> ## Documentation Index
> Fetch the complete documentation index at: https://docs.acasia.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Bare metal reservations

> Dedicated GPU hardware with root access, held by one organization for a defined term. Reservations are invite-only and arranged through the Acasia sales team.

A bare metal customer holds GPU hardware reserved for their organization alone. No other tenant shares the machines, the customer holds root access and runs their own stack, and the hardware is configured around the workload rather than around a listing.

This covers both ends of the range: a large deployment built from the ground up, and a smaller reserved allocation on Acasia's own GPU hardware. Both are arranged through the Acasia sales team.

## What a reservation includes

| Included                      | Description                                                                                              |
| ----------------------------- | -------------------------------------------------------------------------------------------------------- |
| Dedicated hardware            | Machines reserved for your organization for the length of the term                                       |
| Root access                   | Full administrative access to the machines, so you run your own stack on them                            |
| Configuration to the workload | GPU model and count, cluster topology, networking, storage, and interconnect                             |
| Acasia Cloud                  | Node-level telemetry, hardware alerting, machine inventory, credentials, and organization administration |
| Ongoing operations            | Monitoring, capacity changes, and support for the life of the term                                       |

Where the control plane is installed on the reserved hardware, Acasia Cloud also reports per-GPU telemetry and can drive model deployment and cluster sizing. See [The control plane](/marketplace/control-plane).

## Bare metal compared with the Marketplace

|                        | Bare metal                                      | Marketplace                                    |
| ---------------------- | ----------------------------------------------- | ---------------------------------------------- |
| **Tenancy**            | Dedicated to one organization                   | Capacity listed by third-party hardware owners |
| **How you buy**        | Through the Acasia sales team, by invitation    | Self-serve checkout                            |
| **Configuration**      | Customized to the workload                      | Selected from the listed capacity              |
| **Term**               | Defined term, agreed up front                   | End time you set at checkout                   |
| **Commitment**         | Larger capacity, longer term                    | No minimum term                                |
| **Time to compute**    | Scheduled with your Acasia team                 | Immediate after payment                        |
| **Machine access**     | Root access — you run your own stack            | Managed through the control plane              |
| **Telemetry in Cloud** | Node level, or GPU level with the control plane | GPU level, scoped to the GPUs you rented       |

## Who bare metal fits

* Organizations training or fine-tuning models at a scale that needs dedicated interconnect and topology
* Teams running sustained production inference where capacity must be guaranteed for the term
* Organizations with data residency, isolation, or compliance requirements that rule out shared capacity
* Teams that need administrative control of the machines their workloads run on

## Becoming a bare metal customer

Reservations are invite-only. They are not listed, priced, or purchased online — every reservation is scoped with the Acasia sales team before hardware is committed.

Start by contacting Acasia at [acasia.com/contact](https://acasia.com/contact) or emailing [support@acasia.com](mailto:support@acasia.com). The more you can describe up front, the faster the deployment can be scoped:

| Detail                 | Why it matters                                                                  |
| ---------------------- | ------------------------------------------------------------------------------- |
| Workload               | Training, fine-tuning, inference, or a mix — this drives GPU model and topology |
| Scale                  | Target GPU count, and whether it grows over the term                            |
| Term                   | How long the capacity is needed                                                 |
| Timeline               | When the deployment needs to be live                                            |
| Networking and storage | Interconnect, throughput, and dataset size requirements                         |
| Constraints            | Data residency, isolation, compliance, or region requirements                   |

<Note>
  If you need GPU capacity today, the [Marketplace](https://app.acasia.com/marketplace) is the faster path. Rentals there are self-serve and available immediately after checkout.
</Note>

## Next steps

<CardGroup cols={3}>
  <Card title="Contact Acasia" icon="envelope" href="https://acasia.com/contact">
    Start the conversation with the sales team.
  </Card>

  <Card title="Acasia Cloud" icon="cloud" href="/cloud/overview">
    Telemetry, alerting, and administration for your fleet.
  </Card>

  <Card title="Marketplace" icon="cart-shopping" href="/marketplace/overview">
    Rent listed capacity while a deployment is being scoped.
  </Card>
</CardGroup>
