-GPU rates directly, with 208 vCPU and 22 TiB of local SSD on an eight-GPU node.

The rate card is not the platform

CoreWeave's differentiator is the cloud around the GPU. It sells managed Kubernetes, VPC and NAT at no listed charge, free internal and internet transfer, and direct connect at $1,250 a month for 10G, $12,500 for 100G and $50,000 for 400G. Object storage is $0.06 per GB-month for hot, $0.03 for warm and $0.015 for cold; distributed file storage is $0.07. A public IP is $4 a month.

Lambda's pricing page is instance-centric: GPU virtual machines, 1-Click Clusters and Superclusters. It does not describe a managed Kubernetes control plane on that page. If you run Kubernetes, confirm whether you build the control plane, use a partner or buy orchestration with a cluster. That work has a cost even when it does not appear on the GPU invoice.

Egress is the other hidden line. CoreWeave lists internet egress and VPC/NAT as free. Lambda's pricing page does not list an egress rate, so ask for it in the quote. A model-serving workload that pulls weights once and serves many tokens may not care; a distributed training job that moves checkpoints will.

Cost math at eight H100s

Assume eight H100 GPUs running 730 hours a month, no idle time, no storage, no support plan and no tax. That is 5,840 GPU-hours.

8 x 730 = 5,840 GPU-hours

Lambda on-demand:      5,840 x $4.59  = $26,805.60
CoreWeave on-demand:   5,840 x $6.155 = $35,945.20
CoreWeave spot:        5,840 x $2.464 = $14,388.30
CoreWeave reserved:    about 60% off  = about $14,378

Monthly on-demand gap: $9,139.60
Annual on-demand gap:  about $109,675

At list price, Lambda is cheaper on-demand. CoreWeave spot and reserved can beat it, but spot is not a fit for a latency-critical endpoint without fallback, and reserved requires a commitment. The comparison changes again if you value managed Kubernetes and free egress at more than $9,140 a month, which is plausible for a small platform team.

Capacity is sold differently

Lambda sells self-serve instances first-come and 1-Click Clusters from 16 to 2,000+ B200 or H100 GPUs for two weeks to a year. Cluster pricing differs from instance pricing: 16 B200 GPUs list at $9.86 per GPU-hour, 64 at $9.36 and 256+ at $8.87, including interconnect and networking. A self-serve eight-GPU B200 instance is $7.69, so do not compare a cluster rate with an instance rate.

CoreWeave sells on-demand, spot and reserved capacity, with up to 60% off for commits. It lists H100, H200, B200, A100, L40S and RTX PRO 6000; B300 on-demand and GB300 are contact-sales, though B300 spot is listed. Availability varies by region and GPU generation. A published price is not a capacity guarantee: ask for the region, GPU model, interconnect and start date in writing.

When Lambda is the better choice

Lambda wins when you want the lowest published on-demand H100 or B200 rate, simple GPU virtual machines and short-term clusters without a platform abstraction. It also wins for a small team that needs a few GPUs this week and does not want a sales process. If your orchestration already runs on your own Kubernetes, CoreWeave's managed control plane is less valuable.

When CoreWeave is the better choice

CoreWeave wins when you want managed Kubernetes, free egress, reserved and spot capacity, and a cloud designed for multi-node AI infrastructure. It also wins when private networking, direct connect and a committed-capacity path matter more than the last dollar per GPU-hour. For a platform team building an internal inference service, that can be the cheaper total cost even at a higher rate.

When neither is the right shape

If you do not need custom weights, a managed API removes the GPU operations entirely. Plugsky Builder is $60 a month with no monthly token cap and no per-token overage; Scale is $120. Fair-use concurrency and RPM limits apply. That is not comparable to a dedicated GPU, but it is often the correct comparison for a team that is about to rent eight H100s to serve a model it could call through an API. The model catalogue shows what is available, and the fair-use definition explains the limits.

What we did not verify

We did not verify GPU availability, interconnect performance, storage throughput, real egress bills or benchmark tokens per second. We did not read either contract. Neither vendor publishes a numeric GPU uptime SLA on its public pricing page, so we record the SLA as not published; ask for the uptime commitment, measurement window, credit schedule and exclusions. We did not verify taxes, regional pricing or enterprise discounts.

For Plugsky, we did not verify a SOC 2 certification. Request the current SOC 2 report or attestation under NDA and verify the period. Managed API pricing is flat, but it is not a GPU rental and should not be modelled as one.

Frequently asked questions

Is Lambda or CoreWeave cheaper for H100 inference?

On on-demand list prices, Lambda H100 is $4.59 per GPU-hour and CoreWeave HGX H100 is $49.24 per eight-GPU instance, about $6.16 per GPU-hour. Lambda is cheaper at list; CoreWeave spot is about $2.46 and reserved can reach 60% off.

Does CoreWeave offer Kubernetes?

Yes. CoreWeave sells managed Kubernetes and Kubernetes-native networking. Lambda's pricing page focuses on instances, 1-Click Clusters and Superclusters; confirm managed control-plane options with Lambda.

Do Lambda and CoreWeave publish an uptime SLA?

Neither publishes a numeric GPU uptime SLA on its public pricing page. Ask both for the contractual uptime, measurement window, credit schedule and exclusions before you commit.

Which is better for multi-node training?

Both sell multi-node clusters. CoreWeave emphasises Kubernetes, networking and reserved capacity; Lambda sells 1-Click Clusters from 16 to 2,000+ B200 or H100 GPUs for two weeks to a year. Compare interconnect, topology, storage and control plane.

How much does egress cost?

CoreWeave lists internet egress, VPC and NAT as free. Lambda's pricing page does not list an egress rate, so confirm egress, public IP and interconnect charges in your quote.

Can I avoid GPUs entirely?

Yes, if you do not need custom weights or dedicated capacity. A managed API such as Plugsky removes GPU operations and bills a flat monthly plan, with fair-use concurrency and RPM limits.

Sources

Last checked 20 November 2026. GPU prices and capacity change; verify region, term and egress before committing. Not paid.