Summary: Modal, Runpod and Lambda all rent NVIDIA GPUs by the hour, but they bill in very different ways: Modal charges per second for serverless containers and adds CPU and memory on top, Runpod charges per second for dedicated Pods or scale-to-zero Serverless workers, and Lambda charges per minute for whole on-demand instances. The cheapest headline price is not always the cheapest bill, because idle time, minimum balances and storage rules decide what you actually pay. This comparison lines up the list prices from each official pricing page and works through the arithmetic for a few realistic jobs.
Everything here was checked on the official pricing pages and billing docs of Modal, Runpod and Lambda on the day this was written. I did not run paid workloads on these platforms for this post, so the cost examples below are arithmetic from published list prices, not invoices. Prices on GPU clouds move often (Lambda has a price change scheduled for November 16, 2026), so treat the numbers as a snapshot and recheck before you commit.
The short answer
- Pick Modal if you write Python, your traffic is spiky, and you want GPUs to appear when a function is called and disappear afterwards. The Starter plan includes $30 of compute each month.
- Pick Runpod if you want the lowest hourly price for a GPU you control like a normal machine (a Pod), or a simple scale-to-zero inference endpoint (Serverless) with a short default idle timeout.
- Pick Lambda if you need full multi-GPU HGX machines with lots of CPU, RAM and local SSD, and you will keep them busy for hours or days at a time.
How each platform bills you
Modal: per-second serverless, GPU plus CPU plus memory
Modal lists each GPU at a per-second rate. The H100 is $0.001097 per second, which works out to about $3.95 per hour. The A100 80 GB is $0.000694 per second (about $2.50 per hour), the L40S is $0.000542 (about $1.95 per hour) and the L4 is $0.000222 (about $0.80 per hour). On top of the GPU you pay for CPU at $0.0000131 per physical core per second (a physical core counts as 2 vCPUs) and memory at $0.00000222 per GiB per second. According to the pricing FAQ, you are charged for whichever is higher, the amount you requested or the amount you used, and each container has a minimum of 0.125 cores and 128 MiB.
The detail that matters most for your bill is idle time. Modal bills during application load time, while inputs are processed, and for the period a container stays alive after its last input. By default that window is 60 seconds. You can change it with the scaledown_window parameter on the function decorator, anywhere from 2 seconds to 20 minutes. A longer window means fewer cold starts but more billed idle GPU time.
Two more multipliers are listed in the plan comparison: choosing a specific region costs 1.15x to 1.75x the base price, and non-preemptible execution costs 3x. Modal also notes that a gpu="H100" request may be upgraded to an H200 at the same price; pass gpu="H100!" if you need a real H100, for example when benchmarking.
Runpod: Pods and Serverless, both per second
Runpod sells two main products. Pods are dedicated GPU instances. On the pricing page, Secure Cloud lists the H100 SXM at $3.99 per hour and the H100 PCIe at $2.89 per hour, while the page's FAQ says Community Cloud starts from $2.69 and $1.99 per hour for the same two cards. The A100 PCIe is $1.79 per hour on Secure Cloud (from $1.19 on Community Cloud), and an L4 Pod is $0.59 per hour. Pods are billed by the second for compute and storage, with no ingress or egress fees.
Serverless workers cost more per hour but scale to zero. The listed flex-worker rate is $5.39 per hour for an 80 GB H100, $2.72 for an 80 GB A100, and $0.75 for the 24 GB tier (L4, A5000, 3090). The Serverless docs say you are billed from the moment a worker starts until it fully stops, rounded up to the nearest second. That includes the cold start while your handler loads the model, but pulling the container image and downloading cached models normally happen while the worker is scaled down and are not billed. After the last job, a worker waits for the idle timeout, which defaults to just 5 seconds.
Lambda: whole instances, billed per minute
Lambda's On-Demand Cloud prices instances by the hour and bills in one-minute increments. Billing starts once an instance launches and passes health checks, and stops only when you terminate it; the docs are explicit that a running instance is billed whether or not you are using it. The instance table on the pricing page lists per-GPU-hour prices of $4.59 for HGX H100, $3.19 for HGX A100 80 GB, $1.99 for A100 40 GB, $7.69 for HGX B200 and $0.79 for Tesla V100, plus sales tax. Important: the page says these prices take effect on November 16, 2026 at 00:00 UTC (09:00 on November 16 in Korea), and until then current prices are only shown inside a Lambda Cloud account. Lambda invoices weekly.
Side-by-side comparison
| Modal | Runpod | Lambda | |
|---|---|---|---|
| Free allowance | Starter plan: $0/month with $30 of compute included each month | None listed; on-demand Pods need at least one hour's worth of credits to deploy | None listed |
| H100 list price | $0.001097/sec (about $3.95/hr), GPU only | Pod: $3.99/hr SXM, $2.89/hr PCIe (Secure Cloud); Serverless flex: $5.39/hr | $4.59 per GPU-hour (HGX H100, effective Nov 16, 2026) |
| A100 80 GB list price | $0.000694/sec (about $2.50/hr) | Pod: $1.79/hr (Secure Cloud); Serverless flex: $2.72/hr | $3.19 per GPU-hour (effective Nov 16, 2026) |
| Billing unit | Per second, plus CPU and memory | Per second (Serverless rounded up to the next second) | Per minute, for the whole instance |
| Idle behavior | Containers stay up 60 sec after the last input by default (2 sec to 20 min configurable), billed | Serverless idle timeout 5 sec by default; a running Pod is billed until stopped | Billed until you terminate the instance |
| Storage | Volumes $0.09/GiB/month, 1 TiB/month free | Volume disk $0.10/GB/month running, $0.20 stopped; network volume $0.07 (under 1 TB) | Filesystems billed per GiB-month in hourly increments, even when not mounted; no ingress or egress fees |
| Paid plan | Team: $250/month plus compute, $100/month credits, 50 GPU concurrency | Pay as you go; reserved capacity via sales | Pay as you go; 1-Click Clusters reserved from 2 weeks |
Worked examples from the list prices
These are not benchmarks. They only show how billing rules change the total for the same GPU time. I used the H100 because all three list it.
A two-hour fine-tuning run on one H100
- Modal: 2 hours of H100 is about $7.90. If the container also requests 4 physical cores and 32 GiB of memory, CPU adds about $0.38 and memory about $0.51, for roughly $8.79 in total. On the Starter plan, the first $30 each month is covered.
- Runpod Pod: 2 hours of H100 SXM on Secure Cloud is $7.98, or from $5.38 on Community Cloud, plus disk storage. CPU and RAM come with the Pod.
- Lambda: 2 hours at $4.59 per GPU-hour is $9.18 per GPU. Lambda's instance table lists large HGX configurations, so check the instance sizes actually available in your console before assuming you can rent a single GPU at that rate.
For long, steady jobs the hourly rate dominates, and Runpod Community Cloud is the lowest listed H100 price of the three. Before choosing Community Cloud for anything sensitive, read how Runpod describes the two tiers in its docs.
One isolated inference request that takes 3 seconds
- Modal with defaults: 3 seconds of work plus the default 60 seconds of idle time is 63 billed seconds of H100, about $0.069 before CPU and memory, and before any load time. With
scaledown_window=2, it drops to about 5 seconds, roughly $0.0055. - Runpod Serverless: 3 seconds plus the default 5-second idle timeout is 8 seconds at $5.39 per hour, about $0.012, plus any billed cold start while your handler loads the model.
- Lambda: not a good fit for this pattern; you would pay for the instance for every minute it is up between requests.
The lesson: for occasional requests, the idle window matters more than the hourly rate. Modal's longer default makes sense when requests arrive in bursts, because a warm container avoids a cold start; for a truly one-off call, shorten it.
How far the free $30 on Modal goes
Dividing $30 by the GPU-only hourly rate gives about 7.6 H100 hours, 12 A100 80 GB hours, 15 L40S hours or 37 L4 hours per month, minus whatever CPU and memory you use. Modal's own GPU guide recommends starting with the L40S for running (not training) models, since it has 48 GB of GPU memory at a lower price than the A100.
Editor's note
Who each one is for: Modal suits Python developers with bursty jobs and small teams who want a monthly free allowance. Runpod suits people who want a cheap dedicated GPU machine billed by the second, or a plain scale-to-zero endpoint. Lambda suits training jobs that keep 8 GPUs busy for hours and teams that want big local SSDs.
Not for: If you only need to call a popular open model, a per-token API is usually simpler than renting a GPU. Lambda is a poor fit for spiky traffic; Runpod Pods are a poor fit if you forget to stop them.
Gotchas I found in the docs:
- "I'm on Modal Starter, why was I charged?" The FAQ says the $30 covers compute only. Since September 1, 2026, Modal's Shared Endpoints are billed per token from the first request, even if your $30 is unused. Set a workspace budget if you want a hard cap.
- Modal cost higher than expected: check the 60-second default idle window, region selection (1.15x to 1.75x) and non-preemptible execution (3x).
- Runpod Pod won't deploy: on-demand Pods need at least one hour's worth of credits for the chosen configuration in your balance.
- Runpod data lost: when the balance hits $0, Pods are stopped; Pods without a network volume are terminated and their data cannot be recovered. Enable low-balance alerts or auto-pay. Also note stopped Pods' volume disks cost $0.20/GB/month, double the running rate.
- Runpod $80/hour cap: the default spend limit across all resources is $80 per hour; contact support to raise it before a large job.
- Lambda bill after the job finished: instances bill until terminated, and filesystems bill even when not mounted. Refunds come as service credits that expire after 12 months.
FAQ
Which of the three is cheapest for an H100?
On list prices checked today, Runpod Community Cloud has the lowest starting H100 price ($2.69 per hour for SXM, $1.99 for PCIe). For short, irregular jobs, Modal or Runpod Serverless can still cost less overall because nothing is billed once they scale to zero.
Does Modal's $30 free credit roll over?
The pricing page describes it as $30 of included compute per month on the Starter plan. It does not say unused credit carries over, so plan as if it resets monthly.
Why is Runpod Serverless more expensive per hour than a Pod?
You pay a higher rate in exchange for scale-to-zero: an idle Serverless worker is not billed, while a Pod is billed for as long as it runs. If your GPU would be busy most of the day, a Pod is cheaper; if it would sit idle most of the day, Serverless usually wins.
Related reading on this blog: vLLM Errors Fixed: Out of Memory, CUDA and Chat Templates for the serving engine you will likely run on these GPUs, Replicate vs fal.ai vs Fireworks: API Costs if you would rather pay per call than rent a GPU, and Together AI Pricing: Credits, Rates and Limits for another per-token option.
Pick by how many hours your GPU will actually be busy, then set the idle timeout before you run anything.
See the official screenshots at the end of this post. Image credit: my own screenshots of the official pricing pages of Modal (modal.com), Runpod (runpod.io) and Lambda (lambda.ai).
Last verified: 2026-11-10
Sources: Modal pricing, Modal cold start guide, Modal GPU guide, Runpod pricing, Runpod Pods pricing docs, Runpod Serverless pricing docs, Lambda pricing, Lambda billing docs.
댓글
댓글 쓰기