Skip to content
ROBO26
NODE OPERATOR
ML ENGINEER
INDIE RESEARCHER
RENDER FARM
PROTOCOL ENG

ROBO™

By Andrew

COMPUTE

Train, infer and render on compute nobody had to build twice

ROBO — full-stack compute cloud

Independently operated hardware, one control plane

Scroll to explore

SETTLED · 412.6 $ROBO · inference · frankfurtATTACHED · 24 accelerators · são pauloSETTLED · 1,904.2 $ROBO · training · oregonSTAKED · 60,000 $ROBO · operator halomeshSETTLED · 88.4 $ROBO · render · reykjavíkVERIFIED · sampled replication · 312 jobsSETTLED · 2,240.0 $ROBO · training · singaporeDRAIN SIGNALLED · 18 accelerators · torontoSETTLED · 51.9 $ROBO · inference · lagosCLAIMED · 7,412.8 $ROBO · staking rewards

Running on ROBO

HALOMESHCINDERGRIDPROOFWORKSTRATABASELOOMCELLTIDEMARKORRERY

The premise

For a decade, compute meant renting a slice of somebody else’s data center. ROBO makes it a market you route through.

0+

Accelerators live across the network

0

Countries with active capacity

0+

$ROBO settled to operators last month

ROBO is a full-stack compute cloud assembled from hardware that other people own, and run like a single machine.

Independent operators — data centers with idle racks, render studios between projects, universities with overnight capacity — attach their hardware to one network. ROBO handles scheduling, placement, storage, execution proofs and settlement, so the person running a job never has to think about whose GPU they landed on.

That means the same control plane covers all three shapes of work. Distributed training across interconnect-aware clusters. Inference behind an endpoint that autoscales to zero. Frame-parallel rendering that bursts to thousands of nodes and disappears again.

Capacity is priced continuously rather than published in a rate card. When supply is abundant, you pay less; when you need guaranteed placement, you reserve it. Settlement runs on an EVM chain: a job escrows its budget before it is scheduled, and the contract releases $ROBO against a signed record of what actually ran.

Read the technical overview

Andy Wang

Founder, Head of Infrastructure

We stopped treating compute as a place we rent and started treating it as a market we route through. Our cost per trained token fell by two thirds and we deleted an entire capacity-planning function.

01
Network

One grid, ten thousand operators

ROBO does not own a data center. Capacity comes from independent operators who attach hardware to a single scheduler and get paid for what they actually run. That makes supply broader and cheaper than any one provider can be — and makes placement, verification and trust the hard problems.

In this section, we’ll cover:

  • Where the capacity actually comes from
  • How placement decides who runs your job
  • Proving a job ran the way you asked
  • Failure, drain and the long tail
  • Regions, egress and data gravity
Read Network

02
Workloads

Train, infer and render on one substrate

Most platforms are good at one shape of work. ROBO runs all three on the same network because they fail in complementary ways — training is bursty and interconnect-bound, inference is steady and latency-bound, rendering is embarrassingly parallel and price-bound. Mixed together, they fill each other's troughs.

In this section, we’ll cover:

  • The three shapes of work
  • Training across borrowed clusters
  • Inference that scales to zero
  • Rendering as the network's ballast
  • What still does not belong here
Read Workloads

03
Economics

A market, not a price list

Capacity on ROBO is priced continuously by a clearing auction rather than published in a rate card. That makes compute cheaper on average and less predictable in the moment — so most of the product work is in giving customers instruments to buy certainty back when they want it.

In this section, we’ll cover:

  • How a job gets priced
  • Buying certainty back
  • Settlement on an EVM chain
  • Three ways to earn
  • What operators actually earn
  • Emissions, and why they end
  • Disputes
Read Economics

Tokenomics

The network pays for the work it can prove

$ROBO · settled on EVM

ROBO settles on an EVM chain. A job escrows its budget in a contract before it is scheduled; when the network verifies the work actually ran, the contract releases $ROBO to whoever did it. No invoices, no thirty-day terms, no trusting a counterparty you have never met.

Where a job's budget goes

  • Operator who ran the work82%
  • Network fee11%
  • Verification and settlement5%
  • Reliability insurance pool2%

01

Operate

Attach accelerators and get paid per verified accelerator-second. Settlement is continuous rather than monthly, and an operator posts a bond that is slashed for vanishing mid-job.

82%

of gross job value goes to the operator

02

Stake

Back an operator you trust with $ROBO and share their fee stream. You are underwriting their reliability, so you share the slashing too — this is a bond, not a deposit.

18%

of circulating supply is staked behind operators

03

Verify

Run replication checks on sampled jobs. Verifiers are paid from the settlement fee and earn more for catching a mismatch than for confirming one.

0.4%

of settled jobs end up disputed

Read the economics in full

Agent interfaces

Seven agents. Seven ways onto the same network.

Coming soon

Four more agent interfaces in build

Proofwork, Cindergrid, Stratabase and Tidemark cover verification, the operator side, scheduling and treasury. Each is a different front end onto the same settlement contract.

See all seven agent interfaces

Network at a glance

This network runs on

0+

Accelerators

0

Independent operators

0%

Of supply staked behind them

0

Countries