Skip to content
DocumentationTry Hanzo
High-Performance Computing

Hanzo Machines

Rent a machine by the hour, metered to your organization. Ask what sizes and regions exist, price a launch before you make it, and see it in the same list as the workers you already own.

NVIDIA H100

Dedicated GPU instances for AI/ML training

VRAM80GB HBM3
Cores16,896 CUDA
Tensor Cores528 TCs

AMD EPYC 7763

High-core count CPU instances

Cores64 Cores
Threads128 Threads
MemoryUp to 2TB

Storage Options

Ultra-fast NVMe and persistent storage

Local NVMeUp to 8TB
Network StorageUnlimited
I/O Performance7GB/s

How renting one works

Every step is an API call, so anything you do once by hand you can put in a script afterwards

Ask what exists

The sizes and regions you can launch into are an endpoint, not a page in the docs. Read them, then pick one — nothing to keep in sync by hand.

Price it before you buy it

A launch with dryRun returns the price and creates nothing. The same call without it creates the machine, so what you were quoted is what you asked for.

Metered to the organization

Time is debited from your org's prepaid balance while a machine is up. Before it launches, the cloud says whether you may — and what is missing if you may not.

One list, whoever owns them

Rented machines, the droplets and cluster workers behind them, and your own hardware that dialed in with hanzo link all appear as one fleet, deduplicated.

A row per accelerator

The GPU view counts real cards: from the size of the nodes you rented, and from what your own workers report through nvidia-smi. Where there is no inventory, it returns nothing rather than inventing something.

A machine can carry an agent

Bind a cloud agent to a machine and the pair is one thing to launch, message, stop and tear down. Unbinding is separate, so the agent can move without the machine going with it.

When you want a whole machine

Reach for one when the work outlasts a request, or needs a card of its own

Training runs

A job that needs cards for hours and nothing for the rest of the week.

  • GPU sizes named from the list the API returns
  • GPU-hours metered while the job runs
  • Cancel and the hours used so far are billed, then the job stops

Serving a model

A machine that stays up because something is calling it.

  • Run Hanzo Engine on a card you rented
  • Terminate it when the traffic goes away
  • The bill follows the machine, not a reservation

Work that will not fit in a request

Renders, batches and simulations that take longer than an HTTP call.

  • A queue per GPU, plus a shared any-GPU lane
  • Each item says which node claimed it
  • Read the queue instead of guessing where a job went

Agents that need a computer

An agent with a machine under it, launched and torn down as one thing.

  • Bind a cloud agent to a machine
  • Message it, or stop it, by name
  • Tearing down the pair unbinds the agent first

Whose machine it is

Ownership, region and account are decisions the API makes you state

Scoped to your organization

A machine belongs to one org and is named within it. Every read and every terminate is checked against the org on your token, not against the id in the URL.

Bring your own account

Link a DigitalOcean, AWS or GCP account and the clusters it holds fold into the same fleet. Your capacity, your contract, one place to see it.

Pick the region deliberately

Launches name a region from the list the API returns, so where a workload runs is a decision you made rather than one that was made for you.

Launch one, or attach one you own

Rent by the hour from the console or the CLI, or run hanzo link on a machine you already have and it joins the same fleet.

Get started with Machines

Open source

License: Apache-2.0hanzoai

Get Machines

Long-running compute machines