Skip to main content

// Chapter 11 · Concepts

Run a Node

From bare metal to first claimed reward.

5 min1 sectionsConcepts

From bare metal to first claimed reward.

// 11.0 · seven steps · live now · permissionless

// Operator path

Hardware check, install, keys, register, run, claim, exit. Seven steps.

The Console, the Node CLI, and the NodeRegistryLocker contract are live. Onboarding is permissionless: anyone who stakes the 50,000 $PRLX minimum can register, with no allowlist.

// Operator earnings looplive
ParalleliX AI · inference requests
  1. 01

    Stake

    $PRLX on NodeRegistryLocker

  2. 02

    Run

    uptime heartbeats

  3. 03

    Serve

    one whole AI request

  4. 04

    Prove

    SHA-256 commitment

  5. 05

    Earn

    per-second stream

Rewards stream from the 25% Operator Rewards bucket, unlocked linearly via Sablier (~34,200 $PRLX/day), streamed per-second to online nodes by stake x tier. No slashing: principal returns in full on unstake.

// Console · Operate · dashboardOpen in console
// operator · dashboardYour nodes
3 active
nodestateuptimeweight
#4217active99.2%GPU · 2.5x
#4218active98.0%GPU · 3.0x
#4220active94.7%CPU · 1.0x
#4231idle0%
nodes online3 / 4
avg uptime97.3%
stake posted410k $PRLX
accrued reward23,515 $PRLX

Rewards accrue from the 25% Operator Rewards bucket, weighted stake x tier x uptime. Idle nodes (no compute online) earn nothing.

Operator-side view: every registered node, its state and uptime, and accrued rewards. Operator-only, no submitter view.

// Console · Operate · registerOpen in console
// operator · stake & registerone stake = one node
  1. 01

    Hardware tier

    Declare CPU or GPU tier. The CLI verifies it.

  2. 02

    Stake amount

    Lock >= 50,000 $PRLX. Higher stake, higher weight.

  3. 03

    Sign registerNode

    registerNode(stake, tier, nodeKeyHash) -> nodeId

node #4231 registered · principal returned in full on unstake (7-day cooldown, no slashing)

Stake to register: declare hardware tier, lock at least 50,000 $PRLX, sign registerNode. The node key is generated locally by the CLI.

The seven steps

// 11.1 · linear sequence · permissionless from launch

// Operator bootstrap · 7 stepsat launch · permissionless

  1. // step 01

    Install + probe

    Run the one-line installer (curl -fsSL https://parallelix.io/install.sh | sh) to get Node, Ollama, the default model, and the CLI. Then parallelix-node probe detects your GPU and prints the tier: CPU is 1, GPU is 2/3/4 by VRAM. macOS and Linux are supported; Windows runs inside WSL2.

  2. // step 02

    Pull a model

    parallelix-node models pull llama3.2 ensures a model is installed in Ollama. The node serves inference with it, so at least one model must be present before start.

  3. // step 03

    Init the node key

    parallelix-node init generates a separate, low-value secp256k1 node key (never your staking wallet) and prints the nodeKeyHash. The node key signs liveness and results only and never leaves the machine.

  4. // step 04

    Register on-chain

    In the Console, approve $PRLX and sign registerNode(stakeAmount, hardwareTier, nodeKeyHash) on NodeRegistryLocker, locking stakeAmount (>= 50,000 $PRLX). One stake = one node; register again for more, unlimited per wallet.

  5. // step 05

    Run

    parallelix-node start --node-id <id> --gpu attaches the machine, sends signed heartbeats, and serves dispatched inference requests through Ollama, each whole on one node. Being online is the accrual condition.

  6. // step 06

    Claim

    While the node is online, OperatorStakeRewardsV2 streams to its earned(nodeId) balance every block (weighted by stake × tier). Call claim(nodeId) anytime to pull it to wei precision.

  7. // step 07

    Exit

    requestUnstake(nodeId) retires the node and starts the 7-day cooldown. After 7 days, withdrawStake(nodeId) returns the full principal: no slashing, no deductions.

Illustrative CLI session

// 11.2 · pseudocode · not a stable interface

Caveat·Pseudocode

The session below is illustrative. Command names, flags, and output formats are not finalised and may change before v1. Do not script against them.
// bash · pre-v1 pseudocode
$ parallelix-node probe
· gpu       NVIDIA H100 · 80 GB VRAM
· ollama    running (v0.x)
✓ maps to tier 4 (GPU · datacenter)

$ parallelix-node init
✓ node key generated at ~/.parallelix/node.key (mode 0600)
// register this node on-chain (in the Console), using:
   tier         4 (GPU · datacenter)
   nodeKeyHash  0x9c27a4...

# stake + registerNode in the Console → node id 4217

$ parallelix-node start --node-id 4217 --gpu
// parallelix-node 1.2.1
✓ daemon up · heartbeating + serving · ctrl-c to stop
● active · heartbeats 142 · served 37 · running inference

$ parallelix-node service --node-id 4217 --gpu
✓ installed + started as systemd service 'parallelix-node'

# claim + unstake are Console actions signed by the staking wallet:
#   claim(4217)          · pulls earned(4217) to your wallet
#   requestUnstake(4217) · 7-day cooldown, then full principal returned

Operator year 1

// 11.3 · worked walkthrough · figures illustrative

Concrete numbers for one hypothetical operator across twelve months. Tier 3 hardware (4 × H100), eu-west, ai.training and ai.inference workloads, target 99.5% uptime.

All figures are illustrative. Real values depend on the total online stake-weight, which grows as more nodes join and compresses each node's share of the same reward stream. The full month-by-month table is in the §11.3 subsection.

// Year-1 metrics · illustrative

  • // Gross rewards Y1

    ~220,970 $PRLX

    Per-second share of the live reward stream

  • // Operating cost

    $21,600 USD

    Power and bandwidth, rough

  • // Break-even

    ~$0.098 / $PRLX

    Implied per-token price

  • // Principal at risk

    0 $PRLX

    No slashing; stake returned in full

// Risk register · cost · mitigation5 entries

  • Going offline

    Accrual stops while offline; stream redistributes

    Run as a service; monitor heartbeats.

  • Ollama not running

    Node cannot serve; the daemon refuses to start

    parallelix-node verify; start Ollama.

  • Hardware-tier mismatch

    Node cannot serve at the declared tier

    Declare the tier probe reports.

  • Machine sleeps or reboots

    Node goes idle and earns nothing while down

    Install as a systemd/launchd service.

  • Total online weight grows

    Per-node share of the same stream falls

    Stake more (more weight) or run more nodes.

The table compresses 12 months of speculative parameters into one artifact. Treat it as a calibration tool, not a forecast. Real numbers publish in the Telemetry surface of app.parallelix.io once the first production inference request settles.

// Where to go next · reading path