Compute for agents

Beam sources bare-metal servers around the world and turns them into VMs, sandboxes, task queues, functions and endpoints. Run coding agents, background jobs, CI builds, and GPU workloads when you need them. Compute is billed while your code is running.

$ curl -fsSL https://www.beam.cloud/install | sh

We buy hardware in bulk and pass along the savings: VMs from See pricing, H100s from $1.83 per hour, and no egress fees.

Beam powers agents and AI products at:

Customers

“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.”

– Louis Morgner, Co-founder @ Jamie

Sandboxes are isolated computers for untrusted code, like LLM outputs and agent tool calls. Each sandbox starts in about half a second with its own filesystem and networking. It can be paused, snapshotted, and restored.

Run one or thousands. Expose a port to get a preview URL, or attach a GPU when you need one.

your agent

Sandbox(...).create()

sandboxrun_code()

runs untrusted code

sandboxupload_file()

isolated filesystem

sandboxexpose_port()

preview URL

2 vCPU and 8 GiB for See pricing, billed by the millisecond.

from beam import Sandbox

sb = Sandbox(cpu=2, memory="8Gi").create()
print(sb.process.run_code("print(2 + 2)").result)
url = sb.expose_port(3000)

Pricing

Compute is billed by the second and stops when a VM is paused. No egress fees.

Member prices

Sandbox2 vCPU · 8 GiBSee pricing Virtual Machine new4 vCPU · 8 GiB · 50 GiB diskSee pricing CI/CD Runner new4 vCPU · 8 GiB runnerSee pricing GPUsServerless RTX 4090$0.69/hr Whole H100 machine$1.83/hr Function1 vCPU · 1 GiBSee pricing Endpoint2 vCPU · 2 GiBSee pricing Task Queue2 vCPU · 4 GiB workerSee pricing StoragePlanned · per used GB$0.0001/hr

Smallest VM1 vCPU · 1 GiB · 50 GiB diskSee pricing

USD, billed by usage. Storage is not billed yet.

Whole GPU machines from $0.44 per hour, and 8-GPU clusters with InfiniBand by quote. See GPU prices.
Storage is not billed yet. Planned: $0.0001 per GB-hour of data stored, with no free allowance. Storage pricing.
Running a lot of compute? Talk to us about volume pricing, or bring your own cloud.

Speed

Agents spend more time waiting for machines than running on them, so Beam keeps the wait short. We virtualize CPUs and GPUs directly on bare metal, with fast local NVMe disks next to the compute.

New sandbox, 100 launched at once

0.55s

Resume a paused VM, RAM intact

~1s

New VM from a cached image

~2.5s

Wake a paused VM with an HTTP request

~3.6s

Sandboxes start in about half a second, and memory snapshots let new ones start with processes already running.
VMs resume with RAM and running processes intact in about a second.
Set an idle timeout, and paused VMs wake when a request arrives.
Serverless apps keep containers warm for as long as you set with keep_warm_seconds, then scale to zero.

“Beam is amazing. I tested the CLI and in 5 minutes had something running on the cloud.”

– Leonardo Cuco, CTO @ Ween.ai