cloudparts

Oxide Cloud Computer

A whole rack sold as one computer: AMD EPYC sleds, two switches, power shelves and the software to run VMs, disks and private networks through one API, in your own data center.

What it is
Virtual machines,Block storage,Virtual private clouds
Type
Commercial product
Made by
Oxide Computer Company
License
Commercial hardware; its software is open source, mostly under MPL 2.0
Facts checked
13 October 2026
Website

What it is

The unit you buy is a full rack. It holds up to 32 first-generation or 24 second-generation compute sleds, two “Sidecar” switches built on Intel’s Tofino 2 chip, and two power shelves. A second-generation sled has one AMD EPYC 9005 processor with up to 192 cores, up to 1.5 TiB of DDR5 memory, ten NVMe bays of up to 30 TB each and two 100 GbE ports. Sleds plug into a cabled backplane at the back of the rack, so adding capacity means sliding one in: there is no front cabling. Per rack, guests can get up to 7,875 vCPUs, 30.6 TiB of memory and 1.7 PiB of block storage.

On top runs a public-cloud-style service, all of it written by Oxide. Operators split the rack into silos, isolated tenants with their own users and identity provider. Inside a silo, projects hold instances (VMs), disks, snapshots, images and a VPC. A distributed disk keeps three copies on three different sleds and encrypts the data on its way over the rack’s network; a local disk is one unreplicated copy on the VM’s own sled. VPCs are Geneve overlays with subnets, firewall rules, NAT for outbound traffic and floating IPs. Everything is reachable through one HTTP API, which the web console, the oxide CLI, SDKs for Rust, Go and TypeScript, and a Terraform provider all use.

What sets it apart is that the same company designed every layer. Each board has a service processor and a hardware root of trust instead of a traditional BMC, the host boots Oxide’s own illumos distribution, Helios, without a UEFI BIOS, VMs run on bhyve with Oxide’s Propolis, disks are served by Crucible and networking by OPTE and the rack’s own switch. Software arrives as one system release (v23 shipped in September 2026), and there are no per-feature licenses: the software is included with the hardware and published as open source.

Why teams pick it

  • One vendor for compute, storage, networking and the software that ties them together, with no hypervisor or storage licenses to buy and renew
  • A cloud-style API for everything, including silos, projects, quotas, VPCs and floating IPs; the console is built on the same public API, and Terraform, Kubernetes (Cluster API, Omni, Rancher) and Packer integrations use it too
  • Distributed disks keep three copies on three sleds: guests keep reading and writing with one copy offline, and drop to read-only with two offline
  • A root of trust on every server and switch checks its own firmware, data at rest is encrypted, and sleds must form a trust quorum before storage unlocks

Watch out for

  • You buy a whole rack, and it must be installed and configured under the guidance of Oxide technicians
  • One rack is one pool: the docs list joining several racks into a single pool of infrastructure as a roadmap item, not yet available
  • Guests must be x86 UEFI images with VirtIO drivers; nested virtualization is unsupported, and Windows versions that need a TPM, such as Windows 11, are not supported
  • A system update restarts sleds, and running instances on a sled go to the failed state when it restarts, so guests that dislike hard shutdowns should be stopped first

The Oxide Cloud Computer family

1 technologies built from or on Oxide Cloud Computer

Alternatives

What all virtual machines have in common →Compare side by side →
Amazon EC2AWSAWS’s virtual machines: pick an instance type, boot an image, and you have a server in minutes.HashiCorp VagrantCommercialA command-line tool that builds the same development virtual machine for everyone from a Vagrantfile kept in the repo.KubeVirtOpen sourceAdds virtual machines to Kubernetes as ordinary resources, so VMs and containers share one cluster, one API and one set of tools.

How it fits with other technologies

From zero to a running VM with the `oxide` CLI, as in the docs’ CLI quickstart · bash

oxide auth login --host https://engineering.cloud.example.com
oxide project create --name prototypes --description "Experimental work"

# Import a raw cloud image as a disk, snapshot it and turn it into an image
curl -OL https://cloud.debian.org/images/cloud/bookworm/latest/debian-12-genericcloud-amd64.raw
oxide disk import --path debian-12-genericcloud-amd64.raw --project prototypes \
  --disk bookworm-boot --description "Debian bookworm disk" \
  --snapshot bookworm-snapshot --image bookworm \
  --image-description "Debian bookworm image" --image-os Debian --image-version 12

oxide current-user ssh-key create --name personal --description "My personal ssh key" \
  --public-key "$(cat $HOME/.ssh/cloud_rsa.pub)"

# 8 vCPUs, 16 GiB of memory and a 200 GiB boot disk
oxide instance from-image --image bookworm --name pegasus --project prototypes \
  --description "Phasing cloaking device" --hostname pegasus \
  --memory 16g --ncpus 8 --size 200g --start
oxide instance external-ip list --project prototypes --instance pegasus

See it in a real stack

No architecture uses Oxide Cloud Computer yet. Drop it into the playground and wire it up yourself.Open playground →
Where it came from2019–2026

Companies that run their own data centers buy servers, switches, storage and a hypervisor from different vendors, then integrate them themselves, while the public clouds design their own racks and keep them to themselves. When Bryan Cantrill, Steve Tuck and Jess Frazelle started Oxide Computer Company in 2019, Cantrill wrote that the hyperscalers’ rack-level designs were “a clear win” and asked why they were “cloistered among so few”, and that from running a cloud the hardest problems had come from hardware and firmware working at cross purposes with the system software.

Oxide built the whole stack: its own boards, a service processor and root of trust instead of a BMC, its own switch, host OS, hypervisor userspace, distributed storage and control plane. The first system software release, v1.0.0 for “Oxide Computer Model 0”, is dated July 2023, and on 26 October 2023 Cantrill announced general availability of what Oxide calls “the world’s first commercial cloud computer”, shipping in a single crate with all its software included and open source.

Software ships as numbered system releases: v18 (February 2026) added local disks and IPv6, and v23 (September 2026) IPv6-only rack networking. A second-generation sled, Cosmo, with an AMD EPYC 9005 processor, raised the core count per sled from 64 to up to 192. In its October 2026 Series D announcement Oxide said it had become profitable and had scaled manufacturing capacity by 20 times in a year, with demand still exceeding supply.

Why it’s built this way

The rack, not the server, as the unit of design
Cantrill’s launch post argues that a cloud computer means breaking out of the 1U or 2U server. Designing the rack let Oxide use taller sleds with 80 mm fans running as slow as 2,000 RPM, and a cabled backplane so sleds need no cabling at all.
Build the switch too
Integrating third-party switches would have brought back exactly the integration pain Oxide wanted to remove for customers, so it built the Sidecar switch and its networking software despite the extra scope.
All software included, all of it open
On-premises buyers suffer from license management: paying for unused licenses and juggling terms and expiry dates. Oxide includes all system software with no extra licensing, and publishes it as open source with no open-core split.
Everything through the public API
The console is built on the same API as the CLI and SDKs, so anything a user can click is also scriptable, which is what makes the rack feel like a cloud rather than an appliance.

Timeline

  1. 2019

    Oxide Computer Company founded by Bryan Cantrill, Steve Tuck and Jess Frazelle

  2. 2023

    Software v1.0.0 for Oxide Computer Model 0 (July); general availability announced (October 26)

  3. 2025

    $100M Series B (July)

  4. 2026

    Local disks and IPv6 in release v18 (February); IPv6-only rack networking in v23 (September)

Sources16 · facts checked 13 October 2026

Improve this page on GitHub