Serverless GPU Pricing Matrix [2026 Developer Cheat Sheet]

Prices verified on May 7, 2026 from official vendor pricing and documentation. The tricky part is not just the posted $/GPU-hour number: Modal, Replicate, and Lambda meter work very differently. This cheat sheet is optimized for engineering teams that need a fast reference for bursty inference, always-on deployments, and cost modeling before they commit to one platform.

• Modal H100: $3.9492/hr from $0.001097/sec.
• Replicate H100: $5.49/hr for gpu-h100.
• Lambda H100: $2.86/hr for 1x H100 PCIe, $3.78/hr for 1x H100 SXM.
• Idle billing is the real separator: Modal minimizes it, Lambda charges while running, Replicate depends on model type.

Note: the price winner changes with workload shape. Lambda is frequently cheapest per posted GPU-hour, but that does not make it cheapest for spiky traffic if the VM sits idle. Replicate also mixes hardware-time pricing with model-level pricing, so compare the specific model page before locking estimates.

Pricing Matrix

Common SKUs at a glance

• Modal publishes per-second GPU pricing, including B200, H200, H100, A100, L40S, A10, L4, and T4.
• Replicate publishes hardware pricing for A100 80GB, H100, L40S, and T4, plus multi-GPU variants.
• Lambda publishes instance pricing for B200, GH200, H100 PCIe, H100 SXM, A100, A10, A6000, and Quadro RTX 6000.

When To Choose Each

Choose Modal when:

• You want serverless GPU jobs without paying for idle machines.
• You are shipping Python-first inference, batch jobs, queues, or web endpoints.
• You need access to newer GPUs like H200 or B200 with simple code-level configuration.
• You want cleaner cost control for bursty traffic than a permanently running VM.

Choose Replicate when:

• You want an API for hosted models now, not a platform migration project.
• You value official models, stable model APIs, and built-in webhook patterns.
• You are comfortable paying a premium for packaging, distribution, and marketplace convenience.
• You want to expose inference quickly to web or product teams.

Choose Lambda when:

• You need full VM access, SSH, custom services, notebooks, or background daemons.
• Your jobs are long-running enough that VM-style billing is acceptable.
• You want strong raw $/GPU-hour economics and are willing to manage more infrastructure.
• You need a bridge from experimentation to more traditional cluster or multi-node workflows.

Commands By Purpose

Keyboard shortcuts for this reference page

pip install modal
modal token new
modal token info

export REPLICATE_API_TOKEN=r8_******
npm install replicate

import Replicate from "replicate";
const replicate = new Replicate();

const [output] = await replicate.run(
  "black-forest-labs/flux-schnell",
  {
    input: {
      prompt: "An astronaut riding a rainbow unicorn, cinematic, dramatic"
    }
  }
);

modal run app.py::train
modal serve app.py
modal deploy --env=prod --stream-logs app.py

curl -s -X POST \
  -H 'Prefer: wait' \
  -H "Authorization: Bearer $REPLICATE_API_TOKEN" \
  -H 'Content-Type: application/json' \
  -d '{"version": "5c7d5dc6dd8bf75c1acaa8565735e7986bc5b66206b55cca93cb72c9bf15ccaa", "input": {"text": "Alice"}}' \
  https://api.replicate.com/v1/predictions

curl -X POST \
  -H "Authorization: Bearer $REPLICATE_API_TOKEN" \
  -H "Cancel-After: 1m30s" \
  -H "Content-Type: application/json" \
  -d '{"input": {"prompt": "The sun rises slowly between tall buildings."}}' \
  https://api.replicate.com/v1/predictions

ssh -i '<SSH-KEY-FILE-PATH>' ubuntu@<INSTANCE-IP>
ssh -L <LOCAL-PORT>:127.0.0.1:<REMOTE-PORT> ubuntu@<INSTANCE-IP>

rsync -av --info=progress2 <FILES> <USERNAME>@<SERVER-IP>:<REMOTE-PATH>

s5cmd ls s3://<S3-BUCKET>
rclone ls my-remote:example-bucket
rclone -P copy <REMOTE>:<BUCKET-NAME> <LOCAL-DIR>

Configuration

Modal profiles, environments, and secrets

• Authentication can be stored with modal token set or environment variables like MODAL_TOKEN_ID and MODAL_TOKEN_SECRET.
• Environment scoping is first-class: create one with modal environment create dev and select it with modal config set-environment dev or MODAL_ENVIRONMENT.
• Secrets can be created with the CLI and injected into functions.

modal environment create dev
modal config set-environment dev
modal secret create api-creds OPENAI_API_KEY=sk-******

Replicate token hygiene

• Replicate API tokens are 40-character secrets that begin with r8_.
• Store them in environment variables, not source files.
• If you paste examples into docs or tickets, scrub them first with TechBytes' Data Masking Tool.

export REPLICATE_API_TOKEN=r8_******

Lambda account setup

• API keys authenticate Lambda Cloud API operations.
• SSH keys are required before launching instances you want to access directly.
• By default, Lambda allows only incoming ICMP and TCP/22 unless you change firewall rules.
• Attached filesystems mount at /lambda/nfs/<FILESYSTEM_NAME>.

echo '<PUBLIC-KEY>' >> ~/.ssh/authorized_keys
cat ~/.ssh/authorized_keys

Advanced Usage

Modal: pick GPU type in code

Modal keeps GPU selection close to the workload definition, which is useful when you want code review to capture cost-sensitive changes.

import modal

app = modal.App()
image = modal.Image.debian_slim().pip_install("torch")

@app.function(gpu="H100:2", image=image)
def run():
    import torch
    print(torch.cuda.is_available())

• Modal supports appending :n to request multiple GPUs per container.
• The current docs list GPU values including T4, L4, A10, L40S, A100, H100, H200, and B200.

Replicate: use sync for latency, async for orchestration

• Use Prefer: wait for short-lived inference where you want the output in one request.
• Use default async mode plus webhooks for background jobs.
• Use Cancel-After to cap runaway cost on slow generations.
• Remember that official models may use predictable per-output pricing instead of hardware-time pricing.

Lambda: treat storage and connectivity as first-class concerns

• Single-GPU instances usually launch in 3-5 minutes; multi-GPU instances usually take 10-15 minutes.
• Billing starts after the instance boots and passes health checks, then continues while the instance runs.
• Use rsync, rclone, or s5cmd for data movement. The Filesystem S3 Adapter adds S3-compatible tooling support in select regions.
• If you only need an internal service exposed locally, prefer an SSH tunnel over opening public ports.

ip -4 -br addr show enp5s0
nmap -Pn INSTANCE-IP-ADDRESS

Sources And Notes

• Modal Pricing
• Modal GPU Docs
• Modal CLI: run
• Modal CLI: deploy
• Modal CLI: serve
• Modal CLI: token
• Modal CLI: secret
• Modal CLI: environment
• Replicate Pricing
• Replicate Node.js Quickstart
• Replicate Predictions API
• Replicate API Tokens
• Replicate Official Models
• Lambda Pricing
• Lambda: Creating and Managing Instances
• Lambda Billing
• Lambda: Connecting to an Instance
• Lambda: Importing and Exporting Data
• Lambda Filesystems
• Lambda Filesystem S3 Adapter

If you want to turn the snippets above into team-safe examples, run them through the TechBytes Code Formatter before publishing them into internal docs or runbooks.