/ workers

The mesh that runs the models.

ORVEX has no data centers of its own. Every request is served by an independent worker a GPU, somewhere, running our runtime and earning credits for it. This is how that network operates.

/ 01 · what workers are

A worker is compute, not a company.

Anyone with a capable GPU can join the mesh. Workers stake for priority, run the runtime binary, and start receiving jobs within minutes.

/ contribute

Contribute compute

Idle GPU cycles are matched against live inference demand across the network.

/ assign

Jobs get assigned

The scheduler scores every worker on latency, load and price, then routes in under a millisecond.

/ earn

Rewards accrue

Completed inference settles instantly as credits redeemable, stakeable, on-chain.

/ decentralize

No single point of failure

Thousands of independent nodes beat a handful of hyperscale regions on resilience and price.

/ 02 · network overview

The network, right now.

A live dashboard view of the mesh. Every number below moves continuously this is the same telemetry our schedulers see.

workers online
214
available gpus
891
jobs running
3,402
inference requests / s
4,820
avg latency
41ms
success rate
99.60%
network throughput
12,400 tok/s
credits processed / min
1,820 cr
/ 03 · topology

Requests find the best worker.

A simplified view of the mesh: the scheduler holds no state of its own it just routes. Watch a request travel from client to scheduler to worker.

CLIENTrequestSCHEDULERROUTINGNODE·01NODE·02NODE·03NODE·04NODE·05NODE·06NODE·07
/ 04 · fleet

A sample of active workers.

Status, hardware and throughput for a slice of the live fleet. The real fleet is larger and rotates constantly.

wrk_a19f2cStreaming
Portland, US · H100
jobs done48,210
latency32ms
uptime99.98%
reward ×1.4×
reporting now
wrk_c72e91Busy
Ashburn, US · RTX 4090
jobs done21,904
latency44ms
uptime99.81%
reward ×1.1×
reporting now
wrk_9d10bbIdle
Dublin, IE · H100
jobs done61,233
latency38ms
uptime99.95%
reward ×1.3×
reporting now
wrk_88bafeBusy
Frankfurt, DE · RTX 3090
jobs done15,877
latency51ms
uptime99.62%
reward ×1.0×
reporting now
wrk_4501aaStreaming
Tokyo, JP · RTX 5090
jobs done9,021
latency29ms
uptime99.90%
reward ×1.6×
reporting now
wrk_2bebcdOffline
São Paulo, BR · M3 Max
jobs done4,310
latency68ms
uptime97.10%
reward ×0.8×
last seen 14m ago
/ 05 · activity

Live from the network.

A rolling feed of what's happening across the mesh connections, completions, rewards, syncs.

/ live activity
streaming
events / min
38
active regions
11
/ 06 · performance

How the fleet is performing.

Latency, distribution and saturation, generated straight from the runtime no snapshots, no cached charts.

latency
41ms
gpu usage · fleet
75%
queue length
2
success rate
99.58%
throughput
8308
worker distribution
by region
us-west
24%
us-east
19%
eu-west
21%
ap-ne
14%
ap-se
12%
other
10%
/ 07 · job routing

From request to response.

Every job takes the same five-stage path through the runtime the scheduler never becomes the bottleneck.

01
User Request
arrives at the edge
02
Scheduler
scores latency, load, price
03
Best Worker
selected in <1ms
04
Inference
model executes
05
Streaming Response
tokens return live
/ 08 · become a worker

Turn your GPU into revenue.

Four steps between an idle card and a running node.

step 01

Download Worker

Grab the runtime binary for Linux, macOS or Windows. No account required to start.

step 02

Connect GPU

The installer detects your hardware and benchmarks it against the network's baseline.

step 03

Receive Jobs

Once verified, the scheduler starts routing inference and embedding jobs to your node.

step 04

Earn Rewards

Credits settle per completed job and accrue toward your on-chain balance in real time.

/ 09 · benefits

Why run a node.

Decentralized inference isn't just an architecture choice it changes the economics for everyone who contributes compute.

/ 01

Earn passive income

Idle GPU time converts directly into credits no marketplace listing, no negotiation.

/ 02

Privacy focused

Workers never see raw request content beyond what's required to execute the job.

/ 03

No central servers

There is no ORVEX data center to fail. The network is the infrastructure.

/ 04

Global network

Nodes across six continents mean requests are served from wherever they originate.

/ 05

Stake for higher rewards

Staking credits against your node raises its priority tier and reward multiplier.

/ 06

Open infrastructure

The routing protocol and worker spec are open anyone can build a compatible client.

/ 10 · hardware

Supported hardware.

Estimated performance and reward tier for common cards. Actual rewards depend on live demand and uptime.

RTX 5090

tier 1
memory32GB GDDR7
est. performance~410 tok/s
est. reward1.6×

RTX 4090

tier 1
memory24GB GDDR6X
est. performance~340 tok/s
est. reward1.4×

RTX 3090

tier 2
memory24GB GDDR6X
est. performance~210 tok/s
est. reward1.0×

Apple M3 Max

tier 2
memory128GB unified
est. performance~180 tok/s
est. reward0.9×

RX 7900 XTX

tier 2
memory24GB GDDR6
est. performance~195 tok/s
est. reward0.95×

H100 (datacenter)

tier 0
memory80GB HBM3
est. performance~980 tok/s
est. reward2.2×