ServerlessGPUs

Run inference, agents, and task queues on our cloud, or bring your own AWS, GCP, or bare metal. Sub-second cold starts, no rate limits. 4090s from $0.69/hr.

Copy Prompt

InferenceSandboxesTask Queues

• Python
• T
• TypeScript
• Go

from beam import Image, endpoint

# Serve a model on an H100, autoscaled for you
@endpoint(gpu="H100", image=Image().add_python_packages(["vllm"]))
def generate(prompt: str):
    llm = vllm.LLM(model="openai/gpt-oss-20b")
    return {"output": llm.generate(prompt)[0].text}

$ beam deploy Running on B200 in US Central

The runtime for AI workloads at scale

Engineered from the ground up for heavy AI workloads. Sub-second cold starts, massive parallelism, and full observability.

Live Usage

• Containers 179
• Utilization 98%

AWS Account

• 74 Containers

GCP Account

• 39 Containers

Beam Bursting

• 14 Containers

Run AI workloads across clouds

Connect all your cloud accounts to Beam, and run workloads across all of them. Achieve the highest cloud utilization and maximum scale.

Boot time (s)

• Beam (memory snapshots): 0.05s
• Beam: 0.11s
• Other runtimes:
  • Provider A: 1.61s
  • Provider B: 2.28s
  • Kubernetes + EC2: 3.59s

Sub-second cold starts

Memory snapshots restore GPU containers in seconds — up to 35× faster than a traditional cold boot.

• 30+ Regions
  • us-west
  • eu-west
  • ap-south
  • ca-east
  • eu-central
  • us-east

Run near your agents

Workloads route across clouds and regions in real time, for low-latency execution wherever your users are.

Massive parallelization

Snapshot a running sandbox, then restore it into thousands of concurrent isolated runs, each with realtime streaming output.

From serverless inference to sandboxes

Logic and hardware in one place — no YAML, no Dockerfiles, no infra to manage.

1. Inference
2. Task queues
3. Sandboxes

GPU Inference

Sub-second cold starts

We provide a distributed storage layer, memory snapshotting, and GPU checkpoint restore, resulting in lightning fast container boot times.

Scale down to zero, burst to thousands Only pay for what you use

from beam import QueueDepthAutoscaler

# Scale out when queue size > 30 tasks
autoscaling_config = QueueDepthAutoscaler(
    tasks_per_container=30,
    max_containers=300,
)

Durable Task Queues

Retries, Callbacks, and Scheduled Jobs

Control the full lifecycle of a task with automated retries and event-based callbacks to your application.

Sandboxes for AI Agents

Stateful, Persistent Runtimes

Sandboxes are stateful. You can connect to a running process, attach persistent storage volumes, and snapshot the file system to create reusable templates.

Run Docker-in-Docker

from beam import Image, Sandbox

sb = Sandbox().create()
image_id = sb.create_image_from_filesystem()
sb.terminate()

sb = Sandbox(image=Image.from_id(image_id)).create()

Connect any VM, anywhere

Bring your own cloud. No lock-in

Switch Hardware in Seconds

Run your code on any hardware in seconds — just change one line of Python to switch hardware.

Easy Local Debugging

Test your code before deploying it, using the exact configuration you'll run in production.

Multiple Workers Per Container

Scale vertically by running multiple workers on the same container.

Run Docker-in-Docker

Run the full Docker daemon in your containers.

Deploy from GitHub Actions

Deploy your APIs automatically by adding Beam to your existing CI/CD pipeline.

One platform for inference, sandboxes, and training

Use Cases

• Custom Model Inference
  Host any custom model on GPU or CPU. Bring your own image.
• Sandboxed Code Execution
  Run LLM-generated code in secure execution environments.
• RL Environments
  Fork sandboxes from snapshots to run RL rollouts in parallel.
• Training & Fine-Tuning
  Train and fine-tune, from SLMs and LLMs to diffusion models.
• Audio Processing Pipelines
  Deploy task queues to process large amounts of data.
• Streamlit and Gradio UIs
  Run frontend apps, from Streamlit and Gradio apps to Notebooks.

Examples

What will you build?

• Inference
  Serve LLMs at high throughput
  Run vLLM for fast, batched inference with maximum tokens per second.
• Audio
  Transcribe audio at scale
  Deploy Whisper to turn speech into text across thousands of files.
• Image
  Run ComfyUI image pipelines
  Spin up node-based image generation workflows on serverless GPUs.
• Speech
  Generate natural speech
  Deploy Parler-TTS to turn text into expressive, lifelike audio.
• Training
  Fine-tune Gemma models
  Train and customize Gemma on your own data with GPU-backed jobs.

Join our community

Beam is powering hands-down the best developer experience to run models on GPUs easily at scale. Best decision on the infra side for us this year so far.

Testimonials

• Louis Morgner, Co-founder, AI lead @ Jamie
  "@beam_cloud is 🔥. Such a huge workflow improvement over AWS Sagemaker / Google vertex ai"
• Eric Meier, @bitphinix
  "One of the better developer experiences I've had in a while was with @beam_cloud - a serverless GPU and API infra platform. Check them out 👇"
• Brandon Garcia, @BCG
  "I can't recommend Beam highly enough. Their developer experience is top notch."
• Liam Eloie, Machine Learning Engineer
  "Beam has been a huge time-saver by eliminating the need to monitor and manage my own VM infrastructure."

Start shipping on infra

Run sandboxes and GPU workloads on your cloud, and scale out to ours when you need to. No infra to manage.