┌─────────┬─────────────────────────────────────────┐ │ 0xIDE-R3 │ OXIDE CLOUD COMPUTER // REV 3.1 │ ├─────────┼─────────────────────────────────────────┤ │ U01-U02 │ [SLD-00] ████████████████████░░░░ AMD MILAN OK │ │ U03-U04 │ [SLD-01] ██████████████████████░░ AMD MILAN OK │ │ U05-U06 │ [SLD-02] ██████████████░░░░░░░░░░ AMD MILAN & │ │ U07-U08 │ [SLD-03] ██████████████████████░░ AMD MILAN OK │ │ ////// │ ░░░░░░░░░░░░░░░░░░░░░░░░ [CUBBIES 05-16 AVAIL] │ │ SWP-A │ [NET-A] ●●●●●●●●●●●●●●●● 100GbE PRIMARY │ │ SWP-B │ [NET-B] ●●●●●●●●●●●●●●●● 100GbE STANDBY │ │ PSU-SHELF│ [PWR] ████████████ 18kW DC-BUSBAR OK │ └─────────┴─────────────────────────────────────────┘ STATUS: NOMINAL // 24°C // UPTIME: 847d 14h 22m

Servers as they should be █

THE CLOUD COMPUTER

Rack-scale co-design. One-time purchase. No licensing fees. Hyperscaler-class infrastructure you own — deployable in under two hours.

Explore the Rack Product Overview

Product

// WHAT_IS_OXIDE

[HW]

Co-Designed Hardware

Every component — PCB layout, switch ASIC integration, power distribution — is custom-built by Oxide. Blind-mated sleds snap in without a single cable. The rack ships fully assembled and tested.

AMD Milan EPYC // NVMe U.2 // DC Busbar

[SW]

Integrated Software Stack

The entire stack — firmware to cloud control plane — ships with the rack. Virtual compute, elastic block storage, VPC networking, and a full REST API. No VMware. No per-core licensing fees. Ever.

Console // API // CLI // SDK

[NET]

100GbE High-Availability Networking

Dual redundant switches running Delay Driven Multipath. OPTE handles firewalling, routing, NAT, and VPC encapsulation at line rate. Failover in under 100ms — transparent to workloads.

DDM // Geneve Encap // OPTE Engine

[STR]

Resilient Storage

OpenZFS checksums and scrubs all data continuously. Automated rebalancing preserves redundancy when drives or sleds are removed. Up to 10x U.2 NVMe per sled — up to 1PB raw per rack.

OpenZFS // EBS-equivalent // Auto-Rebalance

[OPS]

Front-Serviceable Design

Full serviceability from the cold aisle. Hot-plug cubby bays mean every FRU is reachable without downtime. Power shelf OOB management via RJ-45 on the front panel. No rear access required.

Hot-Plug // Cold Aisle // Zero-Downtime FRU

[API]

Cloud-Parity Developer UX

Provision VMs, block storage, and VPCs through a REST API, CLI, or web console — identical to AWS or GCP, but on iron you own. Multi-tenant, self-service. No ticket, no queue.

REST // Terraform // Python SDK // Rust SDK

Sales

// WHY_OXIDE

Traditional racks force separate procurement of servers, switches, storage, and software — then weeks of wiring, testing, and licensing each layer. Hyperscalers solved this by building it all themselves. Oxide brings that same advantage to enterprises who buy rather than rent.

Capability	Oxide	Traditional
Installation time	< 2 hours	Weeks–Months
Software licensing	Included ($0)	Per-core fees
Cabling	Zero (blind-mated)	Extensive
Space efficiency	33% smaller footprint	Baseline
Cloud-style API	Native, built-in	Add-on / Optional
Data sovereignty	Full ownership	Vendor-dependent
Max rack power	< 18kW DC bus	Variable / higher
Firmware openness	Open source (Hubris)	Proprietary blob

// Ideal Workloads

AI / ML Training

Data-intensive training where your data belongs — on hardware you control at a fixed, predictable cost. No per-GPU-hour surprises.

CI/CD Infrastructure

Dedicated build runners with security baked in. Consistent performance, no noisy neighbors, fixed monthly cost.

Data Engineering

Batch compute, staging, and orchestration on multi-tenant infrastructure you own. Data stays in your authority.

Regulated Industries

Finance, healthcare, and defense workloads requiring hardware verification, air-gap capability, and data residency.

Support

// FRU_REPLACEMENT_GUIDE

Field-replaceable unit tutorials for on-site operators. All procedures are designed for single-technician execution from the cold aisle without scheduled downtime windows.

Compute Sled Replacement▼

DIFFICULTY: LOW | TIME: ~15 MIN | TOOLS: NONE | DOWNTIME: NONE (LIVE MIGRATION)

Compute sleds are hot-pluggable with no tools required. The blind-mate connector self-aligns on insertion. Live migration moves all running VMs before extraction begins.

Open Oxide Console → Sleds → Select target sled
Run: oxide sled evacuate --sled-id $SLED_ID
Confirm zero running instances on sled
Press release lever on sled front panel
Slide sled straight out — no cables to disconnect
Insert replacement sled until lever clicks
Verify LED turns solid green within 90s
Run: oxide sled commission --sled-id $NEW_ID

SLED EVACUATION

$ oxide sled evacuate --sled-id sled-02

Migrating 4 instances...

✓ db-primary → sled-01

✓ api-server → sled-03

✓ cache-01 → sled-04

✓ cache-02 → sled-05

Sled clear. Safe to remove.

ESD wrist strap required. Do not touch PCB contacts. Handle sled by chassis handles only.

NVMe Drive Replacement (U.2)▼

DIFFICULTY: LOW | TIME: ~10 MIN | TOOLS: NONE | DOWNTIME: NONE (OpenZFS AUTO-HEAL)

NVMe drives sit in hot-swap U.2 bays on the sled front face. OpenZFS detects failure, marks the drive FAULTED, and immediately begins rebuilding from parity data. No operator action required until physical swap.

Oxide Console alert: DRIVE FAULT — sled-03, bay-07
Confirm: oxide storage disk list --sled sled-03
Locate bay-07 LED (solid red = faulted)
Press bay eject button — drive unlocks
Pull drive by handle, set aside
Insert replacement drive until audible click
ZFS pool rebuild starts automatically
Monitor: oxide storage pool status --watch

STORAGE STATUS

$ oxide storage pool status

pool: data state: DEGRADED

drive: sled-03/bay-07 FAULTED

↻ resilvering: 847GiB / 3.2TiB (26%)

est. remaining: 38 minutes

Wait for drive LED to extinguish before removal. Drives spin down 10s after last I/O to prevent data corruption.

Network Switch Replacement▼

DIFFICULTY: MEDIUM | TIME: ~30 MIN | TOOLS: NONE | DOWNTIME: NONE (DUAL-SWITCH HA)

Each rack ships with two switches running simultaneously via DDM. Removing one switch shifts all traffic to the peer in milliseconds. No service interruption for running VMs or storage.

Verify peer switch ACTIVE: Console → Network → Switches
Set failing switch to STANDBY via OOB port
Confirm all sled links fail over to peer
Disconnect OOB management cable (front RJ-45)
Press release levers simultaneously on both switch ears
Slide switch out — no other cables connected
Insert replacement, both ears click into position
Reconnect OOB cable, wait for auto-config (<5 min)
Promote: oxide network switch set-active --id swp-01

SWITCH STATUS

$ oxide network switch list

ID ROLE STATE UPTIME

swp-01 primary active 14d 06h

swp-02 standby faulted —

$ oxide network failover --to swp-01

✓ All 32 sleds on swp-01. Safe to remove swp-02.

Never remove both switches simultaneously. Always verify full HA failover before extracting the faulted unit.

Power Shelf PSU Replacement▼

DIFFICULTY: MEDIUM | TIME: ~20 MIN | TOOLS: TORQUE DRIVER | DOWNTIME: NONE (N+1)

Power is distributed via a low-voltage DC bus bar pair. The shelf runs N+1 redundant PSUs — a single failure triggers an alert but remaining units carry full rack load.

Identify faulted PSU via amber LED on power shelf
Confirm remaining PSUs show green LEDs
Connect to power shelf OOB: front RJ-45
Run: oxide power psu status — confirm N+1 sufficient
Loosen 2x captive screws (torque driver, front access)
Pull PSU handle — unit slides free
Insert replacement PSU, tighten captive screws
Verify PSU LED green, shelf shows NOMINAL

POWER STATUS

$ oxide power psu status

PSU STATE OUTPUT TEMP

A nominal 4.2kW 41°C

B nominal 4.2kW 39°C

C faulted 0.0kW —

D nominal 4.1kW 40°C

⚠ N+1 maintained. Replace PSU-C within 72h.

DC bus bar carries high current. Do not bypass safety interlocks. Ensure PSU is fully seated before re-energizing shelf.

Support

// SUPPORT_CHANNELS

[DOC]

Documentation

Full operator and developer docs including architecture guides, API reference, and hardware specifications. Updated with every software release at docs.oxide.computer.

[OPS]

On-Call Engineering

24/7 engineering support for production incidents. Direct line to the same engineers that built the rack. No tier-1 gatekeeping, no ticket queues.

[FRU]

FRU Dispatch

Replacement parts dispatched same business day. Full inventory of sleds, drives, switches, and PSUs maintained at regional depots.

[UPD]

Software Updates

Rolling zero-downtime updates. Firmware, OS, and control plane updated atomically through the Oxide update service — no maintenance windows.

[TRN]

Operator Training

On-site training covering rack operations, all FRU procedures, console management, and API integration patterns for your team.

[COM]

RFD Community

Open Request for Discussion process. Participate in roadmap decisions, browse design documents, and discuss directly with Oxide engineers.

THE CLOUD COMPUTER

// RACK_EXPLORER

// WHAT_IS_OXIDE

// WHY_OXIDE

// FRU_REPLACEMENT_GUIDE

// SUPPORT_CHANNELS