Serverless GPU — no VMs, no SSH, no idle waste
Lambda Labs gives you raw GPU cloud VMs. Velar gives you a serverless Python runtime where you write a function, specify a GPU, and deploy — no infrastructure to manage, no hours wasted on idle instances.
From raw VMs to serverless Python
Lambda Labs is a solid choice for teams that need full control over a GPU machine. But if you are a Python developer who wants to run ML workloads without DevOps overhead, Velar removes the entire infrastructure layer.
No VMs to manage
Lambda Labs gives you a raw GPU instance — you SSH in, set up your environment, and manage the lifecycle yourself. Velar handles all of that so you can write Python and deploy.
No idle GPU cost
Lambda instances keep billing until you stop them. Forget to stop one overnight and you pay for 8 idle hours. Velar scales to zero automatically between function invocations.
Deploy in seconds, not minutes
Lambda VM provisioning takes 1–5 minutes. Velar deploys from Python in under 60 seconds and subsequent cold starts are measured in seconds, not minutes.
Velar vs Lambda Labs
| Feature | Velar | Lambda Labs |
|---|---|---|
| Serverless (no VM management)Lambda Labs provides raw GPU VMs you manage yourself | ||
| Scale to zero when idleLambda instances run and bill until you stop them | ||
| Python-native SDK (no SSH required) | ||
| Sub-minute deploysLambda VM provisioning takes minutes | ||
| Per-second billingLambda bills per hour, instances must be manually stopped | ||
| Automatic auto-scaling | ||
| No Dockerfile required | ||
| Raw VM / SSH accessVelar is serverless — use Lambda if you need root access | ||
| Free starter credits |
Zero infra, pure Python
With Lambda Labs you provision a VM, SSH in, install CUDA, clone your repo, install dependencies, and run your script. With Velar, you write Python and deploy. The runtime handles GPU provisioning, CUDA, dependencies, and scaling.
- No SSH, no VM lifecycle management
- CUDA and drivers managed automatically
- Scales to zero — pay only for actual compute
- Long-running jobs up to 24 hours supported
import velar
app = velar.App("training-run")
# No SSH, no VM setup — just Python
image = velar.Image.from_registry(
"pytorch/pytorch:2.1.0-cuda12.1-cudnn8-runtime"
).pip_install("transformers", "peft", "datasets", "trl")
@app.function(gpu="A100", timeout=7200, image=image)
def fine_tune(dataset_path: str):
from transformers import AutoModelForCausalLM, TrainingArguments
from trl import SFTTrainer
model = AutoModelForCausalLM.from_pretrained(
"meta-llama/Llama-2-7b-hf"
)
# ... your training logic
return {"status": "done", "checkpoint": "/tmp/model"}
# Runs on A100 — scales to zero when done
app.deploy()
result = fine_tune.remote("s3://my-bucket/dataset.jsonl")Frequently asked questions
Is Velar appropriate if I need full root access to a GPU server?
No — if you need to install system packages, run custom CUDA kernels at the OS level, or require persistent SSH access, Lambda Labs is the right tool. Velar is for Python-level GPU workloads.
How does pricing compare to Lambda Labs?
Velar bills per second and scales to zero. Lambda Labs charges hourly and instances stay running until stopped. For bursty or intermittent workloads, Velar is almost always cheaper because you pay for actual compute time, not reserved time.
Can I run long training jobs on Velar?
Yes. Velar functions support timeouts up to 24 hours, making them suitable for fine-tuning runs, embedding generation, and other long batch jobs — without needing to babysit a VM.
Do I need to set up CUDA and drivers?
No. Velar handles GPU driver management. You specify a base image with the CUDA version you need and pip install your dependencies — the rest is handled for you.
Can I use multiple GPUs in parallel?
Yes. Call your function multiple times and Velar provisions separate GPU workers per call. You can fan out across dozens of A100s simultaneously using standard Python list comprehensions.
Less infrastructure. More building.
Velar Cloud is paused. Join the waitlist for the relaunch. Your first deployment takes under 5 minutes.