PARAS MADAN
All writing

Latest in AI

Lambda GPU Pricing: H100 and B200 Instance Costs Explained

The November 16 rates are quoted per GPU, but the bill depends on the whole instance and when you actually terminate it.

Lambda’s instance rates effective November 16, 2026 put an eight-GPU H100 instance at $36.72 an hour and an eight-GPU B200 instance at $61.52 an hour, before tax. The headline rates are per GPU, not per machine, which is the first distinction to get right when budgeting a training run.

Lambda H100, B200 and A100 pricing

The official pricing page lists these eight-GPU configurations, with rates effective from 00:00 UTC on November 16. The hourly and daily totals below are calculated by multiplying the listed GPU rate by eight, then by 24 for a full day.

Eight-GPU instanceVRAM per GPUUSD per GPU/hourUSD per instance/hourUSD per 24 hours
HGX H10080 GB$4.59$36.72$881.28
HGX B200180 GB$7.69$61.52$1,476.48
HGX A10080 GB$3.19$25.52$612.48

These are instance rates, not a quote for Lambda’s separate 1-Click Clusters service. Availability is first-come, and applicable sales tax, VAT or GST is additional. A published price is not a promise that your preferred configuration is available when you need it.

Calculate the billable run, not just the training loop

For a budget estimate, use GPU count × price per GPU/hour × billable hours. Keep storage and tax outside that subtotal so the estimate remains easy to audit.

Consider a hypothetical eight-H100 experiment with 12 billable hours spent doing useful work and another three hours left idle afterward. The useful portion costs $440.64. The idle period adds $110.16, bringing compute alone to $550.80. No performance testing is implied here; this is arithmetic using the published rate.

The mistake would be recording only the training script’s runtime. Environment setup, data preparation and waiting for someone to inspect results can all sit outside that timer. For a team running experiments intermittently, the cleanup process deserves as much attention as a small difference in GPU price.

Does Lambda charge for idle instances?

Yes. Lambda’s billing documentation says on-demand instances are billed in one-minute increments, starting after launch and successful health checks and ending at termination. A running instance remains billable when idle.

Persistent filesystems have a separate meter based on GiB used, charged in one-hour increments. They continue accruing charges while they exist, even when detached. The documentation’s $0.20 example is illustrative, not a verified current storage quote; check the price shown when creating the filesystem. Lambda says there are no ingress or egress charges.

That makes a useful cost review two separate questions: did the compute instance terminate, and which storage resources are still deliberately retained? Keeping a checkpoint may be sensible. Accidentally keeping its GPU machine alive is a different expense.

How to stop compute billing without losing your work

Lambda’s instance-management guide warns that operating-system commands such as sudo shutdown -h now do not terminate the cloud resource. They can put it into Alert status while billing continues. It also says instances cannot currently be suspended.

Before ending a run, verify that checkpoints, outputs and anything else you need have been backed up outside the instance. Then select the instance in the cloud console and use Terminate, or use the documented termination API. Do not treat a disconnected SSH session, an exited process or an OS shutdown as evidence that spending has stopped.

For a repeatable workflow, make the saved output location and the person responsible for cleanup explicit before launching. A job that finishes overnight should not depend on someone remembering which machine they started yesterday. Automating cleanup can help, but it needs a verified backup step rather than blindly terminating after a timer expires.

When is the more expensive GPU worth it?

The right comparison is cost per completed workload, with the same output and quality requirements. An expensive configuration can be the better choice if it finishes enough sooner, while a cheaper one can be wasteful if the workload does not fit or spends most of its time waiting.

Set that comparison up before a long reservation: choose a representative job, record elapsed billable time, check the output and calculate the total. Do not infer a speedup from the hardware name or compare two runs with different model settings. These prices establish a budget, not a benchmark.

The practical advantage comes from making the whole experiment reproducible, including its ending. A team that can reliably save results, terminate compute and restart later has more control over its bill than one that only shops for a lower hourly number.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical teams turn infrastructure choices and product tradeoffs into clear stories developers can use.

Book a 30-minute call.