AI and Machine Learning on the cloud
Your whole pipeline at scale, not just the fancy models. Fine-tune LLMs on A100s. Clean 10 TB of nested JSON files.
Scale your entire AI pipeline
Most AI services fixate on GPUs, but real-world AI is 90% data wrangling and 10% training. Coiled makes it easy to run your entire pipeline on the cloud, from data ingestion to model training and batch inference.
One platform for all your cloud compute needs.
- CPU and GPU scaling in one platform
- Multi-cloud flexibility: AWS, GCP, or Azure
- Python-first: Dask, Hugging Face, XGBoost, PyTorch, and more
- Zero DevOps, no Kubernetes required
Scale AI/ML workflows on the cloud
Skip the 500-line YAML file.
Fine-tune LLMs on GPUs with PyTorch DDP.
- Train on massive datasets with Accelerate and PyTorch
- Multi-node, distributed model training and batch inference
- Get as big a VM as AWS will give you
- No Docker required
#!/usr/bin/env bash
# COILED n-tasks 10
# COILED gpu True
# COILED task-on-scheduler True
accelerate launch \
--multi_gpu \
--machine_rank $COILED_BATCH_TASK_ID \
--main_process_ip $COILED_BATCH_SCHEDULER_ADDRESS \
--main_process_port 12345 \
--num_machines $COILED_BATCH_TASK_COUNT \
--num_processes $COILED_BATCH_TASK_COUNT \
nlp_example.py#!/usr/bin/env bash
# COILED n-tasks 10
# COILED gpu True
# COILED task-on-scheduler True
accelerate launch \
--multi_gpu \
--machine_rank $COILED_BATCH_TASK_ID \
--main_process_ip $COILED_BATCH_SCHEDULER_ADDRESS \
--main_process_port 12345 \
--num_machines $COILED_BATCH_TASK_COUNT \
--num_processes $COILED_BATCH_TASK_COUNT \
nlp_example.pyDemand forecasting at scale
Learn how Guac accelerates ML pipelines with Coiled. For a small, rapidly growing team, the challenge was scaling their pipelines to handle exponentially more data without rebuilding their entire infrastructure.
A complete ETL + ML platform.
- Seamless workflow from data ingestion to prediction
- Works with their ML stack: PyTorch, XGBoost, Optuna
- ETL orchestration with Dagster
- Scaled existing pandas code with Dask
- No complex migration to Spark
Cloud Compute for Python People
Going from your laptop to the cloud should be easy.
Parallel Python that just works.
- Run on a big machine or a cluster of VMs
- Autoscale up and down based on your workload
- Dask clusters for distributed computing
import coiled
@coiled.function()
def simulate(trial: int=0):
return ...
# Run once on the cloud
result = simulate(1)
# Run in parallel on 1000 machines
results = simulate.map(range(1000))
# Retrieve results
list(results)import coiled
@coiled.function()
def simulate(trial: int=0):
return ...
# Run once on the cloud
result = simulate(1)
# Run in parallel on 1000 machines
results = simulate.map(range(1000))
# Retrieve results
list(results)Trusted by Data Teams
Reliable compute for mission-critical workloads that actually stays up
"I've been incredibly impressed with Coiled; it's quite literally the only piece of our entire ETL architecture that I never have to worry about."
Bobby George
Co-founder, Kestrel
"The speed is nice, sure, but the real benefit is taking a multi-day effort and finishing it in an afternoon. Coiled changed the character of our work."
Matt Plough
Software Engineer, KoBold Metals
"My team has started using Coiled this week. Got us up and running with clusters for ad hoc distributed workloads in no time."
Mike Bell
Data Scientist, Titan
"Coiled is natural and fun to use. It's Pythonic."
Lucas Gabriel Balista
Data Science Lead, Online Applications
FAQ
Yes, and not just one.
Coiled supports single-GPU jobs and multi-node distributed GPU training with tools like 🤗 Accelerate.
See GPU Support for more details.
In your AWS, GCP, or Azure account, where they belong.
Your data never leaves your cloud account. We just:
- Turn on VMs when you need them
- Clean them up when you're done
- Put logs in your cloud logging system
Coiled never sees your data. See Security for more details.
Surprisingly little.
- Pay your cloud provider for compute (usually $0.02-0.05 per CPU-hour)
- First 500 CPU hours per month are free
- After that, $0.05 per CPU-hour to Coiled
Most workloads cost just pennies to run. See Pricing for more details.
We handle that automatically.
Coiled handles environment management complexity, but unlike SageMaker or Vertex AI, this doesn't come with constraints. Instead of restrictive, pre-defined environments, you can replicate your local Python environment with Package Sync, or bring your own custom Docker container. Customize GPU driver versions, CUDA configs, or kernel-level optimizations.
Not a chance.
We're a bit obsessive about cleaning up:
- VMs get terminated after your job completes
- Storage gets removed
- Network resources get deleted
No more "oops, I forgot to turn off that instance from last quarter" moments.
Coiled is Python-first and actually easy to use.
With Coiled you have:
- Zero infrastructure at rest. SageMaker and Vertex AI have serverless options that scale down to zero, but other resources (notebooks, endpoints) must be manually stopped. With Azure Machine Learning, everything stays on unless explicitly configured.
- No environment restrictions. SageMaker and Vertex AI abstract away some of the complexities of environment management, at the cost of increased constraints. You are often locked into a predefined environment (unless you build a custom Docker image) and there's limited access to low-level systems (like kernel modules or root privileges). For example, pynvml, may silently fail since it requires root-level or direct NVIDIA driver access.
- No cloud vendor lock-in. Each major cloud provider has it's own specific UX and set of complexities that come with it. Azure ML and SageMaker are notorious for complex onboarding that requires cloud provider-specific expertise. Vertex AI is a little easier to get started with Workbench, but still has a strong GCP-specific UX to learn.
Get started
Know Python? Come use the cloud. Your first 500 CPU hours per month are on us.
$ pip install coiled
$ coiled quickstart
Grant cloud access? (Y/n): Y
... Configuring ...
You're ready to go. 🎉$ pip install coiled
$ coiled quickstart
Grant cloud access? (Y/n): Y
... Configuring ...
You're ready to go. 🎉